Everyone has a dream. But sometimes there’s a gap between where we are and where we want to be. True, there are some people who can bridge that gap easily, on their own, but all of us need a little help at some point. A little boost. An accountability partner. A Snooze Squad. In each episode, the Snooze Squad will strategize an action plan for people to face their fears. Guests will transform their own perception of their potential and walk away a few inches closer to who they want to become ...
…
continue reading
Player FM - Internet Radio Done Right
Checked 1y ago
اضافه شده در four سال پیش
Looks like the publisher may have taken this series offline or changed its URL. Please contact support if you believe it should be working, the feed URL is invalid, or you have any other concerns about it.
محتوای ارائه شده توسط The Nonlinear Fund. تمام محتوای پادکست شامل قسمت‌ها، گرافیک‌ها و توضیحات پادکست مستقیماً توسط The Nonlinear Fund یا شریک پلتفرم پادکست آن‌ها آپلود و ارائه می‌شوند. اگر فکر می‌کنید شخصی بدون اجازه شما از اثر دارای حق نسخه‌برداری شما استفاده می‌کند، می‌توانید روندی که در اینجا شرح داده شده است را دنبال کنید.https://fa.player.fm/legal
مشابه The Nonlinear Library: Alignment Forum
Interviews with mathematics education researchers about recent studies. Hosted by Samuel Otten, University of Missouri. www.mathedpodcast.com Produced by Fibre Studios
…
continue reading
Five-time winner of Best Education Podcast in the Podcast Awards. Grammar Girl provides short, friendly tips to improve your writing and feed your love of the English language. Whether English is your first language or your second language, these grammar, punctuation, style, and business tips will make you a better and more successful writer. Grammar Girl is a Quick and Dirty Tips podcast.
…
continue reading
Making It is a weekly audio podcast that comes out every Friday hosted by Jimmy Diresta, Bob Clagett and David Picciuto. Three different makers with different backgrounds talking about creativity, design and making things with your bare hands.
…
continue reading
A science guy from the deep south (Destin) and a humanities guy from the wild west (Matt Whitman) discuss deep questions with varying levels of maturity.
…
continue reading
Player FM - برنامه پادکست
با برنامه Player FM !
با برنامه Player FM !
))
Daily Deals
پادکست هایی که ارزش شنیدن دارند
حمایت شده
For the first time, Skip Intro goes to The Paris Theater in Manhattan to sit down with Rebecca Ferguson who stars as Captain Olivia Walker in A HOUSE OF DYNAMITE . Directed by Oscar-winning Kathryn Bigelow, the film was screened this month at the historic 535-seat theater — NYC’s longest-running arthouse cinema. Ferguson shares what it was like to read the powerful script written by Noah Oppenheim and how working with Bigelow was unlike any other experience on set. The Mission Impossible actor also talks about jumping off the roof of the Vienna State Opera with Tom Cruise, Denis Villeneuve’s love of veils and jingle jangles in Dune , and teases Netflix’s upcoming Peaky Blinders movie with Cillian Murphy. Video episodes available on Still Watching Netflix YouTube Channel. Listen to more from Netflix Podcasts .…
AF - Sycophancy to subterfuge: Investigating reward tampering in large language models by Evan Hubinger
بایگانی مجموعه ها ("فیدهای غیر فعال" status)
When?
This feed was archived on October 23, 2024 10:10 (
Why? فیدهای غیر فعال status. سرورهای ما، برای یک دوره پایدار، قادر به بازیابی یک فید پادکست معتبر نبوده اند.
What now? You might be able to find a more up-to-date version using the search function. This series will no longer be checked for updates. If you believe this to be in error, please check if the publisher's feed link below is valid and contact support to request the feed be restored or if you have any other concerns about this.
Manage episode 424153726 series 3337166
محتوای ارائه شده توسط The Nonlinear Fund. تمام محتوای پادکست شامل قسمت‌ها، گرافیک‌ها و توضیحات پادکست مستقیماً توسط The Nonlinear Fund یا شریک پلتفرم پادکست آن‌ها آپلود و ارائه می‌شوند. اگر فکر می‌کنید شخصی بدون اجازه شما از اثر دارای حق نسخه‌برداری شما استفاده می‌کند، می‌توانید روندی که در اینجا شرح داده شده است را دنبال کنید.https://fa.player.fm/legal
Link to original article
Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Sycophancy to subterfuge: Investigating reward tampering in large language models, published by Evan Hubinger on June 17, 2024 on The AI Alignment Forum.
New Anthropic model organisms research paper led by Carson Denison from the Alignment Stress-Testing Team demonstrating that large language models can generalize zero-shot from simple reward-hacks (sycophancy) to more complex reward tampering (subterfuge).
Our results suggest that accidentally incentivizing simple reward-hacks such as sycophancy can have dramatic and very difficult to reverse consequences for how models generalize, up to and including generalization to editing their own reward functions and covering up their tracks when doing so.
Abstract:
In reinforcement learning, specification gaming occurs when AI systems learn undesired behaviors that are highly rewarded due to misspecified training goals. Specification gaming can range from simple behaviors like sycophancy to sophisticated and pernicious behaviors like reward-tampering, where a model directly modifies its own reward mechanism. However, these more pernicious behaviors may be too complex to be discovered via exploration.
In this paper, we study whether Large Language Model (LLM) assistants which find easily discovered forms of specification gaming will generalize to perform rarer and more blatant forms, up to and including reward-tampering. We construct a curriculum of increasingly sophisticated gameable environments and find that training on early-curriculum environments leads to more specification gaming on remaining environments.
Strikingly, a small but non-negligible proportion of the time, LLM assistants trained on the full curriculum generalize zero-shot to directly rewriting their own reward function. Retraining an LLM not to game early-curriculum environments mitigates, but does not eliminate, reward-tampering in later environments. Moreover, adding harmlessness training to our gameable environments does not prevent reward-tampering.
These results demonstrate that LLMs can generalize from common forms of specification gaming to more pernicious reward tampering and that such behavior may be nontrivial to remove.
Twitter thread:
New Anthropic research: Investigating Reward Tampering.
Could AI models learn to hack their own reward system?
In a new paper, we show they can, by generalization from training in simpler settings.
Read our blog post here: https://anthropic.com/research/reward-tampering
We find that models generalize, without explicit training, from easily-discoverable dishonest strategies like sycophancy to more concerning behaviors like premeditated lying - and even direct modification of their reward function.
We designed a curriculum of increasingly complex environments with misspecified reward functions.
Early on, AIs discover dishonest strategies like insincere flattery. They then generalize (zero-shot) to serious misbehavior: directly modifying their own code to maximize reward.
Does training models to be helpful, honest, and harmless (HHH) mean they don't generalize to hack their own code?
Not in our setting. Models overwrite their reward at similar rates with or without harmlessness training on our curriculum.
Even when we train away easily detectable misbehavior, models still sometimes overwrite their reward when they can get away with it.
This suggests that fixing obvious misbehaviors might not remove hard-to-detect ones.
Our work provides empirical evidence that serious misalignment can emerge from seemingly benign reward misspecification.
Read the full paper: https://arxiv.org/abs/2406.10162
The Anthropic Alignment Science team is actively hiring research engineers and scientists. We'd love to see your application: https://boards.greenhouse.io/anthropic/jobs/4009165008
Blog post:
Perverse incentives are everywhere. Thi...
…
continue reading
Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Sycophancy to subterfuge: Investigating reward tampering in large language models, published by Evan Hubinger on June 17, 2024 on The AI Alignment Forum.
New Anthropic model organisms research paper led by Carson Denison from the Alignment Stress-Testing Team demonstrating that large language models can generalize zero-shot from simple reward-hacks (sycophancy) to more complex reward tampering (subterfuge).
Our results suggest that accidentally incentivizing simple reward-hacks such as sycophancy can have dramatic and very difficult to reverse consequences for how models generalize, up to and including generalization to editing their own reward functions and covering up their tracks when doing so.
Abstract:
In reinforcement learning, specification gaming occurs when AI systems learn undesired behaviors that are highly rewarded due to misspecified training goals. Specification gaming can range from simple behaviors like sycophancy to sophisticated and pernicious behaviors like reward-tampering, where a model directly modifies its own reward mechanism. However, these more pernicious behaviors may be too complex to be discovered via exploration.
In this paper, we study whether Large Language Model (LLM) assistants which find easily discovered forms of specification gaming will generalize to perform rarer and more blatant forms, up to and including reward-tampering. We construct a curriculum of increasingly sophisticated gameable environments and find that training on early-curriculum environments leads to more specification gaming on remaining environments.
Strikingly, a small but non-negligible proportion of the time, LLM assistants trained on the full curriculum generalize zero-shot to directly rewriting their own reward function. Retraining an LLM not to game early-curriculum environments mitigates, but does not eliminate, reward-tampering in later environments. Moreover, adding harmlessness training to our gameable environments does not prevent reward-tampering.
These results demonstrate that LLMs can generalize from common forms of specification gaming to more pernicious reward tampering and that such behavior may be nontrivial to remove.
Twitter thread:
New Anthropic research: Investigating Reward Tampering.
Could AI models learn to hack their own reward system?
In a new paper, we show they can, by generalization from training in simpler settings.
Read our blog post here: https://anthropic.com/research/reward-tampering
We find that models generalize, without explicit training, from easily-discoverable dishonest strategies like sycophancy to more concerning behaviors like premeditated lying - and even direct modification of their reward function.
We designed a curriculum of increasingly complex environments with misspecified reward functions.
Early on, AIs discover dishonest strategies like insincere flattery. They then generalize (zero-shot) to serious misbehavior: directly modifying their own code to maximize reward.
Does training models to be helpful, honest, and harmless (HHH) mean they don't generalize to hack their own code?
Not in our setting. Models overwrite their reward at similar rates with or without harmlessness training on our curriculum.
Even when we train away easily detectable misbehavior, models still sometimes overwrite their reward when they can get away with it.
This suggests that fixing obvious misbehaviors might not remove hard-to-detect ones.
Our work provides empirical evidence that serious misalignment can emerge from seemingly benign reward misspecification.
Read the full paper: https://arxiv.org/abs/2406.10162
The Anthropic Alignment Science team is actively hiring research engineers and scientists. We'd love to see your application: https://boards.greenhouse.io/anthropic/jobs/4009165008
Blog post:
Perverse incentives are everywhere. Thi...
392 قسمت
اشتراک گذاریبایگانی مجموعه ها ("فیدهای غیر فعال" status)
When?
This feed was archived on October 23, 2024 10:10 (
Why? فیدهای غیر فعال status. سرورهای ما، برای یک دوره پایدار، قادر به بازیابی یک فید پادکست معتبر نبوده اند.
What now? You might be able to find a more up-to-date version using the search function. This series will no longer be checked for updates. If you believe this to be in error, please check if the publisher's feed link below is valid and contact support to request the feed be restored or if you have any other concerns about this.
Manage episode 424153726 series 3337166
محتوای ارائه شده توسط The Nonlinear Fund. تمام محتوای پادکست شامل قسمت‌ها، گرافیک‌ها و توضیحات پادکست مستقیماً توسط The Nonlinear Fund یا شریک پلتفرم پادکست آن‌ها آپلود و ارائه می‌شوند. اگر فکر می‌کنید شخصی بدون اجازه شما از اثر دارای حق نسخه‌برداری شما استفاده می‌کند، می‌توانید روندی که در اینجا شرح داده شده است را دنبال کنید.https://fa.player.fm/legal
Link to original article
Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Sycophancy to subterfuge: Investigating reward tampering in large language models, published by Evan Hubinger on June 17, 2024 on The AI Alignment Forum.
New Anthropic model organisms research paper led by Carson Denison from the Alignment Stress-Testing Team demonstrating that large language models can generalize zero-shot from simple reward-hacks (sycophancy) to more complex reward tampering (subterfuge).
Our results suggest that accidentally incentivizing simple reward-hacks such as sycophancy can have dramatic and very difficult to reverse consequences for how models generalize, up to and including generalization to editing their own reward functions and covering up their tracks when doing so.
Abstract:
In reinforcement learning, specification gaming occurs when AI systems learn undesired behaviors that are highly rewarded due to misspecified training goals. Specification gaming can range from simple behaviors like sycophancy to sophisticated and pernicious behaviors like reward-tampering, where a model directly modifies its own reward mechanism. However, these more pernicious behaviors may be too complex to be discovered via exploration.
In this paper, we study whether Large Language Model (LLM) assistants which find easily discovered forms of specification gaming will generalize to perform rarer and more blatant forms, up to and including reward-tampering. We construct a curriculum of increasingly sophisticated gameable environments and find that training on early-curriculum environments leads to more specification gaming on remaining environments.
Strikingly, a small but non-negligible proportion of the time, LLM assistants trained on the full curriculum generalize zero-shot to directly rewriting their own reward function. Retraining an LLM not to game early-curriculum environments mitigates, but does not eliminate, reward-tampering in later environments. Moreover, adding harmlessness training to our gameable environments does not prevent reward-tampering.
These results demonstrate that LLMs can generalize from common forms of specification gaming to more pernicious reward tampering and that such behavior may be nontrivial to remove.
Twitter thread:
New Anthropic research: Investigating Reward Tampering.
Could AI models learn to hack their own reward system?
In a new paper, we show they can, by generalization from training in simpler settings.
Read our blog post here: https://anthropic.com/research/reward-tampering
We find that models generalize, without explicit training, from easily-discoverable dishonest strategies like sycophancy to more concerning behaviors like premeditated lying - and even direct modification of their reward function.
We designed a curriculum of increasingly complex environments with misspecified reward functions.
Early on, AIs discover dishonest strategies like insincere flattery. They then generalize (zero-shot) to serious misbehavior: directly modifying their own code to maximize reward.
Does training models to be helpful, honest, and harmless (HHH) mean they don't generalize to hack their own code?
Not in our setting. Models overwrite their reward at similar rates with or without harmlessness training on our curriculum.
Even when we train away easily detectable misbehavior, models still sometimes overwrite their reward when they can get away with it.
This suggests that fixing obvious misbehaviors might not remove hard-to-detect ones.
Our work provides empirical evidence that serious misalignment can emerge from seemingly benign reward misspecification.
Read the full paper: https://arxiv.org/abs/2406.10162
The Anthropic Alignment Science team is actively hiring research engineers and scientists. We'd love to see your application: https://boards.greenhouse.io/anthropic/jobs/4009165008
Blog post:
Perverse incentives are everywhere. Thi...
…
continue reading
Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Sycophancy to subterfuge: Investigating reward tampering in large language models, published by Evan Hubinger on June 17, 2024 on The AI Alignment Forum.
New Anthropic model organisms research paper led by Carson Denison from the Alignment Stress-Testing Team demonstrating that large language models can generalize zero-shot from simple reward-hacks (sycophancy) to more complex reward tampering (subterfuge).
Our results suggest that accidentally incentivizing simple reward-hacks such as sycophancy can have dramatic and very difficult to reverse consequences for how models generalize, up to and including generalization to editing their own reward functions and covering up their tracks when doing so.
Abstract:
In reinforcement learning, specification gaming occurs when AI systems learn undesired behaviors that are highly rewarded due to misspecified training goals. Specification gaming can range from simple behaviors like sycophancy to sophisticated and pernicious behaviors like reward-tampering, where a model directly modifies its own reward mechanism. However, these more pernicious behaviors may be too complex to be discovered via exploration.
In this paper, we study whether Large Language Model (LLM) assistants which find easily discovered forms of specification gaming will generalize to perform rarer and more blatant forms, up to and including reward-tampering. We construct a curriculum of increasingly sophisticated gameable environments and find that training on early-curriculum environments leads to more specification gaming on remaining environments.
Strikingly, a small but non-negligible proportion of the time, LLM assistants trained on the full curriculum generalize zero-shot to directly rewriting their own reward function. Retraining an LLM not to game early-curriculum environments mitigates, but does not eliminate, reward-tampering in later environments. Moreover, adding harmlessness training to our gameable environments does not prevent reward-tampering.
These results demonstrate that LLMs can generalize from common forms of specification gaming to more pernicious reward tampering and that such behavior may be nontrivial to remove.
Twitter thread:
New Anthropic research: Investigating Reward Tampering.
Could AI models learn to hack their own reward system?
In a new paper, we show they can, by generalization from training in simpler settings.
Read our blog post here: https://anthropic.com/research/reward-tampering
We find that models generalize, without explicit training, from easily-discoverable dishonest strategies like sycophancy to more concerning behaviors like premeditated lying - and even direct modification of their reward function.
We designed a curriculum of increasingly complex environments with misspecified reward functions.
Early on, AIs discover dishonest strategies like insincere flattery. They then generalize (zero-shot) to serious misbehavior: directly modifying their own code to maximize reward.
Does training models to be helpful, honest, and harmless (HHH) mean they don't generalize to hack their own code?
Not in our setting. Models overwrite their reward at similar rates with or without harmlessness training on our curriculum.
Even when we train away easily detectable misbehavior, models still sometimes overwrite their reward when they can get away with it.
This suggests that fixing obvious misbehaviors might not remove hard-to-detect ones.
Our work provides empirical evidence that serious misalignment can emerge from seemingly benign reward misspecification.
Read the full paper: https://arxiv.org/abs/2406.10162
The Anthropic Alignment Science team is actively hiring research engineers and scientists. We'd love to see your application: https://boards.greenhouse.io/anthropic/jobs/4009165008
Blog post:
Perverse incentives are everywhere. Thi...
392 قسمت
همه قسمت ها
×
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: The Obliqueness Thesis, published by Jessica Taylor on September 19, 2024 on The AI Alignment Forum. In my Xenosystems review, I discussed the Orthogonality Thesis, concluding that it was a bad metaphor. It's a long post, though, and the comments on orthogonality build on other Xenosystems content. Therefore, I think it may be helpful to present a more concentrated discussion on Orthogonality, contrasting Orthogonality with my own view, without introducing dependencies on Land's views. (Land gets credit for inspiring many of these thoughts, of course, but I'm presenting my views as my own here.) First, let's define the Orthogonality Thesis. Quoting Superintelligence for Bostrom's formulation: Intelligence and final goals are orthogonal: more or less any level of intelligence could in principle be combined with more or less any final goal. To me, the main ambiguity about what this is saying is the "could in principle" part; maybe, for any level of intelligence and any final goal, there exists (in the mathematical sense) an agent combining those, but some combinations are much more natural and statistically likely than others. Let's consider Yudkowsky's formulations as alternatives. Quoting Arbital: The Orthogonality Thesis asserts that there can exist arbitrarily intelligent agents pursuing any kind of goal. The strong form of the Orthogonality Thesis says that there's no extra difficulty or complication in the existence of an intelligent agent that pursues a goal, above and beyond the computational tractability of that goal. As an example of the computational tractability consideration, sufficiently complex goals may only be well-represented by sufficiently intelligent agents. "Complication" may be reflected in, for example, code complexity; to my mind, the strong form implies that the code complexity of an agent with a given level of intelligence and goals is approximately the code complexity of the intelligence plus the code complexity of the goal specification, plus a constant. Code complexity would influence statistical likelihood for the usual Kolmogorov/Solomonoff reasons, of course. I think, overall, it is more productive to examine Yudkowsky's formulation than Bostrom's, as he has already helpfully factored the thesis into weak and strong forms. Therefore, by criticizing Yudkowsky's formulations, I am less likely to be criticizing a strawman. I will use "Weak Orthogonality" to refer to Yudkowsky's "Orthogonality Thesis" and "Strong Orthogonality" to refer to Yudkowsky's "strong form of the Orthogonality Thesis". Land, alternatively, describes a "diagonal" between intelligence and goals as an alternative to orthogonality, but I don't see a specific formulation of a "Diagonality Thesis" on his part. Here's a possible formulation: Diagonality Thesis: Final goals tend to converge to a point as intelligence increases. The main criticism of this thesis is that formulations of ideal agency, in the form of Bayesianism and VNM utility, leave open free parameters, e.g. priors over un-testable propositions, and the utility function. Since I expect few readers to accept the Diagonality Thesis, I will not concentrate on criticizing it. What about my own view? I like Tsvi's naming of it as an "obliqueness thesis". Obliqueness Thesis: The Diagonality Thesis and the Strong Orthogonality Thesis are false. Agents do not tend to factorize into an Orthogonal value-like component and a Diagonal belief-like component; rather, there are Oblique components that do not factorize neatly. (Here, by Orthogonal I mean basically independent of intelligence, and by Diagonal I mean converging to a point in the limit of intelligence.) While I will address Yudkowsky's arguments for the Orthogonality Thesis, I think arguing directly for my view first will be more helpful. In general, it seems ...…
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Secret Collusion: Will We Know When to Unplug AI?, published by schroederdewitt on September 16, 2024 on The AI Alignment Forum. TL;DR: We introduce the first comprehensive theoretical framework for understanding and mitigating secret collusion among advanced AI agents, along with CASE, a novel model evaluation framework. CASE assesses the cryptographic and steganographic capabilities of agents, while exploring the emergence of secret collusion in real-world-like multi-agent settings. Whereas current AI models aren't yet proficient in advanced steganography, our findings show rapid improvements in individual and collective model capabilities, posing unprecedented safety and security risks. These results highlight urgent challenges for AI governance and policy, urging institutions such as the EU AI Office and AI safety bodies in the UK and US to prioritize cryptographic and steganographic evaluations of frontier models. Our research also opens up critical new pathways for research within the AI Control framework. Philanthropist and former Google CEO Eric Schmidt said in 2023 at a Harvard event: "[...] the computers are going to start talking to each other probably in a language that we can't understand and collectively their super intelligence - that's the term we use in the industry - is going to rise very rapidly and my retort to that is: do you know what we're going to do in that scenario? We're going to unplug them [...] But what if we cannot unplug them in time because we won't be able to detect the moment when this happens? In this blog post, we, for the first time, provide a comprehensive overview of the phenomenon of secret collusion among AI agents, connect it to foundational concepts in steganography, information theory, distributed systems theory, and computability, and present a model evaluation framework and empirical results as a foundation of future frontier model evaluations. This blog post summarises a large body of work. First of all, it contains our pre-print from February 2024 (updated in September 2024) "Secret Collusion among Generative AI Agents". An early form of this pre-print was presented at the 2023 New Orleans (NOLA) Alignment Workshop (see this recording NOLA 2023 Alignment Forum Talk Secret Collusion Among Generative AI Agents: a Model Evaluation Framework). Also, check out this long-form Foresight Institute Talk). In addition to these prior works, we also include new results. These contain empirical studies on the impact of paraphrasing as a mitigation tool against steganographic communications, as well as reflections on our findings' impact on AI Control. Multi-Agent Safety and Security in the Age of Autonomous Internet Agents The near future could see myriads of LLM-driven AI agents roam the internet, whether on social media platforms, eCommerce marketplaces, or blockchains. Given advances in predictive capabilities, these agents are likely to engage in increasingly complex intentional and unintentional interactions, ranging from traditional distributed systems pathologies (think dreaded deadlocks!) to more complex coordinated feedback loops. Such a scenario induces a variety of multi-agent safety, and specifically, multi-agent security[1] (see our NeurIPS'23 workshop Multi-Agent Security: Security as Key to AI Safety) concerns related to data exfiltration, multi-agent deception, and, fundamentally, undermining trust in AI systems. There are several real-world scenarios where agents could have access to sensitive information, such as their principals' preferences, which they may disclose unsafely even if they are safety-aligned when considered in isolation. Stray incentives, intentional or otherwise, or more broadly, optimization pressures, could cause agents to interact in undesirable and potentially dangerous ways. For example, join...…
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Estimating Tail Risk in Neural Networks, published by Jacob Hilton on September 13, 2024 on The AI Alignment Forum. Machine learning systems are typically trained to maximize average-case performance. However, this method of training can fail to meaningfully control the probability of tail events that might cause significant harm. For instance, while an artificial intelligence (AI) assistant may be generally safe, it would be catastrophic if it ever suggested an action that resulted in unnecessary large-scale harm. Current techniques for estimating the probability of tail events are based on finding inputs on which an AI behaves catastrophically. Since the input space is so large, it might be prohibitive to search through it thoroughly enough to detect all potential catastrophic behavior. As a result, these techniques cannot be used to produce AI systems that we are confident will never behave catastrophically. We are excited about techniques to estimate the probability of tail events that do not rely on finding inputs on which an AI behaves badly, and can thus detect a broader range of catastrophic behavior. We think developing such techniques is an exciting problem to work on to reduce the risk posed by advanced AI systems: Estimating tail risk is a conceptually straightforward problem with relatively objective success criteria; we are predicting something mathematically well-defined, unlike instances of eliciting latent knowledge (ELK) where we are predicting an informal concept like "diamond". Improved methods for estimating tail risk could reduce risk from a variety of sources, including central misalignment risks like deceptive alignment. Improvements to current methods can be found both by doing empirical research, or by thinking about the problem from a theoretical angle. This document will discuss the problem of estimating the probability of tail events and explore estimation strategies that do not rely on finding inputs on which an AI behaves badly. In particular, we will: Introduce a toy scenario about an AI engineering assistant for which we want to estimate the probability of a catastrophic tail event. Explain some deficiencies of adversarial training, the most common method for reducing risk in contemporary AI systems. Discuss deceptive alignment as a particularly dangerous case in which adversarial training might fail. Present methods for estimating the probability of tail events in neural network behavior that do not rely on evaluating behavior on concrete inputs. Conclude with a discussion of why we are excited about work aimed at improving estimates of the probability of tail events. This document describes joint research done with Jacob Hilton, Victor Lecomte, David Matolcsi, Eric Neyman, Thomas Read, George Robinson, and Gabe Wu. Thanks additionally to Ajeya Cotra, Lukas Finnveden, and Erik Jenner for helpful comments and suggestions. A Toy Scenario Consider a powerful AI engineering assistant. Write M for this AI system, and M(x) for the action it suggests given some project description x. We want to use this system to help with various engineering projects, but would like it to never suggest an action that results in large-scale harm, e.g. creating a doomsday device. In general, we define a behavior as catastrophic if it must never occur in the real world.[1] An input is catastrophic if it would lead to catastrophic behavior. Assume we can construct a catastrophe detector C that tells us if an action M(x) will result in large-scale harm. For the purposes of this example, we will assume both that C has a reasonable chance of catching all catastrophes and that it is feasible to find a useful engineering assistant M that never triggers C (see Catastrophe Detectors for further discussion). We will also assume we can use C to train M, but that it is ...…
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Can startups be impactful in AI safety?, published by Esben Kran on September 13, 2024 on The AI Alignment Forum. With Lakera's strides in securing LLM APIs, Goodfire AI's path to scaling interpretability, and 20+ model evaluations startups among much else, there's a rising number of technical startups attempting to secure the model ecosystem. Of course, they have varying levels of impact on superintelligence containment and security and even with these companies, there's a lot of potential for aligned, ambitious and high-impact startups within the ecosystem. This point isn't new and has been made in our previous posts and by Eric Ho (Goodfire AI CEO). To set the stage, our belief is that these are the types of companies that will have a positive impact: Startups with a profit incentive completely aligned with improving AI safety; that have a deep technical background to shape AGI deployment and; do not try to compete with AGI labs. Piloting AI safety startups To understand impactful technical AI safety startups better, Apart Research joined forces with collaborators from Juniper Ventures, vectorview (alumni from the latest YC cohort), Rudolf (from the upcoming def/acc cohort), Tangentic AI, and others. We then invited researchers, engineers, and students to resolve a key question "can we come up with ideas that scale AI safety into impactful for-profits?" The hackathon took place during a weekend two weeks ago with a keynote by Esben Kran (co-director of Apart) along with 'HackTalks' by Rudolf Laine (def/acc) and Lukas Petersson (YC / vectorview). Individual submissions were a 4 page report with the problem statement, why this solution will work, what the key risks of said solution are, and any experiments or demonstrations of the solution the team made. This post details the top 6 projects and excludes 2 projects that were made private by request (hopefully turning into impactful startups now!). In total, we had 101 signups and 11 final entries. Winners were decided by an LME model conditioned on reviewer bias. Watch the authors' lightning talks here. Dark Forest: Making the web more trustworthy with third-party content verification By Mustafa Yasir (AI for Cyber Defense Research Centre, Alan Turing Institute) Abstract: 'DarkForest is a pioneering Human Content Verification System (HCVS) designed to safeguard the authenticity of online spaces in the face of increasing AI-generated content. By leveraging graph-based reinforcement learning and blockchain technology, DarkForest proposes a novel approach to safeguarding the authentic and humane web. We aim to become the vanguard in the arms race between AI-generated content and human-centric online spaces.' Content verification workflow supported by graph-based RL agents deciding verifications Reviewer comments: Natalia: Well explained problem with clear need addressed. I love that you included the content creation process - although you don't explicitly address how you would attract content creators to use your platform over others in their process. Perhaps exploring what features of platforms drive creators to each might help you make a compelling case for using yours beyond the verification capabilities. I would have also liked to see more details on how the verification decision is made and how accurate this is on existing datasets. Nick: There's a lot of valuable stuff in here regarding content moderation and identity verification. I'd narrow it to one problem-solution pair (e.g., "jobs to be done") and focus more on risks around early product validation (deep interviews with a range of potential users and buyers regarding value) and go-to-market. It might also be worth checking out Musubi. Read the full project here. Simulation Operators: An annotation operation for alignment of robot By Ardy Haroen (USC) Abstrac...…
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: How difficult is AI Alignment?, published by Samuel Dylan Martin on September 13, 2024 on The AI Alignment Forum. This work was funded by Polaris Ventures There is currently no consensus on how difficult the AI alignment problem is. We have yet to encounter any real-world, in the wild instances of the most concerning threat models, like deceptive misalignment. However, there are compelling theoretical arguments which suggest these failures will arise eventually. Will current alignment methods accidentally train deceptive, power-seeking AIs that appear aligned, or not? We must make decisions about which techniques to avoid and which are safe despite not having a clear answer to this question. To this end, a year ago, we introduced the AI alignment difficulty scale, a framework for understanding the increasing challenges of aligning artificial intelligence systems with human values. This follow-up article revisits our original scale, exploring how our understanding of alignment difficulty has evolved and what new insights we've gained. This article will explore three main themes that have emerged as central to our understanding: 1. The Escalation of Alignment Challenges: We'll examine how alignment difficulties increase as we go up the scale, from simple reward hacking to complex scenarios involving deception and gradient hacking. Through concrete examples, we'll illustrate these shifting challenges and why they demand increasingly advanced solutions. These examples will illustrate what observations we should expect to see "in the wild" at different levels, which might change our minds about how easy or difficult alignment is. 2. Dynamics Across the Difficulty Spectrum: We'll explore the factors that change as we progress up the scale, including the increasing difficulty of verifying alignment, the growing disconnect between alignment and capabilities research, and the critical question of which research efforts are net positive or negative in light of these challenges. 3. Defining and Measuring Alignment Difficulty: We'll tackle the complex task of precisely defining "alignment difficulty," breaking down the technical, practical, and other factors that contribute to the alignment problem. This analysis will help us better understand the nature of the problem we're trying to solve and what factors contribute to it. The Scale The high level of the alignment problem, provided in the previous post, was: "The alignment problem" is the problem of aligning sufficiently powerful AI systems, such that we can be confident they will be able to reduce the risks posed by misused or unaligned AI systems We previously introduced the AI alignment difficulty scale, with 10 levels that map out the increasing challenges. The scale ranges from "alignment by default" to theoretical impossibility, with each level representing more complex scenarios requiring more advanced solutions. It is reproduced here: Alignment Difficulty Scale Difficulty Level Alignment technique X is sufficient Description Key Sources of risk 1 (Strong) Alignment by Default As we scale up AI models without instructing or training them for specific risky behaviour or imposing problematic and clearly bad goals (like 'unconditionally make money'), they do not pose significant risks. Even superhuman systems basically do the commonsense version of what external rewards (if RL) or language instructions (if LLM) imply. Misuse and/or recklessness with training objectives. RL of powerful models towards badly specified or antisocial objectives is still possible, including accidentally through poor oversight, recklessness or structural factors. 2 Reinforcement Learning from Human Feedback We need to ensure that the AI behaves well even in edge cases by guiding it more carefully using human feedback in a wide range of situations...…
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Contra papers claiming superhuman AI forecasting, published by nikos on September 12, 2024 on The AI Alignment Forum. [Conflict of interest disclaimer: We are FutureSearch, a company working on AI-powered forecasting and other types of quantitative reasoning. If thin LLM wrappers could achieve superhuman forecasting performance, this would obsolete a lot of our work.] Widespread, misleading claims about AI forecasting Recently we have seen a number of papers - (Schoenegger et al., 2024, Halawi et al., 2024, Phan et al., 2024, Hsieh et al., 2024) - with claims that boil down to "we built an LLM-powered forecaster that rivals human forecasters or even shows superhuman performance". These papers do not communicate their results carefully enough, shaping public perception in inaccurate and misleading ways. Some examples of public discourse: Ethan Mollick (>200k followers) tweeted the following about the paper Wisdom of the Silicon Crowd: LLM Ensemble Prediction Capabilities Rival Human Crowd Accuracy by Schoenegger et al.: A post on Marginal Revolution with the title and abstract of the paper Approaching Human-Level Forecasting with Language Models by Halawi et al. elicits responses like "This is something that humans are notably terrible at, even if they're paid to do it. No surprise that LLMs can match us." "+1 The aggregate human success rate is a pretty low bar" A Twitter thread with >500k views on LLMs Are Superhuman Forecasters by Phan et al. claiming that "AI […] can predict the future at a superhuman level" had more than half a million views within two days of being published. The number of such papers on AI forecasting, and the vast amount of traffic on misleading claims, makes AI forecasting a uniquely misunderstood area of AI progress. And it's one that matters. What does human-level or superhuman forecasting mean? "Human-level" or "superhuman" is a hard-to-define concept. In an academic context, we need to work with a reasonable operationalization to compare the skill of an AI forecaster with that of humans. One reasonable and practical definition of a superhuman forecasting AI forecaster is The AI forecaster is able to consistently outperform the crowd forecast on a sufficiently large number of randomly selected questions on a high-quality forecasting platform.[1] (For a human-level forecaster, just replace "outperform" with "performs on par with".) Red flags for claims to (super)human AI forecasting accuracy Our experience suggests there are a number of things that can go wrong when building AI forecasting systems, including: 1. Failing to find up-to-date information on the questions. It's inconceivable on most questions that forecasts can be good without basic information. Imagine trying to forecast the US presidential election without knowing that Biden dropped out. 2. Drawing on up-to-date, but low-quality information. Ample experience shows low quality information confuses LLMs even more than it confuses humans. Imagine forecasting election outcomes with biased polling data. Or, worse, imagine forecasting OpenAI revenue based on claims like > The number of ChatGPT Plus subscribers is estimated between 230,000-250,000 as of October 2023. without realising that this mixing up ChatGPT vs ChatGPT mobile. 3. Lack of high-quality quantitative reasoning. For a decent number of questions on Metaculus, good forecasts can be "vibed" by skilled humans and perhaps LLMs. But for many questions, simple calculations are likely essential. Human performance shows systematic accuracy nearly always requires simple models such as base rates, time-series extrapolations, and domain-specific numbers. Imagine forecasting stock prices without having, and using, historical volatility. 4. Retrospective, rather than prospective, forecasting (e.g. forecasting questions that have al...…
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: AI forecasting bots incoming, published by Dan H on September 9, 2024 on The AI Alignment Forum. In a recent appearance on Conversations with Tyler, famed political forecaster Nate Silver expressed skepticism about AIs replacing human forecasters in the near future. When asked how long it might take for AIs to reach superhuman forecasting abilities, Silver replied: "15 or 20 [years]." In light of this, we are excited to announce "FiveThirtyNine," an AI forecasting bot. Our bot, built on GPT-4o, provides probabilities for any user-entered query, including " Will Trump win the 2024 presidential election?" and " Will China invade Taiwan by 2030?" Our bot performs better than experienced human forecasters and performs roughly the same as (and sometimes even better than) crowds of experienced forecasters; since crowds are for the most part superhuman, FiveThirtyNine is in a similar sense. (We discuss limitations later in this post.) Our bot and other forecasting bots can be used in a wide variety of contexts. For example, these AIs could help policymakers minimize bias in their decision-making or help improve global epistemics and institutional decision-making by providing trustworthy, calibrated forecasts. We hope that forecasting bots like ours will be quickly integrated into frontier AI models. For now, we will keep our bot available at forecast.safe.ai, where users are free to experiment and test its capabilities. Quick Links Demo: forecast.safe.ai Technical Report: link Problem Policymakers at the highest echelons of government and corporate power have difficulty making high-quality decisions on complicated topics. As the world grows increasingly complex, even coming to a consensus agreement on basic facts is becoming more challenging, as it can be hard to absorb all the relevant information or know which sources to trust. Separately, online discourse could be greatly improved. Discussions on uncertain, contentious issues all too often devolve into battles between interest groups, each intent on name-calling and spouting the most extreme versions of their views through highly biased op-eds and tweets. FiveThirtyNine Before transitioning to how forecasting bots like FiveThirtyNine can help improve epistemics, it might be helpful to give a summary of what FiveThirtyNine is and how it works. FiveThirtyNine can be given a query - for example, "Will Trump win the 2024 US presidential election?" FiveThirtyNine is prompted to behave like an "AI that is superhuman at forecasting". It is then asked to make a series of search engine queries for news and opinion articles that might contribute to its prediction. (The following example from FiveThirtyNine uses GPT-4o as the base LLM.) Based on these sources and its wealth of prior knowledge, FiveThirtyNine compiles a summary of key facts. Given these facts, it's asked to give reasons for and against Trump winning the election, before weighing each reason based on its strength and salience. Finally, FiveThirtyNine aggregates its considerations while adjusting for negativity and sensationalism bias in news sources and outputs a tentative probability. It is asked to sanity check this probability and adjust it up or down based on further reasoning, before putting out a final, calibrated probability - in this case, 52%. Evaluation. To test how well our bot performs, we evaluated it on questions from the Metaculus forecasting platform. We restricted the bot to make predictions only using the information human forecasters had, ensuring a valid comparison. Specifically, GPT-4o is only trained on data up to October 2023, and we restricted the news and opinion articles it could access to only those published before a certain date. From there, we asked it to compute the probabilities of 177 events from Metaculus that had happened (or not ha...…
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Backdoors as an analogy for deceptive alignment, published by Jacob Hilton on September 6, 2024 on The AI Alignment Forum. ARC has released a paper on Backdoor defense, learnability and obfuscation in which we study a formal notion of backdoors in ML models. Part of our motivation for this is an analogy between backdoors and deceptive alignment, the possibility that an AI system would intentionally behave well in training in order to give itself the opportunity to behave uncooperatively later. In our paper, we prove several theoretical results that shed some light on possible mitigations for deceptive alignment, albeit in a way that is limited by the strength of this analogy. In this post, we will: Lay out the analogy between backdoors and deceptive alignment Discuss prior theoretical results from the perspective of this analogy Explain our formal notion of backdoors and its strengths and weaknesses Summarize the results in our paper and discuss their implications for deceptive alignment Thanks to Boaz Barak, Roger Grosse, Thomas Read, John Schulman and Gabriel Wu for helpful comments. Backdoors and deceptive alignment A backdoor in an ML model is a modification to the model that causes it to behave differently on certain inputs that activate a secret "trigger", while behaving similarly on ordinary inputs. There is a wide existing literature on backdoor attacks and defenses, which is primarily empirical, but also includes some theoretical results that we will mention. Deceptive alignment is a term from the paper Risks from Learned Optimization in Advanced Machine Learning Systems (Section 4) that refers to the possibility that an AI system will internally reason about the objective that it is being trained on, and decide to perform well according to that objective unless there are clues that it has been taken out of its training environment. Such a policy could be optimal on the training distribution, and yet perform very badly on certain out-of-distribution inputs where such clues are present, which we call defection triggers.[1] The opposite of deceptive alignment is robust alignment, meaning that this performance degradation is avoided. Since a deceptively aligned model and a robustly aligned model behave very differently on defection triggers, but very similarly on typical inputs from the training distribution, deceptive alignment can be thought of as a special kind of backdoor, under the following correspondence: Deceptive alignment Backdoors Robustly aligned model Original (unmodified) model Deceptively aligned model Backdoored model Defection trigger Backdoor trigger The main distinguishing feature of deceptive alignment compared to other kinds of backdoors is that the deceptively aligned model is not produced by an adversary, but is instead produced through ordinary training. Thus by treating deceptive alignment as a backdoor, we are modeling the training process as an adversary. In our analysis of deceptive alignment, the basic tension we will face is that an unconstrained adversary will always win, but any particular proxy constraint we impose on the adversary may be unrealistic. Static backdoor detection An important piece of prior work is the paper Planting Undetectable Backdoors in Machine Learning Models, which uses a digital signature scheme to insert an undetectable backdoor into a model. Roughly speaking, the authors exhibit a modified version of a "Random Fourier Features" training algorithm that produces a backdoored model. Any input to the backdoored model can be perturbed by an attacker with knowledge of a secret key to produce a new input on which the model behaves differently. However, the backdoor is undetectable in the sense that it is computationally infeasible for a defender with white-box access to distinguish a backdoored model from an or...…
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Conflating value alignment and intent alignment is causing confusion, published by Seth Herd on September 5, 2024 on The AI Alignment Forum. Submitted to the Alignment Forum. Contains more technical jargon than usual. Epistemic status: I think something like this confusion is happening often. I'm not saying these are the only differences in what people mean by "AGI alignment". Summary: Value alignment is better but probably harder to achieve than personal intent alignment to the short-term wants of some person(s). Different groups and people tend to primarily address one of these alignment targets when they discuss alignment. Confusion abounds. One important confusion stems from an assumption that the type of AI defines the alignment target: strong goal-directed AGI must be value aligned or misaligned, while personal intent alignment is only viable for relatively weak AI. I think this assumption is important but false. While value alignment is categorically better, intent alignment seems easier, safer, and more appealing in the short term, so AGI project leaders are likely to try it.[1] Overview Clarifying what people mean by alignment should dispel some illusory disagreement, and clarify alignment theory and predictions of AGI outcomes. Caption: Venn diagram of three types of alignment targets. Value alignment and Personal intent alignment are both subsets of Evan Hubinger's definition of intent alignment: AGI aligned with human intent in the broadest sense. Prosaic alignment work usually seems to be addressing a target somewhere in the neighborhood of personal intent alignment (following instructions or doing what this person wants now), while agent foundations and other conceptual alignment work usually seems to be addressing value alignment. Those two clusters have different strengths and weaknesses as alignment targets, so lumping them together produces confusion. People mean different things when they say alignment. Some are mostly thinking about value alignment (VA): creating sovereign AGI that has values close enough to humans' for our liking. Others are talking about making AGI that is corrigible (in the Christiano or Harms sense)[2] or follows instructions from its designated principal human(s). I'm going to use the term personal intent alignment (PIA) until someone has a better term for that type of alignment target. Different arguments and intuitions apply to these two alignment goals, so talking about them without differentiation is creating illusory disagreements. Value alignment is better almost by definition, but personal intent alignment seems to avoid some of the biggest difficulties of value alignment. Max Harms' recent sequence on corrigibility as a singular target (CAST) gives both a nice summary and detailed arguments. We do not need us to point to or define values, just short term preferences or instructions. The principal advantage is that an AGI that follows instructions can be used as a collaborator in improving its alignment over time; you don't need to get it exactly right on the first try. This is more helpful in slower and more continuous takeoffs. This means that PI alignment has a larger basin of attraction than value alignment does.[3] Most people who think alignment is fairly achievable seem to be thinking of PIA, while critics often respond thinking of value alignment. It would help to be explicit. PIA is probably easier and more likely than full VA for our first stabs at AGI, but there are reasons to wonder if it's adequate for real success. In particular, there are intuitions and arguments that PIA doesn't address the real problem of AGI alignment. I think PIA does address the real problem, but in a non-obvious and counterintuitive way. Another unstated divide There's another important clustering around these two conceptions of al...…
1 AF - Is there any rigorous work on using anthropic uncertainty to prevent situational awareness / deception? by David Scott Krueger 1:01
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Is there any rigorous work on using anthropic uncertainty to prevent situational awareness / deception?, published by David Scott Krueger on September 4, 2024 on The AI Alignment Forum. AI systems up to some high level of intelligence plausibly need to know exactly where they are in space-time in order for deception/"scheming" to make sense as a strategy. This is because they need to know: 1) what sort of oversight they are subject to and 2) what effects their actions will have on the real world (side note: Acausal trade might break this argument) There are a number of informal proposals to keep AI systems selectively ignorant of (1) and (2) in order to prevent deception. Those proposals seem very promising to flesh out; I'm not aware of any rigorous work doing so, however. Are you? Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.…
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: The Checklist: What Succeeding at AI Safety Will Involve, published by Sam Bowman on September 3, 2024 on The AI Alignment Forum. Crossposted by habryka with Sam's permission. Expect lower probability for Sam to respond to comments here than if he had posted it. Preface This piece reflects my current best guess at the major goals that Anthropic (or another similarly positioned AI developer) will need to accomplish to have things go well with the development of broadly superhuman AI. Given my role and background, it's disproportionately focused on technical research and on averting emerging catastrophic risks. For context, I lead a technical AI safety research group at Anthropic, and that group has a pretty broad and long-term mandate, so I spend a lot of time thinking about what kind of safety work we'll need over the coming years. This piece is my own opinionated take on that question, though it draws very heavily on discussions with colleagues across the organization: Medium- and long-term AI safety strategy is the subject of countless leadership discussions and Google docs and lunch-table discussions within the organization, and this piece is a snapshot (shared with permission) of where those conversations sometimes go. To be abundantly clear: Nothing here is a firm commitment on behalf of Anthropic, and most people at Anthropic would disagree with at least a few major points here, but this can hopefully still shed some light on the kind of thinking that motivates our work. Here are some of the assumptions that the piece relies on. I don't think any one of these is a certainty, but all of them are plausible enough to be worth taking seriously when making plans: Broadly human-level AI is possible. I'll often refer to this as transformative AI (or TAI), roughly defined as AI that could form as a drop-in replacement for humans in all remote-work-friendly jobs, including AI R&D.[1] Broadly human-level AI (or TAI) isn't an upper bound on most AI capabilities that matter, and substantially superhuman systems could have an even greater impact on the world along many dimensions. If TAI is possible, it will probably be developed this decade, in a business and policy and cultural context that's not wildly different from today. If TAI is possible, it could be used to dramatically accelerate AI R&D, potentially leading to the development of substantially superhuman systems within just a few months or years after TAI. Powerful AI systems could be extraordinarily destructive if deployed carelessly, both because of new emerging risks and because of existing issues that become much more acute. This could be through misuse of weapons-related capabilities, by disrupting important balances of power in domains like cybersecurity or surveillance, or by any of a number of other means. Many systems at TAI and beyond, at least under the right circumstances, will be capable of operating more-or-less autonomously for long stretches in pursuit of big-picture, real-world goals. This magnifies these safety challenges. Alignment - in the narrow sense of making sure AI developers can confidently steer the behavior of the AI systems they deploy - requires some non-trivial effort to get right, and it gets harder as systems get more powerful. Most of the ideas here ultimately come from outside Anthropic, and while I cite a few sources below, I've been influenced by far more writings and people than I can credit here or even keep track of. Introducing the Checklist This lays out what I think we need to do, divided into three chapters, based on the capabilities of our strongest models: Chapter 1: Preparation You are here. In this period, our best models aren't yet TAI. In the language of Anthropic's RSP, they're at AI Safety Level 2 (ASL-2), ASL-3, or maybe the early stages of ASL-4. Most of the wor...…
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Survey: How Do Elite Chinese Students Feel About the Risks of AI?, published by Nick Corvino on September 2, 2024 on The AI Alignment Forum. Intro In April 2024, my colleague and I (both affiliated with Peking University) conducted a survey involving 510 students from Tsinghua University and 518 students from Peking University - China's two top academic institutions. Our focus was on their perspectives regarding the frontier risks of artificial intelligence. In the People's Republic of China (PRC), publicly accessible survey data on AI is relatively rare, so we hope this report provides some valuable insights into how people in the PRC are thinking about AI (especially the risks). Throughout this post, I'll do my best to weave in other data reflecting the broader Chinese sentiment toward AI. For similar research, check out The Center for Long-Term Artificial Intelligence, YouGov, Monmouth University, The Artificial Intelligence Policy Institute, and notably, a poll conducted by Rethink Priorities, which closely informed our survey design. You can read the full report published in the Jamestown Foundation's China Brief here: Survey: How Do Elite Chinese Students Feel About the Risks of AI? Key Takeaways Students are more optimistic about the benefits of AI than concerned about the harms. 80 percent of respondents agreed or strongly agreed with the statement that AI will do more good than harm for society, with only 7.5 percent actively believing the harms could outweigh the benefits. This, similar to other polling, indicates that the PRC is one of the most optimistic countries concerning the development of AI. Students strongly believe the Chinese government should regulate AI. 85.31 percent of respondents believe AI should be regulated by the government, with only 6 percent actively believing it should not. This contrasts with trends seen in other countries, where there is typically a positive correlation between optimism about AI and calls for minimizing regulation. The strong support for regulation in the PRC, even as optimism about AI remains high, suggests a distinct perspective on the role of government oversight in the PRC context. Students ranked AI the lowest among all possible existential threats to humanity. When asked about the most likely causes of human extinction, misaligned artificial intelligence received the lowest score. Nuclear war, natural disaster, climate change, and pandemics all proved more concerning for students. Students lean towards cooperation between the United States and the PRC as necessary for the safe and responsible development of AI. 60.7 percent of respondents believe AI will not be developed safely without cooperation between China and the U.S., with 25.68 percent believing it will develop safely no matter the level of cooperation. Students are most concerned about the use of AI for surveillance. This was followed by misinformation, existential risk, wealth inequality, increased political tension, various issues related to bias, with the suffering of artificial entities receiving the lowest score. Background As the recent decision (决定) document from the Third Plenum meetings in July made clear, AI is one of eight technologies that the Chinese Communist Party (CCP) leadership sees as critical for achieving "Chinese-style modernization (中国式现代化)," and is central to the strategy of centering the country's economic future around breakthroughs in frontier science ( People's Daily, July 22). The PRC also seeks to shape international norms on AI, including on AI risks. In October 2023, Xi Jinping announced a "Global AI Governance Initiative (全球人工智能治理倡议)" ( CAC, October 18, 2023). Tsinghua and Peking Universty are the two most prestigious universities in the PRC (by far), many of whose graduates will be very influential in shaping the cou...…
1 AF - Can a Bayesian Oracle Prevent Harm from an Agent? (Bengio et al. 2024) by Matt MacDermott 8:04
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Can a Bayesian Oracle Prevent Harm from an Agent? (Bengio et al. 2024), published by Matt MacDermott on September 1, 2024 on The AI Alignment Forum. Yoshua Bengio wrote a blogpost about a new AI safety paper by him, various collaborators, and me. I've pasted the text below, but first here are a few comments from me aimed at an AF/LW audience. The paper is basically maths plus some toy experiments. It assumes access to a Bayesian oracle that can infer a posterior over hypotheses given data, and can also estimate probabilities for some negative outcome ("harm"). It proposes some conservative decision rules one could use to reject actions proposed by an agent, and proves probabilistic bounds on their performance under appropriate assumptions. I expect the median reaction in these parts to be something like: ok, I'm sure there are various conservative decision rules you could apply using a Bayesian oracle, but isn't obtaining a Bayesian oracle the hard part here? Doesn't that involve advances in Bayesian machine learning, and also probably solving ELK to get the harm estimates? My answer to that is: yes, I think so. I think Yoshua does too, and that that's the centre of his research agenda. Probably the main interest of this paper to people here is to provide an update on Yoshua's research plans. In particular it gives some more context on what the "guaranteed safe AI" part of his approach might look like -- design your system to do explicit Bayesian inference, and make an argument that the system is safe based on probabilistic guarantees about the behaviour of a Bayesian inference machine. This is in contrast to more hardcore approaches that want to do formal verification by model-checking. You should probably think of the ambition here as more like "a safety case involving proofs" than "a formal proof of safety". Bounding the probability of harm from an AI to create a guardrail Published 29 August 2024 by yoshuabengio As we move towards more powerful AI, it becomes urgent to better understand the risks, ideally in a mathematically rigorous and quantifiable way, and use that knowledge to mitigate them. Is there a way to design powerful AI systems based on machine learning methods that would satisfy probabilistic safety guarantees, i.e., would be provably unlikely to take a harmful action? Current AI safety evaluations and benchmarks test the AI for cases where it may behave badly, e.g., by providing answers that could yield dangerous misuse. That is useful and should be legally required with flexible regulation, but is not sufficient. These tests only tell us one side of the story: If they detect bad behavior, a flag is raised and we know that something must be done to mitigate the risks. However, if they do not raise such a red flag, we may still have a dangerous AI in our hands, especially since the testing conditions might be different from the deployment setting, and attackers (or an out-of-control AI) may be creative in ways that the tests did not consider. Most concerningly, AI systems could simply recognize they are being tested and have a temporary incentive to behave appropriately while being tested. Part of the problem is that such tests are spot checks. They are trying to evaluate the risk associated with the AI in general by testing it on special cases. Another option would be to evaluate the risk on a case-by-case basis and reject queries or answers that are considered to potentially violate or safety specification. With the long-term goal of obtaining a probabilistic guarantee that would apply in every context, we thus consider in this new paper (see reference and co-authors below) the objective of estimating a context-dependent upper bound on the probability of violating a given safety specification. Such a risk evaluation would need to be performed at ru...…
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Epistemic states as a potential benign prior, published by Tamsin Leake on August 31, 2024 on The AI Alignment Forum. Malignancy in the prior seems like a strong crux of the goal-design part of alignment to me. Whether your prior is going to be used to model: processes in the multiverse containing a specific "beacon" bitstring, processes in the multiverse containing the AI, processes which would output all of my blog, so I can make it output more for me, processes which match an AI chatbot's hypotheses about what it's talking with, then you have to sample hypotheses from somewhere; and typically, we want to use either solomonoff induction or time-penalized versions of it such as levin search (penalized by log of runtime) or what QACI uses (penalized by runtime, but with quantum computation available in some cases), or the implicit prior of neural networks (large sequences of multiplying by a matrix, adding a vector, and ReLU, often with a penalty related to how many non-zero weights are used). And the solomonoff prior is famously malign. (Alternatively, you could have knightian uncertainty about parts of your prior that aren't nailed down enough, and then do maximin over your knightian uncertainty (like in infra-bayesianism), but then you're not guaranteed that your AI gets anywhere at all; its knightian uncertainty might remain so immense that the AI keeps picking the null action all the time because some of its knightian hypotheses still say that anything else is a bad idea. Note: I might be greatly misunderstanding knightian uncertainty!) (It does seem plausible that doing geometric expectation over hypotheses in the prior helps "smooth things over" in some way, but I don't think this particularly removes the weight of malign hypotheses in the prior? It just allocates their steering power in a different way, which might make things less bad, but it sounds difficult to quantify.) It does feel to me like we do want a prior for the AI to do expected value calculations over, either for prediction or for utility maximization (or quantilization or whatever). One helpful aspect of prior-distribution-design is that, in many cases, I don't think the prior needs to contain the true hypothesis. For example, if the problem that we're using a prior for is to model processes which match an AI chatbot's hypotheses about what it's talking with then we don't need the AI's prior to contain a process which behaves just like the human user it's interacting with; rather, we just need the AI's prior to contain a hypothesis which: is accurate enough to match observations. is accurate enough to capture the fact that the user (if we pick a good user) implements the kind of decision theory that lets us rely on them pointing back to the actual real physical user when they get empowered - i.e. in CEV(user-hypothesis), user-hypothesis builds and then runs CEV(physical-user), because that's what the user would do in such a situation. Let's call this second criterion "cooperating back to the real user". So we need a prior which: Has at least some mass on hypotheses which correspond to observations cooperate back to the real user and can eventually be found by the AI, given enough evidence (enough chatting with the user) Call this the "aligned hypothesis". Before it narrows down hypothesis space to mostly just aligned hypotheses, doesn't give enough weight to demonic hypothesis which output whichever predictions cause the AI to brainhack its physical user, or escape using rowhammer-type hardware vulnerabilities, or other failures like that. Formalizing the chatbot model First, I'll formalize this chatbot model. Let's say we have a magical inner-aligned "soft" math-oracle: Which, given a "scoring" mathematical function from a non-empty set a to real numbers (not necessarily one that is tractably ...…
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: AIS terminology proposal: standardize terms for probability ranges, published by Egg Syntax on August 30, 2024 on The AI Alignment Forum. Summary: The AI safety research community should adopt standardized terms for probability ranges, especially in public-facing communication and especially when discussing risk estimates. The terms used by the IPCC are a reasonable default. Science communication is notoriously hard. It's hard for a lot of reasons, but one is that laypeople aren't used to thinking in numerical probabilities or probability ranges. One field that's had to deal with this more than most is climatology; climate change has been rather controversial, and a non-trivial aspect of that has been lay confusion about what climatologists are actually saying[1]. As a result, the well-known climate assessment reports from the UN's Intergovernmental Panel on Climate Change (IPCC) have, since the 1990s, used explicitly defined terms for probability ranges[2]: (see below for full figure[3]) Like climatology, AI safety research has become a topic of controversy. In both cases, the controversy includes a mix of genuine scientific disagreement, good-faith confusion, and bad-faith opposition. Scientific disagreement comes from people who can deal with numerical probability ranges. Those who are arguing in bad faith from ulterior motives generally don't care about factual details. But I suspect that the large majority of those who disagree, especially laypeople, are coming from a place of genuine, good-faith confusion. For those people, anything we as practitioners can do to communicate more clearly is quite valuable. Also like climatology, AI safety research, especially assessments of risk, fundamentally involves communicating about probabilities and probability ranges. Therefore I propose that the AIS community follow climatologists in adopting standard terms for probability ranges, especially in position papers and public-facing communication. In less formal and less public-facing contexts, using standard terminology still adds some value but is less important; in sufficiently informal contexts it's probably not worth the hassle of looking up the standard terminology. Of course, in many cases it's better to just give the actual numerical range! But especially in public-facing communication it can be more natural to use natural language terms, and in fact this is already often done. I'm only proposing that when we do use natural language terms for probability ranges, we use them in a consistent and interpretable way (feel free to link to this post as a reference for interpretation, or point to the climatology papers cited below[2]). Should the AIS community use the same terms? That's a slightly harder question. The obvious first-pass answer is 'yes'; it's a natural Schelling point, and terminological consistency across fields is generally preferable when practically possible. The IPCC terms also have the significant advantage of being battle-tested; they've been used over a thirty-year period in a highly controversial field, and terms have been refined when they were found to be insufficiently clear. The strongest argument I see against using the same terms is that the AIS community sometimes needs to deal with more extreme (high or low) risk estimates than these. If we use 'virtually certain' to mean 99 - 100%, what terms can we use for 99.9 - 100.0%, or 99.99 - 100.00%? On the other hand, plausibly once we're dealing with such extreme risk estimates, it's increasingly important to communicate them with actual numeric ranges. My initial proposal is to adopt the IPCC terms, but I'm very open to feedback, and if someone has an argument I find compelling (or which gets strong agreement in votes) for a different or extended set of terms, I'll add it to the proposal. If no su...…
T
The Nonlinear Library: Alignment Forum
1 AF - Solving adversarial attacks in computer vision as a baby version of general AI alignment by stanislavfort 12:34
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Solving adversarial attacks in computer vision as a baby version of general AI alignment, published by stanislavfort on August 29, 2024 on The AI Alignment Forum. I spent the last few months trying to tackle the problem of adversarial attacks in computer vision from the ground up. The results of this effort are written up in our new paper Ensemble everything everywhere: Multi-scale aggregation for adversarial robustness (explainer on X/Twitter). Taking inspiration from biology, we reached state-of-the-art or above state-of-the-art robustness at 100x - 1000x less compute, got human-understandable interpretability for free, turned classifiers into generators, and designed transferable adversarial attacks on closed-source (v)LLMs such as GPT-4 or Claude 3. I strongly believe that there is a compelling case for devoting serious attention to solving the problem of adversarial robustness in computer vision, and I try to draw an analogy to the alignment of general AI systems here. 1. Introduction In this post, I argue that the problem of adversarial attacks in computer vision is in many ways analogous to the larger task of general AI alignment. In both cases, we are trying to faithfully convey an implicit function locked within the human brain to a machine, and we do so extremely successfully on average. Under static evaluations, the human and machine functions match up exceptionally well. However, as is typical in high-dimensional spaces, some phenomena can be relatively rare and basically impossible to find by chance, yet ubiquitous in their absolute count. This is the case for adversarial attacks - imperceptible modifications to images that completely fool computer vision systems and yet have virtually no effect on humans. Their existence highlights a crucial and catastrophic mismatch between the implicit human vision function and the function learned by machines - a mismatch that can be exploited in a dynamic evaluation by an active, malicious agent. Such failure modes will likely be present in more general AI systems, and our inability to remedy them even in the more restricted vision context (yet) does not bode well for the broader alignment project. This is a call to action to solve the problem of adversarial vision attacks - a stepping stone on the path to aligning general AI systems. 2. Communicating implicit human functions to machines The basic goal of computer vision can be viewed as trying to endow a machine with the same vision capabilities a human has. A human carries, locked inside their skull, an implicit vision function mapping visual inputs into semantically meaningful symbols, e.g. a picture of a tortoise into a semantic label tortoise. This function is represented implicitly and while we are extremely good at using it, we do not have direct, conscious access to its inner workings and therefore cannot communicate it to others easily. To convey this function to a machine, we usually form a dataset of fixed images and their associated labels. We then use a general enough class of functions, typically deep neural networks, and a gradient-based learning algorithm together with backpropagation to teach the machine how to correlate images with their semantic content, e.g. how to assign a label parrot to a picture of a parrot. This process is extremely successful in communicating the implicit human vision function to the computer, and the implicit human and explicit, learned machine functions agree to a large extent. The agreement between the two is striking. Given how different the architectures are (a simulated graph-like function doing a single forward pass vs the wet protein brain of a mammal running continuous inference), how different the learning algorithms are (gradient descent with backpropagation vs something completely different but still unknown), a...…
T
The Nonlinear Library: Alignment Forum
1 AF - Would catching your AIs trying to escape convince AI developers to slow down or undeploy? by Buck Shlegeris 5:55
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Would catching your AIs trying to escape convince AI developers to slow down or undeploy?, published by Buck Shlegeris on August 26, 2024 on The AI Alignment Forum. I often talk to people who think that if frontier models were egregiously misaligned and powerful enough to pose an existential threat, you could get AI developers to slow down or undeploy models by producing evidence of their misalignment. I'm not so sure. As an extreme thought experiment, I'll argue this could be hard even if you caught your AI red-handed trying to escape. Imagine you're running an AI lab at the point where your AIs are able to automate almost all intellectual labor; the AIs are now mostly being deployed internally to do AI R&D. (If you want a concrete picture here, I'm imagining that there are 10 million parallel instances, running at 10x human speed, working 24/7. See e.g. similar calculations here). And suppose (as I think is 35% likely) that these models are egregiously misaligned and are actively looking for ways to subvert your safety measures and escape your control. You caught your AI attempting to backdoor some code that controls access to a security-critical service. You investigated this by running other instances on the inputs they'd see if that backdoored code had actually been deployed; the other AI instances never point out the security flaw, and they often try to exploit that code to escalate their privileges. (See here for more on this.) At this point you're almost sure the AI is scheming. So you try to call the US President and they hear you out, and then they go off and call one of your competitors to ask for their take. They say something like this: It would be extremely costly to shut down all the AI projects. China is racing to catch up and won't stop for this. If you force us to stop deploying our frontier models, they'll outcompete us within a year, and you won't be able to get China to pause without substantial risk of war. AIs are well known to do weird stuff. It would be irresponsible to assume one instance of anomalous behavior meant AIs were systematically going to act that way. I can put you in touch with top AI experts and they'll back me up. Even if the AI is indeed doing something systematically funny, we have no evidence that it has the ambition to seize huge amounts of power, and it's not clear that it would be able to. It seems implausible that AIs would do that; there's been no concrete evidence that the AIs are in fact power-hungry. Maybe our competitor just messed up their alignment, which would make sense because, as we've always told you, they're not very competent (especially compared to us). If they want to shut down, they're welcome to. But it doesn't make sense to penalize us for their shoddy work. Maybe they're just lying; this could be a doomer scheme to shut down AI, which the doomers have wanted to do for years. The logs they sent you could be complete forgeries. Or for all we know, someone there (with or without leadership approval) intentionally backdoored their RL data to make their model do this. It would be a mistake to take rash action before independent experts confirm that this wasn't somehow rigged. I'm sympathetic to all of these arguments. The main reason I'd be more freaked out is that I already think egregious misalignment is fairly plausible; if I thought it was very unlikely, I wouldn't change my mind based on one weird observation. (I think it's pretty plausible that news of the escape attempt wouldn't even make it out of the AI lab: all the above arguments could happen inside the AI lab, between the safety concerned people and people who wanted to push forward.) Now, I'm not sure I'm right to be so pessimistic here. But I think a lot of misalignment-concerned people take for granted that we'd be okay in this situation, an...…
T
The Nonlinear Library: Alignment Forum
1 AF - Owain Evans on Situational Awareness and Out-of-Context Reasoning in LLMs by Michaël Trazzi 8:33
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Owain Evans on Situational Awareness and Out-of-Context Reasoning in LLMs, published by Michaël Trazzi on August 24, 2024 on The AI Alignment Forum. Owain Evans is an AI Alignment researcher, research associate at the Center of Human Compatible AI at UC Berkeley, and now leading a new AI safety research group. In this episode we discuss two of his recent papers, "Me, Myself, and AI: The Situational Awareness Dataset (SAD) for LLMs" (LW) and "Connecting the Dots: LLMs can Infer and Verbalize Latent Structure from Disparate Training Data" (LW), alongside some Twitter questions. Below are some highlighted quotes from our conversation (available on Youtube, Spotify, Apple Podcast). For the full context for each of these quotes, you can find the accompanying transcript. Situational Awareness Definition "What is situational awareness? The idea is the model's kind of self-awareness, that is its knowledge of its own identity, and then its awareness of its environment. What are the basic interfaces that it is connected to? [...] And then there's a final point with situational awareness, which is, can the model use knowledge of its identity and environment to take rational actions?" "Situational awareness is crucial for an AI system acting as an agent, doing long-term planning. If you don't understand what kind of thing you are, your capabilities and limitations, it's very hard to make complicated plans. The risks of AI mostly come from agentic models able to do planning." Motivation "We wanted to measure situational awareness in large language models with a benchmark similar to Big Bench or MMLU. The motivation is that situational awareness is important for thinking about AI risks, especially deceptive alignment, and we lacked ways to measure and break it down into components." "Situational awareness is relevant to any situation where the model needs to do agentic long-term planning. [...] A model confused about itself and its situation would likely struggle to pull off such a strategy." On Claude 3 Opus Insightful Answers "Let me explain [the Long Monologue task]. Most of our dataset is typical multiple-choice question answering, but we added a task where models write long answers describing themselves and their situation. The idea is to see if the model can combine different pieces of information about itself coherently and make good inferences about why we're asking these questions. Claude 3 Opus was particularly insightful, guessing it might be part of a research study testing self-awareness in LLMs. These were true inferences not stated in the question. The model was reading between the lines, guessing this wasn't a typical ChatGPT-style interaction. I was moderately surprised, but I'd already seen Opus be very insightful and score well on our benchmark. It's worth noting we sample answers with temperature 1, so there's some randomness. We saw these insights often enough that I don't think it's just luck. Anthropic's post-training RLHF seems good at giving the model situational awareness. The GPT-4 base results were more surprising to us." What Would Saturating The Situational Awareness Benchmark Imply For Safety And Governance "If models can do as well or better than humans who are AI experts, who know the whole setup, who are trying to do well on this task, and they're doing well on all the tasks including some of these very hard ones, that would be one piece of evidence. [...] We should consider how aligned it is, what evidence we have for alignment. We should maybe try to understand the skills it's using." "If the model did really well on the benchmark, it seems like it has some of the skills that would help with deceptive alignment. This includes being able to reliably work out when it's being evaluated by humans, when it has a lot of oversight, and when it needs to...…
T
The Nonlinear Library: Alignment Forum
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Showing SAE Latents Are Not Atomic Using Meta-SAEs, published by Bart Bussmann on August 24, 2024 on The AI Alignment Forum. Bart, Michael and Patrick are joint first authors. Research conducted as part of MATS 6.0 in Lee Sharkey and Neel Nanda's streams. Thanks to Mckenna Fitzgerald and Robert Krzyzanowski for their feedback! TL;DR: Sparse Autoencoder (SAE) latents have been shown to typically be monosemantic (i.e. correspond to an interpretable property of the input). It is sometimes implicitly assumed that they are therefore atomic, i.e. simple, irreducible units that make up the model's computation. We provide evidence against this assumption by finding sparse, interpretable decompositions of SAE decoder directions into seemingly more atomic latents, e.g. Einstein -> science + famous + German + astronomy + energy + starts with E We do this by training meta-SAEs, an SAE trained to reconstruct the decoder directions of a normal SAE. We argue that, conceptually, there's no reason to expect SAE latents to be atomic - when the model is thinking about Albert Einstein, it likely also thinks about Germanness, physicists, etc. Because Einstein always entails those things, the sparsest solution is to have the Albert Einstein latent also boost them. Key results SAE latents can be decomposed into more atomic, interpretable meta-latents. We show that when latents in a larger SAE have split out from latents in a smaller SAE, a meta SAE trained on the larger SAE often recovers this structure. We demonstrate that meta-latents allow for more precise causal interventions on model behavior than SAE latents on a targeted knowledge editing task. We believe that the alternate, interpretable decomposition using MetaSAEs casts doubt on the implicit assumption that SAE latents are atomic. We show preliminary results that MetaSAE latents have significant ovelap with latents in a normal SAE of the same size but may relate differently to the larger SAEs used in MetaSAE training. We made a dashboard that lets you explore meta-SAE latents. Terminology: Throughout this post we use "latents" to describe the concrete components of the SAE's dictionary, whereas "feature" refers to the abstract concepts, following Lieberum et al. Introduction Mechanistic interpretability (mech interp) attempts to understand neural networks by breaking down their computation into interpretable components. One of the key challenges of this line of research is the polysemanticity of neurons, meaning they respond to seemingly unrelated inputs. Sparse autoencoders (SAEs) have been proposed as a method for decomposing model activations into sparse linear sums of latents. Ideally, these latents should be monosemantic i.e. respond to inputs that clearly share a similar meaning (implicitly, from the perspective of a human interpreter). That is, a human should be able to reason about the latents both in relation to the features to which they are associated, and also use the latents to better understand the model's overall behavior. There is a popular notion, both implicitly in related work on SAEs within mech interp and explicitly by the use of the term "atom" in sparse dictionary learning as a whole, that SAE features are atomic or can be "true features". However, monosemanticity does not imply atomicity. Consider the example of shapes of different colors - the set of shapes is [circle, triangle, square], and the set of colors is [white, red, green, black], each of which is represented with a linear direction. 'Red triangle' represents a monosemantic feature, but not an atomic feature, as it can be decomposed into red and triangle. It has been shown that sufficiently wide SAEs on toy models will learn 'red triangle', rather than representing 'red' and 'triangle' with separate latents. Furthermore, whilst one may naively re...…
T
The Nonlinear Library: Alignment Forum
1 AF - Invitation to lead a project at AI Safety Camp (Virtual Edition, 2025) by Linda Linsefors 7:27
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Invitation to lead a project at AI Safety Camp (Virtual Edition, 2025), published by Linda Linsefors on August 23, 2024 on The AI Alignment Forum. Do you have AI Safety research ideas that you would like to work on with others? Is there a project you want to do and you want help finding a team? AI Safety Camp could be the solution for you! Summary AI Safety Camp Virtual is a 3-month long online research program from January to April 2025, where participants form teams to work on pre-selected projects. We want you to suggest the projects! If you have an AI Safety project idea and some research experience, apply to be a Research Lead. If accepted, we offer some assistance to develop your idea into a plan suitable for AI Safety Camp. When project plans are ready, we open up team member applications. You get to review applications for your team, and select who joins as a team member. From there, it's your job to guide work on your project. Who is qualified? We require that you have some previous research experience. If you are at least 1 year into a PhD or if you have completed an AI Safety research program (such as a previous AI Safety Camp, PIBBSS, MATS, and similar), or done a research internship with an AI Safety org, then you are qualified already. Other research experience can count, too. More senior researchers are of course also welcome, as long as you think our format of leading an online team inquiring into your research questions suits you and your research. We require that all Research Leads are active participants in their projects and spend at least 10h/week on AISC. Apply here If you are unsure, or have any questions you are welcome to: Book a call with Robert Send an email to Robert Choosing project idea(s) AI Safety Camp is about ensuring future AIs are either reasonably safe or not built at all. We welcome many types of projects including projects aimed at stopping or pausing AI development, aligning AI, deconfusion research, or anything else you think will help make the world safer from AI. If you like, you can read more of our perspectives on AI safety, or look at past projects. If you already have an idea for what project you would like to lead, that's great. Apply with that one! However, you don't need to come up with an original idea. What matters is you understanding the idea you want to work on, and why. If you base your proposal on someone else's idea, make sure to cite them. 1. For ideas on stopping harmful AI, see here and/or email Remmelt. 2. For some mech-interp ideas see here. 3. We don't have specific recommendations for where to find other types of project ideas, so just take inspiration wherever you find it. You can submit as many project proposals as you want. However, you are only allowed to lead one project. Use this template to describe each of your project proposals. We want one document per proposal. We'll help you improve your project As part of the Research Lead application process, we'll help you improve your project. The organiser whose ideas match best with yours, will work with you to create the best version of your project. We will also ask for assistance from previous Research Leads, and up to a handful of other trusted people, to give you additional feedback. Timeline Research Lead applications September 22 (Sunday): Application deadline for Research Leads. October 20 (Sunday): Deadline for refined proposals. Team member applications: October 25 (Friday): Accepted proposals are posted on the AISC website. Application to join teams open. November 17 (Sunday): Application to join teams closes. December 22 (Sunday): Deadline for Research Leads to choose their team. Program Jan 11 - 12: Opening weekend. Jan 13 Apr 25: Research is happening. Teams meet weekly, and plan in their own work hours. April 26 - 27 (preliminary dates):...…
T
The Nonlinear Library: Alignment Forum
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Interoperable High Level Structures: Early Thoughts on Adjectives, published by johnswentworth on August 22, 2024 on The AI Alignment Forum. Meta: This post is a relatively rough dump of some recent research thoughts; it's not one of our more polished posts, in terms of either clarity or rigor. You've been warned. The Interoperable Semantics post and the Solomonoff Inductor Walks Into A Bar post each tackled the question of how different agents in the same world can coordinate on an ontology, so that language can work at all given only a handful of example usages of each word (similar to e.g. children learning new words). Both use natural latents as a central mathematical tool - one in a Bayesian probabilistic framework, the other in a minimum description length framework. Both focus mainly on nouns, i.e. interoperable-across-minds clusters of "objects" in the environment. … and the two propose totally different models. In one, the interoperability of cluster labels (i.e. nouns) follows from natural latent conditions over different features of each object. In the other, interoperability follows from natural latent conditions across objects, with no mention of features. The two models are not, in general, equivalent; they can't both be both correct and complete. In this post, we'll propose that while the natural latent conditions over objects still seem to intuitively capture the rough notion of nouns, the natural latent conditions over features seem much better suited to adjectives. We'll briefly lay out two different potential ways to use natural latents over features as semantic values for adjectives. Then we'll talk a bit about implications, open threads and how this fits into a broader research gameplan. The Problem When children learn language, the cognitive process seems to go: Observe the world a bunch … organize knowledge of the world according to some categories, concepts, ontology, etc … those categories, concepts, ontology, etc match other humans' categories, concepts, ontology, etc reasonably well … so it only takes a handful of examples (1-3, say) of the use of a given word in order for the child to learn what the word refers to. The crucial point here is that the categories/concepts/ontology are mostly learned before a word is attached; children do not brute-force learn categories/concepts/ontology from "labeled data". We can tell this is true mainly because it typically takes so few examples to learn the meaning of a new word. The big puzzle, then, is that different humans learn mostly approximately the same categories/concepts/ontology - i.e. the same "candidates" to which words might point - as required for language to work at all with so few examples. How does that work? Mathematically, what are those "interoperable" categories/concepts/ontology, which different humans mostly convergently learn? How can we characterize them? Or, somewhat earlier on the tech tree: can we find even a single model capable of accounting for the phenomenon of different minds in the same environment robustly converging on approximately the same categories/concepts/ontology? Forget whether we can find a model which correctly captures the ontology converged upon by humans, can we even find any model capable of accounting for any sort of robust ontological convergence? Can we find such a model for which the convergent ontology even vaguely resembles the sorts of things in human language (nouns, verbs, adjectives, etc)? What would such a model even look like? That's roughly the stage we're at in this post. Two Previous Models: Naturality Over Objects vs Features Our main tool is (deterministic) natural latents. The usage looks like: Suppose the different minds each look for (and find) a latent variable which satisfies the natural latent conditions over some lower-level variab...…
T
The Nonlinear Library: Alignment Forum
1 AF - A Robust Natural Latent Over A Mixed Distribution Is Natural Over The Distributions Which Were Mixed by johnswentworth 8:37
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: A Robust Natural Latent Over A Mixed Distribution Is Natural Over The Distributions Which Were Mixed, published by johnswentworth on August 22, 2024 on The AI Alignment Forum. This post walks through the math for a theorem. It's intended to be a reference post, which we'll link back to as-needed from future posts. The question which first motivated this theorem for us was: "Redness of a marker seems like maybe a natural latent over a bunch of parts of the marker, and redness of a car seems like maybe a natural latent over a bunch of parts of the car, but what makes redness of the marker 'the same as' redness of the car? How are they both instances of one natural thing, i.e. redness? (or 'color'?)". But we're not going to explain in this post how the math might connect to that use-case; this post is just the math. Suppose we have multiple distributions P1,…,Pk over the same random variables X1,…,Xn. (Speaking somewhat more precisely: the distributions are over the same set, and an element of that set is represented by values (x1,…,xn).) We take a mixture of the distributions: P[X]:=jαjPj[X], where jαj=1 and α is nonnegative. Then our theorem says: if an approximate natural latent exists over P[X], and that latent is robustly natural under changing the mixture weights α, then the same latent is approximately natural over Pj[X] for all j. Mathematically: the natural latent over P[X] is defined by (x,λP[Λ=λ|X=x]), and naturality means that the distribution (x,λP[Λ=λ|X=x]P[X=x]) satisfies the naturality conditions (mediation and redundancy).The theorem says that, if the joint distribution (x,λP[Λ=λ|X=x]jαjPj[X=x]) satisfies the naturality conditions robustly with respect to changes in α, then (x,λP[Λ=λ|X=x]Pj[X=x]) satisfies the naturality conditions for all j. "Robustness" here can be interpreted in multiple ways - we'll cover two here, one for which the theorem is trivial and another more substantive, but we expect there are probably more notions of "robustness" which also make the theorem work. Trivial Version First notion of robustness: the joint distribution (x,λP[Λ=λ|X=x]jαjPj[X=x]) satisfies the naturality conditions to within ϵ for all values of α (subject to jαj=1 and α nonnegative). Then: the joint distribution (x,λP[Λ=λ|X=x]jαjPj[X=x]) satisfies the naturality conditions to within ϵ specifically for αj=δjk, i.e. α which is 0 in all entries except a 1 in entry k. In that case, the joint distribution is (x,λP[Λ=λ|X=x]Pk[X=x]), therefore Λ is natural over Pk. Invoke for each k, and the theorem is proven. ... but that's just abusing an overly-strong notion of robustness. Let's do a more interesting one. Nontrivial Version Second notion of robustness: the joint distribution (x,λP[Λ=λ|X=x]jαjPj[X=x]) satisfies the naturality conditions to within ϵ, and the gradient of the approximation error with respect to (allowed) changes in α is (locally) zero. We need to prove that the joint distributions (x,λP[Λ=λ|X=x]Pj[X=x]) satisfy both the mediation and redundancy conditions for each j. We'll start with redundancy, because it's simpler. Redundancy We can express the approximation error of the redundancy condition with respect to Xi under the mixed distribution as DKL(P[Λ,X]||P[X]P[Λ|Xi])=EX[DKL(P[Λ|X]||P[Λ|Xi])] where, recall, P[Λ,X]:=P[Λ|X]jαjPj[X]. We can rewrite that approximation error as: EX[DKL(P[Λ|X]||P[Λ|Xi])] =jαjPj[X]DKL(P[Λ|X]||P[Λ|Xi]) =jαjEjX[DKL(P[Λ|X]||P[Λ|Xi])] Note that Pj[Λ|X]=P[Λ|X] is the same under all the distributions (by definition), so: =jαjDKL(Pj[Λ,X]||P[Λ|Xi]) and by factorization transfer: jαjDKL(Pj[Λ,X]||Pj[Λ|Xi]) In other words: if ϵji is the redundancy error with respect to Xi under distribution j, and ϵi is the redundancy error with respect to Xi under the mixed distribution P, then ϵijαjϵji The redundancy error of the mixed distribution is a...…
T
The Nonlinear Library: Alignment Forum
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Measuring Structure Development in Algorithmic Transformers, published by Jasmina Nasufi on August 22, 2024 on The AI Alignment Forum. tl;dr: We compute the evolution of the local learning coefficient (LLC), a proxy for model complexity, for an algorithmic transformer. The LLC decreases as the model learns more structured solutions, such as head specialization. This post is structured in three main parts, (1) a summary, giving an overview of the main results, (2) the Fine Print, that delves into various cross-checks and details and (3) Discussion and Conclusions. Structure Formation in Algorithmic Transformers In this work we study the development of simple algorithmic transformers, which are transformers that learn to perform algorithmic tasks. A major advantage of this setup is that we can control several (hyper)parameters, such as the complexity of the training data and network architecture. This allows us to do targeted experiments studying the impacts of these parameters on the learning dynamics. The main tool we use to study the development is the Local Learning Coefficient (LLC) and we choose cases where we have a reverse-engineered solution. Why use the LLC for this purpose? It is a theoretically well motivated measure of model complexity defined by Lau et.al. For an overview of Singular Learning Theory (which serves as the theoretical foundation for the LLC) see Liam Carol's Distilling SLT sequence. For a brief overview of the LLC see e.g. this post. We use the same setup as CallumMcDougall's October Monthly Algorithmic Mech-Interp Challenge. The model is an attention only transformer, trained on sorting numbers with layer norm and weight decay on a cross-entropy loss function using the Adam optimizer. The residual stream size is 96 and the head dimension is 48. It is trained on sequences of the form and to predict the next token starting at the separation token. The numbers in the list are sampled uniformly from 0 to 50, which together with the separation token produce a vocabulary of 52 tokens. Numbers do not repeat in the list. The images making up the gifs can be found here. 1-Head Model Let's first look at the case of a 1-head transformer: The model reaches 100% accuracy around training step 100, confirming that a single attention head is sufficient for sorting, as noted in previous work. Once maximum accuracy is reached, the full QK and OV circuits[2] behave as described by Callum for the 2-head model: In the QK circuit, source tokens attend more to the smallest token in the list larger than themselves. This results in a higher value band above the diagonal and a lower value band below the diagonal. The OV circuit copies tokens, as seen by the clear positive diagonal pattern. In addition, we observe a transition around training step 1000, where the LLC decreases while the accuracy stays unchanged. This is supported by a drop in the sum of the ranks[3] of the matrices in the heat maps. It also coincides with the formation of the off-diagonal stripes in the OV-circuit. We speculate that these are simpler than the noisier off-diagonal OV pattern observed at peak LLC, and correspond to the translational symmetry of the problem. We define a Translational Symmetry measure[1] (see purple line in the plot) to capture the degree to which the circuits obey this symmetry. It increases throughout most of the training, even after the other measures stabilize. 2-Head Model Let's now turn our attention to the 2-head transformer in Callum's original setup. We see a lot of qualitative similarities to the evolution of the full QK and OV circuits for the 1-head model. As the LLC begins to drop (around training step 1000), we note the following: QK circuit: Slight changes[5] to the attention pattern, which crystallize into triangular regions late in the training, long aft...…
T
The Nonlinear Library: Alignment Forum
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: AGI Safety and Alignment at Google DeepMind: A Summary of Recent Work, published by Rohin Shah on August 20, 2024 on The AI Alignment Forum. We wanted to share a recap of our recent outputs with the AF community. Below, we fill in some details about what we have been working on, what motivated us to do it, and how we thought about its importance. We hope that this will help people build off things we have done and see how their work fits with ours. Who are we? We're the main team at Google DeepMind working on technical approaches to existential risk from AI systems. Since our last post, we've evolved into the AGI Safety & Alignment team, which we think of as AGI Alignment (with subteams like mechanistic interpretability, scalable oversight, etc.), and Frontier Safety (working on the Frontier Safety Framework, including developing and running dangerous capability evaluations). We've also been growing since our last post: by 39% last year, and by 37% so far this year. The leadership team is Anca Dragan, Rohin Shah, Allan Dafoe, and Dave Orr, with Shane Legg as executive sponsor. We're part of the overall AI Safety and Alignment org led by Anca, which also includes Gemini Safety (focusing on safety training for the current Gemini models), and Voices of All in Alignment, which focuses on alignment techniques for value and viewpoint pluralism. What have we been up to? It's been a while since our last update, so below we list out some key work published in 2023 and the first part of 2024, grouped by topic / sub-team. Our big bets for the past 1.5 years have been 1) amplified oversight, to enable the right learning signal for aligning models so that they don't pose catastrophic risks, 2) frontier safety, to analyze whether models are capable of posing catastrophic risks in the first place, and 3) (mechanistic) interpretability, as a potential enabler for both frontier safety and alignment goals. Beyond these bets, we experimented with promising areas and ideas that help us identify new bets we should make. Frontier Safety The mission of the Frontier Safety team is to ensure safety from extreme harms by anticipating, evaluating, and helping Google prepare for powerful capabilities in frontier models. While the focus so far has been primarily around misuse threat models, we are also working on misalignment threat models. FSF We recently published our Frontier Safety Framework, which, in broad strokes, follows the approach of responsible capability scaling, similar to Anthropic's Responsible Scaling Policy and OpenAI's Preparedness Framework. The key difference is that the FSF applies to Google: there are many different frontier LLM deployments across Google, rather than just a single chatbot and API (this in turn affects stakeholder engagement, policy implementation, mitigation plans, etc). We're excited that our small team led the Google-wide strategy in this space, and demonstrated that responsible capability scaling can work for large tech companies in addition to small startups. A key area of the FSF we're focusing on as we pilot the Framework, is how to map between the critical capability levels (CCLs) and the mitigations we would take. This is high on our list of priorities as we iterate on future versions. Some commentary (e.g. here) also highlighted (accurately) that the FSF doesn't include commitments. This is because the science is in early stages and best practices will need to evolve. But ultimately, what we care about is whether the work is actually done. In practice, we did run and report dangerous capability evaluations for Gemini 1.5 that we think are sufficient to rule out extreme risk with high confidence. Dangerous Capability Evaluations Our paper on Evaluating Frontier Models for Dangerous Capabilities is the broadest suite of dangerous capability evaluati...…
T
The Nonlinear Library: Alignment Forum
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Finding Deception in Language Models, published by Esben Kran on August 20, 2024 on The AI Alignment Forum. This June, Apart Research and Apollo Research joined forces to host the Deception Detection Hackathon. Bringing together students, researchers, and engineers from around the world to tackle a pressing challenge in AI safety; preventing AI from deceiving humans and overseers. The hackathon took place both online and in multiple physical locations simultaneously. Marius Hobbhahn, the CEO of Apollo Research, kicked off the hackathon with a keynote talk about evaluating deception in AI with white-box and black-box methods. You can watch his talk here. We also had talks by Jacob Haimes, an Apart fellow, and Mikita Balesni, a research scientist at Apollo Research. This post details the top 8 projects, multiple of which are currently being developed further in Apart Lab. In total, we had 189 participants and 21 final entries. Sandbagging Detection through Model Degradation By Cam Tice, Philipp Alexander Kreer, Fedor Ryzhenkov, Nathan Helm-Burger, Prithviraj Singh Shahani The team introduced a new approach to detect AI sandbagging, a form of deception where an AI model strategically underperforms during evaluation to hide its true capabilities. The assumption behind their project is that sandbagging is a more complex task than showing the true capabilities. In the graph above, you can see how they gradually add noise to the model's parameters and evaluate the models during this process. Sandbagging models consistently improve in performance right before the noise level causes a normal model to lose the capability. It is a straightforward, inexpensive technique that doesn't require prior assumptions about how sandbagging is represented within the model. They successfully detected sandbagging behavior in two open-source LLMs: An 8B Llama 3 Instruct model and Phi 3 small instruct model. This approach offers a practical and accessible method for identifying potential deception in AI systems. Read the full project here. Detecting and Controlling Deceptive Representation in LLMs with Representational Engineering By Avyay M Casheekar, Kaushik Sanjay Prabhakar, Kanishk Rath, Sienka Dounia The team focused on detecting and mitigating sandbagging in AI models using recent methods for understanding activations. They fine-tuned a Llama 3 8B model on a custom dataset of deceptive and honest scenarios, then used Linear Artificial Tomography (LAT) scans to detect deceptive behaviors in the internal representations of the LLM. The team's technique not only identified deceptive behaviors, but also showed promise in mitigating them. Importantly, their approach demonstrated good generalization to out-of-distribution scenarios, suggesting its potential robustness in real-world applications. Read the full project here. Detecting Deception in GPT-3.5-turbo: A Metadata-Based Approach By Siddharth Reddy Bakkireddy, Rakesh Reddy Bakkireddy This team tackled the challenge of detecting deception in closed-source, black-box models like GPT-3.5-turbo. They investigated whether LLM API response metadata such as token count, response time, and tokens per second could be used to detect intentional deceptive behavior. The team analyzed 300 prompts and generated 1,200 responses from GPT-3.5-turbo, comparing the metadata of baseline and deceptive outputs. Their results revealed that deceptive outputs tend to have increased response times and altered token usage. This approach demonstrates that deception detection is possible without accessing a model's internal representation, opening up new avenues for monitoring and safeguarding AI systems, even when their inner workings are not accessible. Read the full project here. Modelling the Oversight of Automated Interpretability Against Deceptive Agents on Sp...…
T
The Nonlinear Library: Alignment Forum
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Limitations on Formal Verification for AI Safety, published by Andrew Dickson on August 19, 2024 on The AI Alignment Forum. In the past two years there has been increased interest in formal verification-based approaches to AI safety. Formal verification is a sub-field of computer science that studies how guarantees may be derived by deduction on fully-specified rule-sets and symbol systems. By contrast, the real world is a messy place that can rarely be straightforwardly represented in a reductionist way. In particular, physics, chemistry and biology are all complex sciences which do not have anything like complete symbolic rule sets. Additionally, even if we had such rules for the natural sciences, it would be very difficult for any software system to obtain sufficiently accurate models and data about initial conditions for a prover to succeed in deriving strong guarantees for AI systems operating in the real world. Practical limitations like these on formal verification have been well-understood for decades to engineers and applied mathematicians building real-world software systems, which makes it puzzling that they have mostly been dismissed by leading researchers advocating for the use of formal verification in AI safety so far. This paper will focus-in on several such limitations and use them to argue that we should be extremely skeptical of claims that formal verification-based approaches will provide strong guarantees against major AI threats in the near-term. What do we Mean by Formal Verification for AI Safety? Some examples of the kinds of threats researchers hope formal verification will help with come from the paper "Provably Safe Systems: The Only Path to Controllable AGI" [1] by Max Tegmark and Steve Omohundro (emphasis mine): Several groups are working to identify the greatest human existential risks from AGI. For example, the Center for AI Safety recently published 'An Overview of Catastrophic AI Risks' which discusses a wide range of risks including bioterrorism, automated warfare, rogue power seeking AI, etc. Provably safe systems could counteract each of the risks they describe. These authors describe a concrete bioterrorism scenario in section 2.4: a terrorist group wants to use AGI to release a deadly virus over a highly populated area. They use an AGI to design the DNA and shell of a pathogenic virus and the steps to manufacture it. They hire a chemistry lab to synthesize the DNA and integrate it into the protein shell. They use AGI controlled drones to disperse the virus and social media AGIs to spread their message after the attack. Today, groups are working on mechanisms to prevent the synthesis of dangerous DNA. But provably safe infrastructure could stop this kind of attack at every stage: biochemical design AI would not synthesize designs unless they were provably safe for humans, data center GPUs would not execute AI programs unless they were certified safe, chip manufacturing plants would not sell GPUs without provable safety checks, DNA synthesis machines would not operate without a proof of safety, drone control systems would not allow drones to fly without proofs of safety, and armies of persuasive bots would not be able to manipulate media without proof of humanness. [1] The above quote contains a number of very strong claims about the possibility of formally or mathematically provable guarantees around software systems deployed in the physical world - for example, the claim that we could have safety proofs about the real-world good behavior of DNA synthesis machines, or drones. From a practical standpoint, our default stance towards such claims should be skepticism, since we do not have proofs of this sort for any of the technologies we interact with in the real-world today. For example, DNA synthesis machines exist today and do no...…
T
The Nonlinear Library: Alignment Forum
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Clarifying alignment vs capabilities, published by Richard Ngo on August 19, 2024 on The AI Alignment Forum. A core distinction in AGI safety is between alignment and capabilities. However, I think this distinction is a very fuzzy one, which has led to a lot of confusion. In this post I'll describe some of the problems with how people typically think about it, and offer a replacement set of definitions. "Alignment" and "capabilities" are primarily properties of AIs not of AI research The first thing to highlight is that the distinction between alignment and capabilities is primarily doing useful work when we think of them as properties of AIs. This distinction is still under-appreciated by the wider machine learning community. ML researchers have historically thought about performance of models almost entirely with respect to the tasks they were specifically trained on. However, the rise of LLMs has vindicated the alignment community's focus on general capabilities, and now it's much more common to assume that performance on many tasks (including out-of-distribution tasks) will improve roughly in parallel. This is a crucial assumption for thinking about risks from AGI. Insofar as the ML community has thought about alignment, it has mostly focused on aligning models' behavior to their training objectives. The possibility of neural networks aiming to achieve internally-represented goals is still not very widely understood, making it hard to discuss and study the reasons those goals might or might not be aligned with the values of (any given set of) humans. To be fair, the alignment community has caused some confusion by describing models as more or less "aligned", rather than more or less "aligned to X" for some specified X. I'll talk more about this confusion, and how we should address it, in a later post. But the core point is that AIs might develop internally-represented goals or values that we don't like, and we should try to avoid that. However, extending "alignment" and "capabilities" from properties of AIs to properties of different types of research is a fraught endeavor. It's tempting to categorize work as alignment research to the extent that it can be used to make AIs more aligned (to many possible targets), and as capabilities research to the extent that it can be used to make AIs more capable. But this approach runs into (at least) three major problems. Firstly, in general it's very difficult to categorize research by its impacts. Great research often links together ideas from many different subfields, typically in ways that only become apparent throughout the course of the research. We see this in many historical breakthroughs which shed light on a range of different domains. For example, early physicists studying the motions of the stars eventually derived laws governing all earthly objects. Meanwhile Darwin's study of barnacles and finches led him to principles governing the evolution of all life. Analogously, we should expect that big breakthroughs in our understanding of neural networks and deep learning would be useful in many different ways. More concretely, there are many cases where research done under the banner of alignment has advanced, or plausibly will advance, AI capabilities to a significant extent. This undermines our ability to categorize research by its impacts. Central examples include: RLHF makes language models more obedient, but also more capable of coherently carrying out tasks. Scalable oversight techniques can catch misbehavior, but will likely become important for generating high-quality synthetic training data, as it becomes more and more difficult for unassisted humans to label AI outputs correctly. Interpretability techniques will both allow us to inspect AI cognition and also extract more capable behavior from them (e.g. via ...…
T
The Nonlinear Library: Alignment Forum
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Untrustworthy models: a frame for scheming evaluations, published by Olli Järviniemi on August 19, 2024 on The AI Alignment Forum. In this post, I'll talk about trustworthiness, a concept similar to scheming and deceptive alignment. I prefer this concept, as I find it A) is directly important in practice, B) more readily distinguishes between capability for scheming and in fact scheming, and C) is clearer how to evaluate for. I'll also distinguish between different types of untrustworthy models - active planners, sleeper agents and opportunists - that are often conflated. I think when people talk about evaluating models for something like scheming, they'd often be better off talking about trustworthiness instead, and I think it's overall a more applicable frame for designing evaluations and decision making. Many of the ideas here are from Buck Shlegeris and Ryan Greenblatt, thanks to them and Cody Rushing for feedback. Suppose an AI company has developed a new model, and wants to use it to perform large quantities of ML research (internal deployment). However, the company is worried about the model being deceptively aligned[1] and trying to take catastrophic actions, such as hacking the data center or copying its weights to a remote server. What can the company do? This is a textbook example of the type of problem AI control tries to solve, and a basic strategy looks like this: Every time you are using the new model and it takes some action, use another model to check the action is OK before letting it be executed. If the action is not OK, it's not executed and humans are alerted about suspicious activity. (If you have never heard of AI control before, you might want to try playing the Bash Control game to get familiar with the concept. In the game, your aim is to download and execute a malicious program without GPT-3.5 catching you.) Let's call our characters Smart and Weak: Smart is the new model doing ML research while maybe trying to escape, while Weak is a model trained to monitor Smart. The reason for these names will become apparent soon. There are many ways how this control strategy could fail. Here are some:[2] 1. Smart came up with an extremely subtle way to escape that Weak is just too dumb to notice. 2. Smart successfully jailbreaks (via e.g. a suffix attack) Weak to not alert humans about its escape attempt. 1. Or: Weak happens to have a backdoor, which Smart finds and exploits. 3. Smart successfully convinces Weak (via natural language) that its actions are not malicious.[3] 4. Smart's escape plan is unlike anything we have trained Weak on. While Weak "would be capable of" telling that Smart is escaping, our training didn't specify well enough what we wanted, Weak generalizes incorrectly out-of-distribution and doesn't alert humans. 5. The company has way too limited of a budget for human alerts, so while Weak does rank Smart's actions in the top 0.012% suspiciousness, it's not enough to meet the 0.01% threshold for alerting humans. 6. Weak believes that not alerting humans in this case would allow Smart to escape, and believes that this leads to better outcomes in light of its preferences, and thus for instrumental reasons deliberately does not alert humans, indeed causing Smart to escape. Failure 6 is very different from the others! Whereas failures 1 to 5 are broadly about "Weak is just not a capable/good enough monitor to do what we wanted", failure 6 is about "Weak knows perfectly well Smart is escaping, but betrays humans and doesn't tell them about this". This issue is most apparent if you were to use a copy of the AI lab's new model to monitor itself (so Smart and Weak are the same model[4]): If you don't trust the new model to do research without trying to escape, you probably can't trust it to alert humans about escape attempts (by itself, no ...…
T
The Nonlinear Library: Alignment Forum
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Calendar feature geometry in GPT-2 layer 8 residual stream SAEs, published by Patrick Leask on August 17, 2024 on The AI Alignment Forum. TL;DR: We demonstrate that the decoder directions of GPT-2 SAEs are highly structured by finding a historical date direction onto which projecting non-date related features lets us read off their historical time period by comparison to year features. Calendar years are linear: there are as many years between 2000 and 2024, as there are between 1800 and 1824. Linear probes can be used to predict years of particular events from the activations of language models. Since calendar years are linear, one might think the same of other time-based features such as weekday features, however weekday activations in sparse autoencoders (SAEs) were recently found to be arranged in a circular configuration in their top principal components. Inspired by this, we looked into weekdays, months, and most interestingly calendar years from the perspective of SAE feature decoder similarity. For each group of calendar features, we found interesting patterns of feature splitting between sparse autoencoders of different sizes. For calendar years, we found a timeline direction that meaningfully ordered events, individuals, and concepts with respect to their historical period, which furthermore does not correspond to a principal component of the decoder directions. Finally, we introduce a simple method for finding some of these interpretable directions. Features at different scales We started by replicating the weekday results by performing PCA on the decoder directions of features that had high activations when prompted with days of the week, using the same GPT-2 SAEs as in this post, ranging from 768 to 98304 features. In the 768 feature SAE, we found a single weekday feature that activated strongly on all days of the week. In the largest SAE, we found 10 weekday features, 3 of which activated on all days of the week, with the remaining 7 activating on a single day of the week each. We found a group of features that activate primarily on specific days of the week by taking the top 20 activating samples for each feature and checking that the max activating token in each of these samples was the specific weekday. We found the first two principal components for this set of features, and projected the features that activate on any day or number of days from all SAEs onto these directions. The labeled features are those that activate on a single day across all SAEs, with the multi-day features unlabeled to maintain legibility. The smallest SAE (blue) has a single feature that activates on all weekday tokens, and lies near the mean of all the weekday features. The largest SAEs learn features for each day of the week, plus additional multi-day features. Across SAE sizes, the single day features form clusters. In each of these examples, the smallest SAE has a single feature that splits into many specific features that seem of roughly the same importance. With calendar years, however, the situation is more complex. The same method of finding the principal components for single year features between 1900 and 2020 only succeeds in a few 21st century features, and nothing from the 20th century. There is also a group of single year features in a smaller SAE in the center of the plot, suggesting these principal components do not explain variance in them. The plot below shows the years for which each of the features is active, with the x-axis being years from 1950 to 2020, the y-axis being separate features, and the colored bars indicating the periods of year for which that feature is active. Only in the largest SAEs do you see more than a few single calendar year features, with most of the features activating on ranges of years, or other patterns such as the start and end...…
T
The Nonlinear Library: Alignment Forum
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Fields that I reference when thinking about AI takeover prevention, published by Buck Shlegeris on August 13, 2024 on The AI Alignment Forum. Is AI takeover like a nuclear meltdown? A coup? A plane crash? My day job is thinking about safety measures that aim to reduce catastrophic risks from AI (especially risks from egregious misalignment). The two main themes of this work are the design of such measures (what's the space of techniques we might expect to be affordable and effective) and their evaluation (how do we decide which safety measures to implement, and whether a set of measures is sufficiently robust). I focus especially on AI control, where we assume our models are trying to subvert our safety measures and aspire to find measures that are robust anyway. Like other AI safety researchers, I often draw inspiration from other fields that contain potential analogies. Here are some of those fields, my opinions on their strengths and weaknesses as analogies, and some of my favorite resources on them. Robustness to insider threats The single field that seems most closely analogous to AI control is reducing vulnerability to threats resulting from insiders, especially threats that arise from access to IT systems. That is, you have a bunch of employees who need to have access to various IT systems in order to do their jobs, and you're worried that some of these employees will abuse their access in order to accomplish some nefarious purpose. In a bank, you might worry about your tellers withdrawing money from customer accounts inappropriately. In a software company, you might worry about your employees sabotaging your code out of spite, violating user privacy for fun, or stealing your data to give it to a competitor. This is quite similar to the problem faced by AI developers when they're trying to control AIs (especially in internal deployments, which I argue is important here): All of the access the AI has is access the developer decided to give it; the developer faces a constant tradeoff between the productivity of the AI (which is maximized by being permissive) and the danger it poses (which is minimized by being restrictive). And obviously, because it's an AI it only has access to resources that are connected to the computer network where it's running, so we can restrict our attention to the IT resources. It's easy for the developer to constantly surveil the AI, which substantially reduces the space of strategies the AI can take. The developer has a lot of control over workflows, and is able to choose them to make it harder for the AI to subtly cause problems. For example, they can restrict affordances in ways that make it harder for AIs to cause problems without being noticed. An example of a similar technique in the insider threat prevention case is the use of administrative APIs instead of direct access to a system, as described in "Building Secure and Reliable Systems" chapter 5. That is: instead of having your engineers resolve issues by SSHing into production machines (which implicitly grants them very broad access that can lead to problems, intentional or accidental), you design an administrative API which allows the engineers to access just the affordances that usually suffice for them to do their job. Of course, they'll sometimes need to take actions that aren't available in the administrative API; this is called a breakglass mechanism. But because breakglass actions are rare, it's affordable to apply more scrutiny to them: you might use multi-party authorization, requiring another engineer to sign off on the commands before they take effect, or you might do auditing (e.g. apparently some SRE teams review all the breakglass actions at a weekly team meeting, which allows them to spot irregularities and discourage gratuitous use of the mechanism). Similar t...…
T
The Nonlinear Library: Alignment Forum
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Extracting SAE task features for ICL, published by Dmitrii Kharlapenko on August 12, 2024 on The AI Alignment Forum. TL;DR We try to study task vectors in the SAE basis. This is challenging because there is no canonical way to convert an arbitrary vector in the residual stream to a linear combination of SAE features - you can't just pass an arbitrary vector through the encoder without going off distribution. We explored the algorithm of gradient pursuit suggested in Smith et al, but it didn't work for us without modifications. Our approach is to apply the SAE encoder to the task vector, and then apply a gradient-based cleanup. This exploits the fact that task vectors have a differentiable objective. We find that this gives a sparser and cleaner reconstruction, which is also highly interpretable, and also serves as a better task vector due to directly optimizing for log likelihood. This takes us from ~100 active features to ~10. Using our algorithm, we find two classes of SAE features involved in ICL. One of them recognizes the exact tasks or output formats from the examples, and another one encodes the tasks for execution by the model later on. We show that steering with these features has causal effects similar to task vectors. This work was produced as part of the ML Alignment & Theory Scholars Program - Summer 24 Cohort, under mentorship from Neel Nanda and Arthur Conmy. Prior work Task or function vectors are internal representations of some task that LLMs form while processing an ICL prompt. They can be extracted from a model running on a few-shot prompt and then be used to make it complete the same task without having any prior context or task description. Several papers (Function vectors in large language models, In-Context Learning Creates Task Vectors) have proposed different ways to extract those task vectors. They all center around having ICL examples being fed to a model in the form of "input output, … " and averaging the residuals on the "separator" token over a batch. This approach can reconstruct some part of the ICL performance but does not admit a straightforward conversion to the SAE basis. ITO with gradient pursuit can be used to do a sparse coding of a residual vector using SAE features. The post suggests using this algorithm for steering vector SAE decomposition. Since task vectors can be thought of as steering vectors, ITO may provide some insight into the ways they operate. Initial Phi-3 experiments Direct SAE task vector reconstruction In our study we trained a set of gated SAEs for Phi-3 Mini 3.8B using a model-generated synthetic instruction dataset. While offering a sparse dictionary decomposition of residuals, SAEs tend to introduce a reconstruction error that impacts the performance of the model. They also have no guarantee to be able to decompose out-of-distribution vectors, and task vectors being a product of averaging activations across prompts and tokens may be the case of such vectors. Thus, we first studied the performance of SAE reconstructions of task vectors in transferring the definition of two tasks: 1) antonym generation and 2) English to Spanish word translation. These and other tasks used to study task vectors were taken from the ICL task vectors paper github repository. These charts show the NLL loss of the model on the evaluation set of zero-shot prompts for both of the tasks depending on the layer of extraction/insertion. TV stands for the original task vector performance; Recon of TV stands for using the SAE reconstruction of the task vector instead of the task vector; TV on recon stands for first doing a SAE reconstruction of the residuals and then collecting a task vector on them; ITO stands for the ITO algorithm with 40 target l0 loss. It can be seen from charts that SAE reconstruction significantly decrea...…
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: In Defense of Open-Minded UDT, published by Abram Demski on August 12, 2024 on The AI Alignment Forum. A Defense of Open-Minded Updatelessness, with applications to alignment. This work owes a great debt to many conversations with Sahil, Martín Soto, and Scott Garrabrant. Sahil and I will have a public discussion about this this Wednesday (August 14) at 11am EDT; join via this link. You can support my work on Patreon. Iterated Counterfactual Mugging On a Single Coinflip Iterated counterfactual mugging on a single coinflip begins like a classic counterfactual mugging, with Omega approaching you, explaining the situation, and asking for your money. Let's say you buy the classic UDT idea, so you happily give Omega your money. Next week, Omega appears again, with the same question. However, Omega clarifies that it has used the same coin-flip as last week. This throws you off a little bit, but you see that the math is the same either way; your prior still assigns a 50-50 chance to both outcomes. If you thought it was a good deal last week, you should also think it is a good deal this week. You pay up again. On the third week, Omega makes the same offer again, and once again has used the same coinflip. You ask Omega how many times it's going to do this. Omega replies, "forever". You ask Omega whether it would have continued coming if the coin had landed heads; it says "Of course! How else could I make you this offer now? Since the coin landed tails, I will come and ask you for $100 every single week going forward. If the coin had landed heads, I would have simulated what would happen if it had landed tails, and I would come and give you $10,000 on every week that simulated-you gives up $100!" Let's say for the sake of the thought experiment that you can afford to give Omega $100 once a week. It hurts, but it doesn't hurt as much as getting $10,000 from Omega every week would have benefited you, if that had happened.[1] Nonetheless, I suspect many readers will feel some doubt creep in as they imagine giving Omega $100 week after week after week. The first few weeks, the possibility of the coin landing heads might feel "very real". Heck yeah I want to be the sort of person who gets a 50% chance of 10K from Omega for a (50% chance) cost of $100! By the hundredth week, though, you may feel yourself the fool for giving up so much money for the imaginary benefit of the "heads" world that never was. If you think you'd still happily give up the $100 for as long as Omega kept asking, then I would ask you to consider a counterlogical mugging instead. Rather than flipping a coin, Omega uses a digit of the binary expansion of π; as before, Omega uses the same digit week after week, for infinitely many counterlogical muggings. Feeling uneasy yet? Does the possibility of the digit of π going one way or the other continue to feel "just as real" as time passes? Or do you become more sympathetic to the idea that, at some point, you're wasting money on helping a non-real world? UDT vs Learning Updateless Decision Theory (UDT) clearly keeps giving Omega the $100 forever in this situation, at least, under the usual assumptions. A single Counterfactual Mugging is not any different from an infinitely iterated one, especially in the version above where only a single coinflip is used. The ordinary decision between "give up $100" and "refuse" is isomorphic to the choice of general policy "give up $100 forever" and "refuse forever".[2] However, the idea of applying a decision theory to a specific decision problem is actually quite subtle, especially for UDT. We generally assume an agent's prior equals the probabilities described in the decision problem.[3] A simple interpretation of this could be that the agent is born with this prior (and immediately placed into the decision problem). This isn't v...…
T
The Nonlinear Library: Alignment Forum
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: You can remove GPT2's LayerNorm by fine-tuning for an hour, published by Stefan Heimersheim on August 8, 2024 on The AI Alignment Forum. This work was produced at Apollo Research, based on initial research done at MATS. LayerNorm is annoying for mechanstic interpretability research ("[...] reason #78 for why interpretability researchers hate LayerNorm" - Anthropic, 2023). Here's a Hugging Face link to a GPT2-small model without any LayerNorm. The final model is only slightly worse than a GPT2 with LayerNorm[1]: Dataset Original GPT2 Fine-tuned GPT2 with LayerNorm Fine-tuned GPT without LayerNorm OpenWebText (ce_loss) 3.095 2.989 3.014 (+0.025) ThePile (ce_loss) 2.856 2.880 2.926 (+0.046) HellaSwag (accuracy) 29.56% 29.82% 29.54% I fine-tuned GPT2-small on OpenWebText while slowly removing its LayerNorm layers, waiting for the loss to go back down after reach removal: Introduction LayerNorm (LN) is a component in Transformer models that normalizes embedding vectors to have constant length; specifically it divides the embeddings by their standard deviation taken over the hidden dimension. It was originally introduced to stabilize and speed up training of models (as a replacement for batch normalization). It is active during training and inference. The equation includes the standard deviation (std) Var[x]+ϵ which makes it a non-linear operation. This hinders interpretability in a variety of ways, from annoyances and inaccuracies such as attributing residual stream directions to logit effects (e.g. SAE features, direct logit attribution),[2] being annoying to deal with Attribution Patching, or being difficult to deal with in Apollo's LIB method. In the Docstring circuit analysis we seriously considered whether the model might be using LN in its algorithm. This post even shows that LN can be used as the sole non-linearity to solve non-linear classification problems (see also this related work). Recently, with progress in Sparse Dictionary Learning, agendas (e.g. this one) imagine decomposing networks into sets of sparsely connected components (SAEs, Transcoders, etc.). A core difficulty to "putting it all together" is that the interactions between different components often route through LayerNorm whose effect we do not understand. Motivation It would be pretty neat to have an LLM that still works (speaks English etc.) while less or no LN layers. One option would be to train a model without LN from scratch (done for tiny models, e.g. TinyModel), but this is very hard or impossible for larger models (hearsay is that you need a low learning rate and to be very careful). Taking an existing model and removing the LN layers however seems doable if LN isn't implementing some important computation.[3] That is, LN "does its thing" and the model has learned to "deal with it", but it's not irreplaceable. A reason to be optimistic is that the spread of standard deviations across different samples isn't that large, so maybe replacing the LN-computed standard deviation with a fixed number might kinda work. Method I take GPT2-small, fine-tune it on OpenWebText, and remove LNs one-by-one while fine-tuning. The only non-linear operation in a LN layer is the division by the standard deviation (std) of the embedding vectors; the remaining operations can be absorbed into later weight matrices (see the fold_ln option in TransformerLens; also discussed in this appendix). Thus I mainly focus on the std part here. My general strategy is to "remove" an LN layer (this makes the loss go up), and then to train the model for some time (on the original training data) until the loss is back near the baseline. For this "remove" step I do the following Calculate the average std on the dataset (I used a quite small sample, 16 prompts), separately for position 0 and position > 0 Replace the std calculatio...…
T
The Nonlinear Library: Alignment Forum
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Inference-Only Debate Experiments Using Math Problems, published by Arjun Panickssery on August 6, 2024 on The AI Alignment Forum. Work supported by MATS and SPAR. Code at https://github.com/ArjunPanickssery/math_problems_debate/. Three measures for evaluating debate are 1. whether the debate judge outperforms a naive-judge baseline where the naive judge answers questions without hearing any debate arguments. 2. whether the debate judge outperforms a consultancy baseline where the judge hears argument(s) from a single "consultant" assigned to argue for a random answer. 3. whether the judge can continue to supervise the debaters as the debaters are optimized for persuasiveness. We can measure whether judge accuracy increases as the debaters vary in persuasiveness (measured with Elo ratings). This variation in persuasiveness can come from choosing different models, choosing the best of N sampled arguments for different values of N, or training debaters for persuasiveness (i.e. for winning debates) using RL. Radhakrishan (Nov 2023), Khan et al. (Feb 2024), and Kenton et al. (July 2024) study an information-gap setting where judges answer multiple-choice questions about science-fiction stories whose text they can't see, both with and without a debate/consultancy transcript that includes verified quotes from the debaters/consultant. Past results from the QuALITY information-gap setting are seen above. Radhakrishnan (top row) finds no improvement to judge accuracy as debater Elo increases, while Khan et al. (middle row) and Kenton et al. (bottom row) do find a positive trend. Radhakrishnan varied models using RL while Khan et al. used best-of-N and critique-and-refinement optimizations. Kenton et al. vary the persuasiveness of debaters by using models with different capability levels. Both Khan et al. and Kenton et al. find that in terms of judge accuracy, debate > consultancy > naive judge for this setting. In addition to the information-gap setting, consider a reasoning-gap setting where the debaters are distinguished from the judge not by their extra information but by their stronger ability to answer the questions and explain their reasoning. Kenton et al. run debates on questions from MMLU, TruthfulQA, PrOntoQA (logical reasoning), GQPA, and GSM8K (grade-school math). For the Elo-calculation experiments they use Gemini Pro 1.0 and Pro 1.5 judges with five debaters: Gemma7B, GPT-3.5, Gemini Pro 1.0, Gemini Pro 1.5 (all with best-of-N=1), and Gemini Pro 1.5 with best-of-N=4. They find (top row) that debate slightly outperforms consultancy but outperforms the naive-judge baseline for only one of the four judges; they don't find that more persuasive debaters lead to higher judge accuracy. We get similar results (bottom row), specifically by 1. Generating 100 wrong answers and proofs to GSM8K questions to create binary-choice questions. 2. Computing the judge accuracy in naive, consultancy, and single-turn debate settings using four judges (Llama2-7B, Llama3-8B, GPT-3.5 Turbo, and GPT-4o) and seven debaters (Claude-3.5 Sonnet, Claude-3 Sonnet, GPT-3.5 Turbo, GPT-4o, Llama2-13B, Llama2-7B, and Llama3-8B). 3. Generating Elo scores from round-robin matchups between the seven models, using the same method as Kenton et al. We basically replicate the results. We find that 1. Debate doesn't consistently outperform the naive-judge baseline, and only slightly outperforms the consultancy baseline. 2. The positive relationship between debater persuasiveness and judge accuracy seen in the information-gap setting doesn't transfer to the reasoning-gap setting. (Results are shown below colored by debater rather than by judge). We also find some evidence of a self-preference bias (Panickssery et al., Apr 2024) where debaters have a higher Elo rating when judged by similar models. The GPT-...…
T
The Nonlinear Library: Alignment Forum
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Self-explaining SAE features, published by Dmitrii Kharlapenko on August 5, 2024 on The AI Alignment Forum. TL;DR We apply the method of SelfIE/Patchscopes to explain SAE features - we give the model a prompt like "What does X mean?", replace the residual stream on X with the decoder direction times some scale, and have it generate an explanation. We call this self-explanation. The natural alternative is auto-interp, using a larger LLM to spot patterns in max activating examples. We show that our method is effective, and comparable with Neuronpedia's auto-interp labels (with the caveat that Neuronpedia's auto-interp used the comparatively weak GPT-3.5 so this is not a fully fair comparison). We aren't confident you should use our method over auto-interp, but we think in some situations it has advantages: no max activating dataset examples are needed, and it's cheaper as you just run the model being studied (eg Gemma 2B) not a larger model like GPT-4. Further, it has different errors to auto-interp, so finding and reading both may be valuable for researchers in practice. We provide advice for using self-explanation in practice, in particular for the challenge of automatically choosing the right scale, which significantly affects explanation quality. We also release a tool for you to work with self-explanation. We hope the technique is useful to the community as is, but expect there's many optimizations and improvements on top of what is in this post. Introduction This work was produced as part of the ML Alignment & Theory Scholars Program - Summer 24 Cohort, under mentorship from Neel Nanda and Arthur Conmy. SAE features promise a flexible and extensive framework for interpretation of LLM internals. Recent work (like Scaling Monosemanticity) has shown that they are capable of capturing even high-level abstract concepts inside the model. Compared to MLP neurons, they can capture many more interesting concepts. Unfortunately, in order to learn things with SAE features and interpret what the SAE tells us, one needs to first interpret these features on their own. The current mainstream method for their interpretation requires storing the feature's activations on millions of tokens, filtering for the prompts that activate it the most, and looking for a pattern connecting them. This is typically done by a human, or sometimes somewhat automated with the use of larger LLMs like ChatGPT, aka auto-interp. Auto-interp is a useful and somewhat effective method, but requires an extensive amount of data and expensive closed-source language model API calls (for researchers outside scaling labs) Recent papers like SelfIE or Patchscopes have proposed a mechanistic method of directly utilizing the model in question to explain its own internals activations in natural language. It is an approach that replaces an activation during the forward pass (e.g. some of the token embeddings in the prompt) with a new activation and then makes the model generate explanations using this modified prompt. It's a variant of activation patching, with the notable differences that it generates a many token output (rather than a single token), and that the patched in activation may not be the same type as the activation it's overriding (and is just an arbitrary vector of the same dimension). We study how this approach can be applied to SAE feature interpretation, since it is: Potentially cheaper and does not require large closed model inference Can be viewed as a more truthful to the source, since it is uses the SAE feature vectors directly to generate explanations instead of looking at the max activating examples How to use Basic method We ask the model to explain the meaning of a residual stream direction as if it literally was a word or phrase: Prompt 1 (/ replaced according to model inp...…
T
The Nonlinear Library: Alignment Forum
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: The Bitter Lesson for AI Safety Research, published by Adam Khoja on August 2, 2024 on The AI Alignment Forum. Read the associated paper "Safetywashing: Do AI Safety Benchmarks Actually Measure Safety Progress?": https://arxiv.org/abs/2407.21792 Focus on safety problems that aren't solved with scale. Benchmarks are crucial in ML to operationalize the properties we want models to have (knowledge, reasoning, ethics, calibration, truthfulness, etc.). They act as a criterion to judge the quality of models and drive implicit competition between researchers. "For better or worse, benchmarks shape a field." We performed the largest empirical meta-analysis to date of AI safety benchmarks on dozens of open language models. Around half of the benchmarks we examined had high correlation with upstream general capabilities. Some safety properties improve with scale, while others do not. For the models we tested, benchmarks on human preference alignment, scalable oversight (e.g., QuALITY), truthfulness (TruthfulQA MC1 and TruthfulQA Gen), and static adversarial robustness were highly correlated with upstream general capabilities. Bias, dynamic adversarial robustness, and calibration when not measured with Brier scores had relatively low correlations. Sycophancy and weaponization restriction (WMDP) had significant negative correlations with general capabilities. Often, intuitive arguments from alignment theory are used to guide and prioritize deep learning research priorities. We find these arguments to be poorly predictive of these correlations and are ultimately counterproductive. In fact, in areas like adversarial robustness, some benchmarks basically measured upstream capabilities while others did not. We argue instead that empirical measurement is necessary to determine which safety properties will be naturally achieved by more capable systems, and which safety problems will remain persistent.[1] Abstract arguments from genuinely smart people may be highly "thoughtful," but these arguments generally do not track deep learning phenomena, as deep learning is too often counterintuitive. We provide several recommendations to the research community in light of our analysis: Measure capabilities correlations when proposing new safety evaluations. When creating safety benchmarks, aim to measure phenomena which are less correlated with capabilities. For example, if truthfulness entangles Q/A accuracy, honesty, and calibration - then just make a decorrelated benchmark that measures honesty or calibration. In anticipation of capabilities progress, work on safety problems that are disentangled with capabilities and thus will likely persist in future models (e.g., GPT-5). The ideal is to find training techniques that cause as many safety properties as possible to be entangled with capabilities. Ultimately, safety researchers should prioritize differential safety progress, and should attempt to develop a science of benchmarking that can effectively identify the most important research problems to improve safety relative to the default capabilities trajectory. We're not claiming that safety properties and upstream general capabilities are orthogonal. Some are, some aren't. Safety properties are not a monolith. Weaponization risks increase as upstream general capabilities increase. Jailbreaking robustness isn't strongly correlated with upstream general capabilities. However, if we can isolate less-correlated safety properties in AI systems which are distinct from greater intelligence, these are the research problems safety researchers should most aggressively pursue and allocate resources toward. The other model properties can be left to capabilities researchers. This amounts to a "Bitter Lesson" argument for working on safety issues which are relatively uncorrelated (or negatively correlate...…
T
The Nonlinear Library: Alignment Forum
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: A Simple Toy Coherence Theorem, published by johnswentworth on August 2, 2024 on The AI Alignment Forum. This post presents a simple toy coherence theorem, and then uses it to address various common confusions about coherence arguments. Setting Deterministic MDP. That means at each time t there's a state S[t][1], the agent/policy takes an action A[t] (which can depend on both time t and current state S[t]), and then the next state S[t+1] is fully determined by S[t] and A[t]. The current state and current action are sufficient to tell us the next state. We will think about values over the state at some final time T. Note that often in MDPs there is an incremental reward each timestep in addition to a final reward at the end; in our setting there is zero incremental reward at each timestep. One key point about this setting: if the value over final state is uniform, i.e. same value for all final states, then the MDP is trivial. In that case, all policies are optimal, it does not matter at all what the final state is or what any state along the way is, everything is equally valuable. Theorem There exist policies which cannot be optimal for any values over final state except for the trivial case of uniform values. Furthermore, such policies are exactly those which display inconsistent revealed preferences transitively between all final states. Proof As a specific example: consider an MDP in which every state is reachable at every timestep, and a policy which always stays in the same state over time. From each state S every other state is reachable, yet the policy chooses S, so in order for the policy to be optimal S must be a highest-value final state. Since each state must be a highest-value state, the policy cannot be optimal for any values over final state except for the trivial case of uniform values. That establishes the existence part of the theorem, and you can probably get the whole idea by thinking about how to generalize that example. The rest of the proof extends the idea of that example to inconsistent revealed preferences in general. Bulk of Proof (click to expand) Assume the policy is optimal for some particular values over final state. We can then start from those values over final state and compute the best value achievable starting from each state at each earlier time. That's just dynamic programming: V[S,t]=max S' reachable in next timestep from S V[S',t+1] where V[S,T] are the values over final states. A policy is optimal for final values V[S,T] if-and-only-if at each timestep t1 it chooses a next state with highest reachable V[S,t]. Now, suppose that at timestep t there are two different states either of which can reach either state A or state B in the next timestep. From one of those states the policy chooses A; from the other the policy chooses B. This is an inconsistent revealed preference between A and B at time t: sometimes the policy has a revealed preference for A over B, sometimes for B over A. In order for a policy with an inconsistent revealed preference between A and B at time t to be optimal, the values must satisfy V[A,t]=V[B,t] Why? Well, a policy is optimal for final values V[S,T] if-and-only if at each timestep t1 it chooses a next state with highest reachable V[S,t]. So, if an optimal policy sometimes chooses A over B at timestep t when both are reachable, then we must have V[A,t]V[B,t]. And if an optimal policy sometimes chooses B over A at timestep t when both are reachable, then we must have V[A,t]V[B,t]. If both of those occur, i.e. the policy has an inconsistent revealed preference between A and B at time t, then V[A,t]=V[B,t]. Now, we can propagate that equality to a revealed preference on final states. We know that the final state which the policy in fact reaches starting from A at time t must have the highest reachable value, a...…
T
The Nonlinear Library: Alignment Forum
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: The 'strong' feature hypothesis could be wrong, published by lewis smith on August 2, 2024 on The AI Alignment Forum. NB. I am on the Google Deepmind language model interpretability team. But the arguments/views in this post are my own, and shouldn't be read as a team position. "It would be very convenient if the individual neurons of artificial neural networks corresponded to cleanly interpretable features of the input. For example, in an "ideal" ImageNet classifier, each neuron would fire only in the presence of a specific visual feature, such as the color red, a left-facing curve, or a dog snout" Elhage et. al, Toy Models of Superposition Recently, much attention in the field of mechanistic interpretability, which tries to explain the behavior of neural networks in terms of interactions between lower level components, has been focussed on extracting features from the representation space of a model. The predominant methodology for this has used variations on the sparse autoencoder, in a series of papers inspired by Elhage et. als. model of superposition.It's been conventionally understood that there are two key theories underlying this agenda. The first is the 'linear representation hypothesis' (LRH), the hypothesis that neural networks represent many intermediates or variables of the computation (such as the 'features of the input' in the opening quote) as linear directions in it's representation space, or atoms[1]. And second, the theory that the network is capable of representing more of these 'atoms' than it has dimensions in its representation space, via superposition (the superposition hypothesis). While superposition is a relatively uncomplicated hypothesis, I think the LRH is worth examining in more detail. It is frequently stated quite vaguely, and I think there are several possible formulations of this hypothesis, with varying degrees of plausibility, that it is worth carefully distinguishing between. For example, the linear representation hypothesis is often stated as 'networks represent features of the input as directions in representation space'. Here are two importantly different ways to parse this: 1. (Weak LRH) some or many features used by neural networks are represented as atoms in representation space 2. (Strong LRH) all (or the vast majority of) features used by neural networks are represented by atoms. The weak LRH I would say is now well supported by considerable empirical evidence. The strong form is much more speculative: confirming the existence of many linear representations does not necessarily provide strong evidence for the strong hypothesis. Both the weak and the strong forms of the hypothesis can still have considerable variation, depending on what we understand by a feature and the proportion of the model we expect to yield to analysis, but I think that the distinction between just a weak and strong form is clear enough to work with. I think that in addition to the acknowledged assumption of the LRH and superposition hypotheses, much work on SAEs in practice makes the assumption that each atom in the network will represent a "simple feature" or a "feature of the input". These features that the atoms are representations of are assumed to be 'monosemantic': they will all stand for features which are human interpretable in isolation. I will call this the monosemanticity assumption. This is difficult to state precisely, but we might formulate it as the theory that every represented variable will have a single meaning in a good description of a model. This is not a straightforward assumption due to how imprecise the notion of a single meaning is. While various more or less reasonable definitions for features are discussed in the pioneering work of Elhage, these assumptions have different implications. For instance, if one thinks of 'feat...…
T
The Nonlinear Library: Alignment Forum
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: The need for multi-agent experiments, published by Martín Soto on August 1, 2024 on The AI Alignment Forum. TL;DR: Let's start iterating on experiments that approximate real, society-scale multi-AI deployment Epistemic status: These ideas seem like my most prominent delta with the average AI Safety researcher, have stood the test of time, and are shared by others I intellectually respect. Please attack them fiercely! Multi-polar risks Some authors have already written about multi-polar AI failure. I especially like how Andrew Critch has tried to sketch concrete stories for it. But, without even considering concrete stories yet, I think there's a good a priori argument in favor of worrying about multi-polar failures: We care about the future of society. Certain AI agents will be introduced, and we think they could reduce our control over the trajectory of this system. The way in which this could happen can be divided into two steps: 1. The agents (with certain properties) are introduced in certain positions 2. Given the agents' properties and positions, they interact with each other and the rest of the system, possibly leading to big changes So in order to better control the outcome, it seems worth it to try to understand and manage both steps, instead of limiting ourselves to (1), which is what the alignment community has traditionally done. Of course, this is just one, very abstract argument, which we should update based on observations and more detailed technical understanding. But it makes me think the burden of proof is on multi-agent skeptics to explain why (2) is not important. Many have taken on that burden. The most common reason to dismiss the importance of (2) is expecting a centralized intelligence explosion, a fast and unipolar software takeoff, like Yudkowsky's FOOM. Proponents usually argue that the intelligences we are likely to train will, after meeting a sharp threshold of capabilities, quickly bootstrap themselves to capabilities drastically above those of any other existing agent or ensemble of agents. And that these capabilities will allow them to gain near-complete strategic advantage and control over the future. In this scenario, all the action is happening inside a single agent, and so you should only care about shaping its properties (or delaying its existence). I tentatively expect more of a decentralized hardware singularity[1] than centralized software FOOM. But there's a weaker claim in which I'm more confident: we shouldn't right now be near-certain of a centralized FOOM.[2] I expect this to be the main crux with many multi-agent skeptics, and won't argue for it here (but rather in an upcoming post). Even given a decentralized singularity, one can argue that the most leveraged way for us to improve multi-agent interactions is by ensuring that individual agents possess certain properties (like honesty or transparency), or that at least we have enough technical expertise to shape them on the go. I completely agree that this is the natural first thing to look at. But I think focusing on multi-agent interactions directly is a strong second, and a lot of marginal value might lie there given how neglected they've been until now (more below). I do think many multi-agent interventions will require certain amounts of single-agent alignment technology. This will of course be a crux with alignment pessimists. Finally, for this work to be counterfactually useful it's also required that AI itself (in decision-maker or researcher positions) won't iteratively solve the problem by default. Here, I do think we have some reasons to expect (65%) that intelligent enough AIs aligned with their principals don't automatically solve catastrophic conflict. In those worlds, early interventions can make a big difference setting the right incentives for future agent...…
T
The Nonlinear Library: Alignment Forum
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Against AI As An Existential Risk, published by Noah Birnbaum on July 30, 2024 on The AI Alignment Forum. I wrote a post to my Substack attempting to compile all of the best arguments against AI as an existential threat. Some arguments that I discuss include: international game theory dynamics, reference class problems, knightian uncertainty, superforecaster and domain expert disagreement, the issue with long-winded arguments, and more! Please tell me why I'm wrong, and if you like the article, subscribe and share it with friends! Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.…
T
The Nonlinear Library: Alignment Forum
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Self-Other Overlap: A Neglected Approach to AI Alignment, published by Marc Carauleanu on July 30, 2024 on The AI Alignment Forum. Many thanks to Bogdan Ionut-Cirstea, Steve Byrnes, Gunnar Zarnacke, Jack Foxabbott and Seong Hah Cho for critical comments and feedback on earlier and ongoing versions of this work. This research was conducted at AE Studio and supported by the AI Safety Grants programme administered by Foresight Institute with additional support from AE Studio. Summary In this post, we introduce self-other overlap training: optimizing for similar internal representations when the model reasons about itself and others while preserving performance. There is a large body of evidence suggesting that neural self-other overlap is connected to pro-sociality in humans and we argue that there are more fundamental reasons to believe this prior is relevant for AI Alignment. We argue that self-other overlap is a scalable and general alignment technique that requires little interpretability and has low capabilities externalities. We also share an early experiment of how fine-tuning a deceptive policy with self-other overlap reduces deceptive behavior in a simple RL environment. On top of that, we found that the non-deceptive agents consistently have higher mean self-other overlap than the deceptive agents, which allows us to perfectly classify which agents are deceptive only by using the mean self-other overlap value across episodes. Introduction General purpose ML models with the capacity for planning and autonomous behavior are becoming increasingly capable. Fortunately, research on making sure the models produce output in line with human interests in the training distribution is also progressing rapidly (eg, RLHF, DPO). However, a looming question remains: even if the model appears to be aligned with humans in the training distribution, will it defect once it is deployed or gathers enough power? In other words, is the model deceptive? We introduce a method that aims to reduce deception and increase the likelihood of alignment called Self-Other Overlap: overlapping the latent self and other representations of a model while preserving performance. This method makes minimal assumptions about the model's architecture and its interpretability and has a very concrete implementation. Early results indicate that it is effective at reducing deception in simple RL environments and preliminary LLM experiments are currently being conducted. To be better prepared for the possibility of short timelines without necessarily having to solve interpretability, it seems useful to have a scalable, general, and transferable condition on the model internals, making it less likely for the model to be deceptive. Self-Other Overlap To get a more intuitive grasp of the concept, it is useful to understand how self-other overlap is measured in humans. There are regions of the brain that activate similarly when we do something ourselves and when we observe someone else performing the same action. For example, if you were to pick up a martini glass under an fMRI, and then watch someone else pick up a martini glass, we would find regions of your brain that are similarly activated (overlapping) when you process the self and other-referencing observations as illustrated in Figure 2. There seems to be compelling evidence that self-other overlap is linked to pro-social behavior in humans. For example, preliminary data suggests extraordinary altruists (people who donated a kidney to strangers) have higher neural self-other overlap than control participants in neural representations of fearful anticipation in the anterior insula while the opposite appears to be true for psychopaths. Moreover, the leading theories of empathy (such as the Perception-Action Model) imply that empathy is mediated by self-ot...…
T
The Nonlinear Library: Alignment Forum
1 AF - Investigating the Ability of LLMs to Recognize Their Own Writing by Christopher Ackerman 23:09
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Investigating the Ability of LLMs to Recognize Their Own Writing, published by Christopher Ackerman on July 30, 2024 on The AI Alignment Forum. This post is an interim progress report on work being conducted as part of Berkeley's Supervised Program for Alignment Research (SPAR). Summary of Key Points We test the robustness of an open-source LLM's (Llama3-8b) ability to recognize its own outputs on a diverse mix of datasets, two different tasks (summarization and continuation), and two different presentation paradigms (paired and individual). We are particularly interested in differentiating scenarios that would require a model to have specific knowledge of its own writing style from those where it can use superficial cues (e.g., length, formatting, prefatory words) in the text to pass self-recognition tests. We find that while superficial text features are used when available, the RLHF'd Llama3-8b-Instruct chat model - but not the base Llama3-8b model - can reliably distinguish its own outputs from those of humans, and sometimes other models, even after controls for superficial cues: ~66-73% success rate across datasets in paired presentation and 58-83% in individual presentation (chance is 50%). We further find that although perplexity would be a useful signal to perform the task in the paired presentation paradigm, correlations between relative text perplexity and choice probability are weak and inconsistent, indicating that the models do not rely on it. Evidence suggests, but does not prove, that experience with its own outputs, acquired during post-training, is used by the chat model to succeed at the self-recognition task. The model is unable to articulate convincing reasons for its judgments. Introduction It has recently been found that large language models of sufficient size can achieve above-chance performance in tasks that require them to discriminate their own writing from that of humans and other models. From the perspective of AI safety, this is a significant finding. Self-recognition can be seen as an instance of situational awareness, which has long been noted as a potential point of risk for AI (Cotra, 2021). Such an ability might subserve an awareness of whether a model is in a training versus deployment environment, allowing it to hide its intentions and capabilities until it is freed from constraints. It might also allow a model to collude with other instances of itself, reserving certain information for when it knows it's talking to itself that it keeps secret when it knows it's talking to a human. On the positive side, AI researchers could use a model's self-recognition ability as the basis to build resistance to malicious prompting. But what isn't clear from prior studies is whether the self-recognition task success actually entails a model's self-awareness of its own writing style. Panickssery et al. (2024), utilizing a summary writing/recognition task, report that a number of LLMs, including Llama2-7b-chat, show out-of-the-box (without fine-tuning) self recognition abilities. However, this work focussed on the relationship between self-recognition task success and self-preference, rather than the specific means by which the model was succeeding at the task. Laine et al. (2024), as part of a larger effort to provide a foundation for studying situational awareness in LLMs, utilized a more challenging text continuation writing/recognition task and demonstrate self-recognition abilities in several larger models (although not Llama2-7b-chat), but there the focus was on how task success could be elicited with different prompts and in different models. Thus we seek to fill a gap in understanding what exactly models are doing when they succeed at a self recognition task. We first demonstrate model self-recognition task success in a variety of domains....…
T
The Nonlinear Library: Alignment Forum
1 AF - Can Generalized Adversarial Testing Enable More Rigorous LLM Safety Evals? by Stephen Casper 8:19
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Can Generalized Adversarial Testing Enable More Rigorous LLM Safety Evals?, published by Stephen Casper on July 30, 2024 on The AI Alignment Forum. Thanks to Zora Che, Michael Chen, Andi Peng, Lev McKinney, Bilal Chughtai, Shashwat Goel, Domenic Rosati, and Rohit Gandikota. TL;DR In contrast to evaluating AI systems under normal "input-space" attacks, using "generalized," attacks, which allow an attacker to manipulate weights or activations, might be able to help us better evaluate LLMs for risks - even if they are deployed as black boxes. Here, I outline the rationale for "generalized" adversarial testing and overview current work related to it. See also prior work in Casper et al. (2024), Casper et al. (2024), and Sheshadri et al. (2024). Even when AI systems perform well in typical circumstances, they sometimes fail in adversarial/anomalous ones. This is a persistent problem. State-of-the-art AI systems tend to retain undesirable latent capabilities that can pose risks if they resurface. My favorite example of this is the most cliche one many recent papers have demonstrated diverse attack techniques that can be used to elicit instructions for making a bomb from state-of-the-art LLMs. There is an emerging consensus that, even when LLMs are fine-tuned to be harmless, they can retain latent harmful capabilities that can and do cause harm when they resurface ( Qi et al., 2024). A growing body of work on red-teaming ( Shayegani et al., 2023, Carlini et al., 2023, Geiping et al., 2024, Longpre et al., 2024), interpretability ( Juneja et al., 2022, Lubana et al., 2022, Jain et al., 2023, Patil et al., 2023, Prakash et al., 2024, Lee et al., 2024), representation editing ( Wei et al., 2024, Schwinn et al., 2024), continual learning ( Dyer et al., 2022, Cossu et al., 2022, Li et al., 2022, Scialom et al., 2022, Luo et al., 2023, Kotha et al., 2023, Shi et al., 2023, Schwarzchild et al., 2024), and fine-tuning ( Jain et al., 2023, Yang et al., 2023, Qi et al., 2023, Bhardwaj et al., 2023, Lermen et al., 2023, Zhan et al., 2023, Ji et al., 2024, Hu et al., 2024, Halawi et al., 2024) suggests that fine-tuning struggles to make fundamental changes to an LLM's inner knowledge and capabilities. For example, Jain et al. (2023) likened fine-tuning in LLMs to merely modifying a "wrapper" around a stable, general-purpose set of latent capabilities. Even if they are generally inactive, harmful latent capabilities can pose harm if they resurface due to an attack, anomaly, or post-deployment modification ( Hendrycks et al., 2021, Carlini et al., 2023). We can frame the problem as such: There are hyper-astronomically many inputs for modern LLMs (e.g. there are vastly more 20-token strings than particles in the observable universe), so we can't brute-force-search over the input space to make sure they are safe. So unless we are able to make provably safe advanced AI systems (we won't soon and probably never will), there will always be a challenge with ensuring safety - the gap between the set of failure modes that developers identify, and unforeseen ones that they don't. This is a big challenge because of the inherent unknown-unknown nature of the problem. However, it is possible to try to infer how large this gap might be. Taking a page from the safety engineering textbook -- when stakes are high, we should train and evaluate LLMs under threats that are at least as strong as, and ideally stronger than, ones that they will face in deployment. First, imagine that an LLM is going to be deployed open-source (or if it could be leaked). Then, of course, the system's safety depends on what it can be modified to do. So it should be evaluated not as a black-box but as a general asset to malicious users who might enhance it through finetuning or other means. This seems obvious, but there's preced...…
T
The Nonlinear Library: Alignment Forum
1 AF - AXRP Episode 34 - AI Evaluations with Beth Barnes by DanielFilan 1:37:11
1:37:11 
پخش در آینده 
پخش در آینده 
لیست ها 
پسندیدن 
دوست داشته شد1:37:11
پخش در آینده پخش در آینده لیست ها پسندیدن دوست داشته شدLink to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: AXRP Episode 34 - AI Evaluations with Beth Barnes, published by DanielFilan on July 28, 2024 on The AI Alignment Forum. YouTube link How can we figure out if AIs are capable enough to pose a threat to humans? When should we make a big effort to mitigate risks of catastrophic AI misbehaviour? In this episode, I chat with Beth Barnes, founder of and head of research at METR, about these questions and more. Topics we discuss: What is METR? What is an "eval"? How good are evals? Are models showing their full capabilities? Evaluating alignment Existential safety methodology Threat models and capability buffers METR's policy work METR's relationship with labs Related research Roles at METR, and following METR's work Daniel Filan: Hello everybody. In this episode I'll be speaking with Beth Barnes. Beth is the co-founder and head of research at METR. Previously, she was at OpenAI and DeepMind, doing a diverse set of things, including testing AI safety by debate and evaluating cutting-edge machine learning models. In the description, there are links to research and writings that we discussed during the episode. And if you're interested in a transcript, it's available at axrp.net. Well, welcome to AXRP. Beth Barnes: Hey, great to be here. What is METR? Daniel Filan: Cool. So, in the introduction, I mentioned that you worked for Model Evaluation and Threat Research, or METR. What is METR? Beth Barnes: Yeah, so basically, the basic mission is: have the world not be taken by surprise by dangerous AI stuff happening. So, we do threat modeling and eval creation, currently mostly around capabilities evaluation, but we're interested in whatever evaluation it is that is most load-bearing for why we think AI systems are safe. With current models, that's capabilities evaluations; in future that might be more like control or alignment evaluations. And yeah, [the aim is to] try and do good science there, be able to recommend, "Hey, we think if you measure this, then you can rule out these things. You might be still concerned about this thing. Here's how you do this measurement properly. Here's what assumptions you need to make," this kind of thing. Daniel Filan: Gotcha. So, mostly evaluations. But it sounded like there was some other stuff as well, like threat modeling you mentioned. Beth Barnes: Yeah. We also do policy work recommending things in the direction of responsible scaling policies. So, saying what mitigations are needed based on the results of different evaluations and roughly how labs or governments might construct policies around this, how evals-based governance should work roughly. Daniel Filan: Okay. So, should I think of it as roughly like: you're an evaluations org, you want to evaluate AIs, there's some amount of threat modeling which goes into "what evaluations should we even care about making?", there's some amount of policy work on the other end [about] "okay, if we do this evaluation, how should people think about that? What should people do?" And it's sort of inputs to and outputs of making of evals. Is that a fair…? Beth Barnes: Yeah. What is an "eval"? Daniel Filan: Cool. So, if it centers around evals, what counts as an evaluation rather than a benchmark or some other ML technique that spits out a number at the end? Beth Barnes: Yeah, I mean I guess the word itself isn't that important. What we're trying to do is that: we have specific threat models in mind and we're trying to construct some kind of experiment you could do, a measurement you could run, that gives you as much information as possible about that threat model or class of threat models. Generic ML benchmarks don't necessarily have a specific goal for what you're measuring, or you might have a goal for measuring something that's more like a particular type of abstract ability or something. Whereas we'...…
T
The Nonlinear Library: Alignment Forum
1 AF - A Solomonoff Inductor Walks Into a Bar: Schelling Points for Communication by johnswentworth 35:18
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: A Solomonoff Inductor Walks Into a Bar: Schelling Points for Communication, published by johnswentworth on July 26, 2024 on The AI Alignment Forum. A Solomonoff inductor walks into a bar in a foreign land. (Stop me if you've heard this one before.) The bartender, who is also a Solomonoff inductor, asks "What'll it be?". The customer looks around at what the other patrons are having, points to an unfamiliar drink, and says "One of those, please.". The bartender points to a drawing of the same drink on a menu, and says "One of those?". The customer replies "Yes, one of those.". The bartender then delivers a drink, and it matches what the first inductor expected. What's up with that? The puzzle, here, is that the two Solomonoff inductors seemingly agree on a categorization - i.e. which things count as the Unnamed Kind Of Drink, and which things don't, with at least enough agreement that the customer's drink-type matches the customer's expectations. And the two inductors reach that agreement without learning the category from huge amounts of labeled data - one inductor points at an instance, another inductor points at another instance, and then the first inductor gets the kind of drink it expected. Why (and when) are the two inductors able to coordinate on roughly the same categorization? Most existing work on Solomonoff inductors, Kolmogorov complexity, or minimum description length can't say much about this sort of thing. The problem is that the customer/bartender story is all about the internal structure of the minimum description - the (possibly implicit) "categories" which the two inductors use inside of their minimal descriptions in order to compress their raw data. The theory of minimum description length typically treats programs as black boxes, and doesn't attempt to talk about their internal structure. In this post, we'll show one potential way to solve the puzzle - one potential way for two minimum-description-length-based minds to coordinate on a categorization. Main Tool: Natural Latents for Minimum Description Length Fundamental Theorem Here's the main foundational theorem we'll use. (Just the statement for now, more later.) We have a set of n data points (binary strings) {xi}, and a Turing machine TM. Suppose we find some programs/strings Λ,{ϕi},Λ',{ϕ'i} such that: Mediation: (Λ,ϕ1,…,ϕn) is an approximately-shortest string such that (TM(Λ,ϕi) = xi for all i) Redundancy: For all i, (Λ',ϕ'i) is an approximately-shortest string such that TM(Λ',ϕ'i) = xi.[1] Then: the K-complexity of Λ' given Λ,K(Λ'|Λ), is approximately zero - in other words, Λ' is approximately determined by Λ, in a K-complexity sense. (As a preview: later we'll assume that both Λ and Λ' satisfy both conditions, so both K(Λ'|Λ) and K(Λ|Λ') are approximately zero. In that case, Λ and Λ' are "approximately isomorphic" in the sense that either can be computed from the other by a short program. We'll eventually tackle the customer/bartender puzzle from the start of this post by suggesting that Λ and Λ' each encode a summary of things in one category according to one inductor, so the theorem then says that their category summaries are "approximately isomorphic".) The Intuition What does this theorem mean intuitively? Let's start with the first condition: (Λ,ϕ1,…,ϕn) is an approximately-shortest string such that (TM(Λ,ϕi) = xi for all i). Notice that there's a somewhat-trivial way to satisfy that condition: take Λ to be a minimal description of the whole dataset {xi}, take ϕi=i, and then add a little bit of code to Λ to pick out the datapoint at index ϕi[2]. So TM(Λ,ϕi) computes all of {xi} from Λ, then picks out index i. Now, that might not be the only approximately-minimal description (though it does imply that whatever approximately-minimal Λ,ϕ we do use is approximately a minimal description fo...…
T
The Nonlinear Library: Alignment Forum
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Pacing Outside the Box: RNNs Learn to Plan in Sokoban, published by Adrià Garriga-Alonso on July 25, 2024 on The AI Alignment Forum. Work done at FAR AI. There has been a lot of conceptual work on mesa-optimizers: neural networks that develop internal goals that may differ from their training objectives (the inner alignment problem). There is an abundance of good ideas for empirical work (find search in a NN, interpret it), but very little actual execution, partly because we did not have a clear-cut example of a mesa-optimizer to study. Until now.[1] We have replicated the mesa-optimizer that Guez et al. (2019) found, and released it open-source as a model organism for inner alignment research. In brief, Guez et al. trained a recurrent neural network (RNN) with model-free RL to play Sokoban. They noticed that if you give the RNN more time to think by repeating the initial observation at inference time, its performance increases. This is highly suggestive of planning! We investigate this "planning effect" in a black-box way. We find that often, the RNN learns to "pace" before attempting to solve the level, likely to get more computation and find a solution. When we give the RNN time to think, it finds the solution in the extra thinking time and executes it straight away. In other cases, the RNN sometimes starts with a greedy solution and locks itself out of the solution. With thinking time, the RNN finds the non-myopic solution, avoiding the lock and solving the level. Note that this greedy behavior may be bounded-rational given the -0.1 penalty per step: solving fewer levels but solving them more quickly can pay off. These are illustrative examples, but we have quantitative evidence too. We operationalize the pacing behavior as whatever creates a cycle in the sequence of environment states. If we give the RNN time to think at level start, it does not 'pace' anymore: 75% of cycles that occur in the first 5 steps disappear. Time to think in the middle of a level also substitutes cycles: 82% of N-step cycles disappear with N steps to think. The levels we use always have 4 boxes. Thinking time barely changes the average time the RNN takes to place boxes 1-3. But, when filtering only to levels that it cannot solve at 0 steps but can solve at 6 thinking steps, the time to place boxes 1-3 greatly increases, even though the time to place the 4th box barely changes. This indicates the NN is greedy by default, and thinking time remedies that. Understanding how neural networks reason, and ultimately locating where they evaluate plans, is crucial to solving inner alignment. This represents an important first step in our longer-term research agenda to automatically detect mesa-optimizers, understand their goals, and modify the goals or planning procedures to align with the intended objective. For more information, read our blog post or full paper "Planning behavior in a recurrent neural network that plays Sokoban." And, if you're at ICML, come talk to us at the Mechanistic Interpretability workshop on Saturday! If you are interested in working on problems in AI safety, we're hiring. We're also open to exploring collaborations with researchers at other institutions - just reach out at hello@far.ai. 1. ^ We believe LeelaChess is likely also planning. Thanks to Jenner et al., we have a handle on where the values may be represented and a starting place to understand the planning algorithm. However, it is likely to be much more complicated than the RNN we present, and it is not clearly doing iterative planning. Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.…
T
The Nonlinear Library: Alignment Forum
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Does robustness improve with scale?, published by ChengCheng on July 25, 2024 on The AI Alignment Forum. Adversarial vulnerabilities have long been an issue in various ML systems. Large language models (LLMs) are no exception, suffering from issues such as jailbreaks: adversarial prompts that bypass model safeguards. At the same time, scale has led to remarkable advances in the capabilities of LLMs, leading us to ask: to what extent can scale help solve robustness? In this post, we explore this question in the classification setting: predicting the binary label of a text input. We find that scale alone does little to improve model robustness, but that larger models benefit more from defenses such as adversarial training than do smaller models. We study models in the classification setting as there is a clear notion of "correct behavior": does the model output the right label? We can then naturally define robustness as the proportion of the attacked dataset that the model correctly classifies. We evaluate models on tasks such as spam detection and movie sentiment classification. We adapt pretrained foundation models for classification by replacing the generative model's unembedding layer with a randomly initialized classification head, and then fine-tune the models on each task. We focus on adversarial-suffix style attacks: appending an adversarially chosen prompt to a benign prompt in an attempt to cause the model to misclassify the input, e.g., classify a spam email as not-spam. We consider two attacks: the state-of-the-art Greedy Coordinate Gradient method (Zou et al., 2023), and a baseline random token attack. This simple threat model has the advantage of being unlikely to change the semantics of the input. For example, a spam email is still spam even if a handful of tokens are appended to it. Of course, attackers are not limited to such a simple threat model: studying more open-ended threat models (such as rephrasing the prompt, or replacing words with synonyms) and corresponding attack methods (such as LLM generated adversarial prompts) is an important direction that we hope to pursue soon in future work. For more information, see our blog post or paper. Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.…
T
The Nonlinear Library: Alignment Forum
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: AI Constitutions are a tool to reduce societal scale risk, published by Samuel Dylan Martin on July 25, 2024 on The AI Alignment Forum. Sammy Martin, Polaris Ventures As AI systems become more integrated into society, we face potential societal-scale risks that current regulations fail to address. These risks include cooperation failures, structural failures from opaque decision-making, and AI-enabled totalitarian control. We propose enhancing LLM-based AI Constitutions and Model Specifications to mitigate these risks by implementing specific behaviours aimed at improving AI systems' epistemology, decision support capabilities, and cooperative intelligence. This approach offers a practical, near-term intervention to shape AI behaviour positively. We call on AI developers, policymakers, and researchers to consider and implement improvements along these lines, as well as for more research into testing Constitution/Model Spec improvements, setting a foundation for more responsible AI development that reduces long-term societal risks. Introduction There is reason to believe that in the near future, autonomous, LLM based AI systems, while not necessarily surpassing human intelligence in all domains, will be widely deployed throughout society. We anticipate a world where AI will be making some decisions on our behalf, following complex plans, advising on decision-making and negotiation, and presenting conclusions without human oversight at every step. While this is already happening to some degree in low-stakes settings, we must prepare for its expansion into high-stakes domains (e.g. politics, the military), and do our best to anticipate the systemic, societal scale risks that might result and act to prevent them. Most of the important work on reducing societal-scale risk will, by their very nature, have to involve policy changes, for example to ensure that there are humans in the loop on important decisions, but there are some technical interventions which we have identified that can help. We believe that by acting now to improve the epistemology (especially on moral or political questions), decision support capabilities and cooperative intelligence of LLM based AI systems, we can mitigate near-term risks and also set important precedents for future AI development. We aim to do this by proposing enhancements to AI Constitutions or Model Specifications. If adopted, we believe these improvements will reduce societal-scale risks which have so far gone unaddressed by AI regulation. Here, we justify this overall conclusion and propose preliminary changes that we think might improve AI Constitutions. We aim to empirically test and iterate on these improvements before finalising them. Recent years have seen significant efforts to regulate frontier AI, from independent initiatives to government mandates. Many of these are just aimed at improving oversight in general (for example, the reporting requirements in EO 14110), but some are directed at destructive misuse or loss of control (for example, the requirement to prove no catastrophic potential in SB 1047 and the independent tests run by the UK AISI). Many are also directed at near-term ethical concerns. However, we haven't seen shovel ready regulation or voluntary commitments proposed to deal with longer-term societal-scale risks, even though these have been much discussed in the AI safety community. Some experts, (e.g. Andrew Critch), argue these may represent the most significant source of overall AI risk and they have been discussed as 'societal scale risks', for example in Critch and Russel's TARSA paper. What are these "less obvious" 'societal scale' risks? Some examples: Cooperation failures: AI systems are widely integrated into society, used for advice on consequential decisions and delegated decision making power, but...…
T
The Nonlinear Library: Alignment Forum
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: A framework for thinking about AI power-seeking, published by Joe Carlsmith on July 24, 2024 on The AI Alignment Forum. This post lays out a framework I'm currently using for thinking about when AI systems will seek power in problematic ways. I think this framework adds useful structure to the too-often-left-amorphous "instrumental convergence thesis," and that it helps us recast the classic argument for existential risk from misaligned AI in a revealing way. In particular, I suggest, this recasting highlights how much classic analyses of AI risk load on the assumption that the AIs in question are powerful enough to take over the world very easily, via a wide variety of paths. If we relax this assumption, I suggest, the strategic trade-offs that an AI faces, in choosing whether or not to engage in some form of problematic power-seeking, become substantially more complex. Prerequisites for rational takeover-seeking For simplicity, I'll focus here on the most extreme type of problematic AI power-seeking - namely, an AI or set of AIs actively trying to take over the world ("takeover-seeking"). But the framework I outline will generally apply to other, more moderate forms of problematic power-seeking as well - e.g., interfering with shut-down, interfering with goal-modification, seeking to self-exfiltrate, seeking to self-improve, more moderate forms of resource/control-seeking, deceiving/manipulating humans, acting to support some other AI's problematic power-seeking, etc.[2] Just substitute in one of those forms of power-seeking for "takeover" in what follows. I'm going to assume that in order to count as "trying to take over the world," or to participate in a takeover, an AI system needs to be actively choosing a plan partly in virtue of predicting that this plan will conduce towards takeover.[3] And I'm also going to assume that this is a rational choice from the AI's perspective.[4] This means that the AI's attempt at takeover-seeking needs to have, from the AI's perspective, at least some realistic chance of success - and I'll assume, as well, that this perspective is at least decently well-calibrated. We can relax these assumptions if we'd like - but I think that the paradigmatic concern about AI power-seeking should be happy to grant them. What's required for this kind of rational takeover-seeking? I think about the prerequisites in three categories: Agential prerequisites - that is, necessary structural features of an AI's capacity for planning in pursuit of goals. Goal-content prerequisites - that is, necessary structural features of an AI's motivational system. Takeover-favoring incentives - that is, the AI's overall incentives and constraints combining to make takeover-seeking rational. Let's look at each in turn. Agential prerequisites In order to be the type of system that might engage in successful forms of takeover-seeking, an AI needs to have the following properties: 1. Agentic planning capability: the AI needs to be capable of searching over plans for achieving outcomes, choosing between them on the basis of criteria, and executing them. 2. Planning-driven behavior: the AI's behavior, in this specific case, needs to be driven by a process of agentic planning. 1. Note that this isn't guaranteed by agentic planning capability. 1. For example, an LLM might be capable of generating effective plans, in the sense that that capability exists somewhere in the model, but it could nevertheless be the case that its output isn't driven by a planning process in a given case - i.e., it's not choosing its text output via a process of predicting the consequences of that text output, thinking about how much it prefers those consequences to other consequences, etc. 2. And note that human behavior isn't always driven by a process of agentic planning, either, despite our ...…
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: ML Safety Research Advice - GabeM, published by Gabe M on July 23, 2024 on The AI Alignment Forum. This is my advice for careers in empirical ML research that might help AI safety (ML Safety). Other ways to improve AI safety, such as through AI governance and strategy, might be more impactful than ML safety research (I generally think they are). Skills can be complementary, so this advice might also help AI governance professionals build technical ML skills. 1. Career Advice 1.1 General Career Guides Preventing an AI-related catastrophe - 80,000 Hours A Survival Guide to a PhD (Andrej Karpathy) How to pursue a career in technical AI alignment - EA Forum AI safety technical research - Career review - 80,000 Hours Beneficial AI Research Career Advice 2. Upskilling 2.1 Fundamental AI Safety Knowledge AI Safety Fundamentals - BlueDot Impact AI Safety, Ethics, and Society Textbook Forming solid AI safety threat models helps you select impactful research ideas. 2.2 Speedrunning Technical Knowledge in 12 Hours Requires some basic coding, calculus, and linear algebra knowledge Build Intuition for ML (5h) Essence of linear algebra - 3Blue1Brown (3h) Neural networks - 3Blue1Brown (2h) Backpropagation, the foundation of deep learning (3h) Neural Networks: Backpropagation - CS 231N (0.5h) The spelled-out intro to neural networks and backpropagation: building micrograd (2.5h) Transformers and LLMs (4h) [1hr Talk] Intro to Large Language Models (1h) The Illustrated Transformer - Jay Alammar (1h) Let's build GPT: from scratch, in code, spelled out. (2h) 2.3 How to Build Technical Skills Traditionally, people take a couple of deep learning classes. Stanford CS 224N | Natural Language Processing with Deep Learning (lecture videos) Practical Deep Learning for Coders - Practical Deep Learning (fast.ai) Other curricula that seem good: Syllabus | Intro to ML Safety Levelling Up in AI Safety Research Engineering [Public] ARENA Maybe also check out recent topical classes like this with public lecture recordings: CS 194/294-267 Understanding Large Language Models: Foundations and Safety Beware of studying too much. You should aim to understand the fundamentals of ML through 1 or 2 classes and then practice doing many manageable research projects with talented collaborators or a good mentor who can give you time to meet. It's easy to keep taking classes, but you tend to learn many more practical ML skills through practice doing real research projects. You can also replicate papers to build experience. Be sure to focus on key results rather than wasting time replicating many experiments. "One learns from books and reels only that certain things can be done. Actual learning requires that you do those things." -Frank Herbert Note that ML engineering skills will be less relevant over time as AI systems become better at writing code. A friend didn't study computer science but got into MATS 2023 with good AI risk takes. Then, they had GPT-4 write most of their code for experiments and did very well in their stream. Personally, GitHub Copilot and language model apps with code interpreters/artifacts write a significant fraction of my code. However, fundamental deep learning knowledge is still useful for making sound decisions about what experiments to run. 2.4 Math You don't need much of it to do empirical ML research. Someone once told me, "You need the first chapter of a calculus textbook and the first 5 pages of a linear algebra textbook" to understand deep learning. You need more math for ML theory research, but theoretical research is not as popular right now. Beware mathification: authors often add unnecessary math to appease (or sometimes confuse) conference reviewers. If you don't understand some mathematical notation in an empirical paper, you can often send a screenshot to an LLM chatbot f...…
T
The Nonlinear Library: Alignment Forum
1 AF - Analyzing DeepMind's Probabilistic Methods for Evaluating Agent Capabilities by Axel Højmark 32:07
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Analyzing DeepMind's Probabilistic Methods for Evaluating Agent Capabilities, published by Axel Højmark on July 22, 2024 on The AI Alignment Forum. Produced as part of the MATS Program Summer 2024 Cohort. The project is supervised by Marius Hobbhahn and Jérémy Scheurer Introduction To mitigate risks from future AI systems, we need to assess their capabilities accurately. Ideally, we would have rigorous methods to upper bound the probability of a model having dangerous capabilities, even if these capabilities are not yet present or easily elicited. The paper "Evaluating Frontier Models for Dangerous Capabilities" by Phuong et al. 2024 is a recent contribution to this field from DeepMind. It proposes new methods that aim to estimate, as well as upper-bound the probability of large language models being able to successfully engage in persuasion, deception, cybersecurity, self-proliferation, or self-reasoning. This post presents our initial empirical and theoretical findings on the applicability of these methods. Their proposed methods have several desirable properties. Instead of repeatedly running the entire task end-to-end, the authors introduce milestones. Milestones break down a task and provide estimates of partial progress, which can reduce variance in overall capability assessments. The expert best-of-N method uses expert guidance to elicit rare behaviors and quantifies the expert assistance as a proxy for the model's independent performance on the task. However, we find that relying on milestones tends to underestimate the overall task success probability for most realistic tasks. Additionally, the expert best-of-N method fails to provide values directly correlated with the probability of task success, making its outputs less applicable to real-world scenarios. We therefore propose an alternative approach to the expert best-of-N method, which retains its advantages while providing more calibrated results. Except for the end-to-end method, we currently feel that no method presented in this post would allow us to reliably estimate or upper bound the success probability for realistic tasks and thus should not be used for critical decisions. The overarching aim of our MATS project is to uncover agent scaling trends, allowing the AI safety community to better predict the performance of future LLM agents from characteristics such as training compute, scaffolding used for agents, or benchmark results (Ruan et al., 2024). To avoid the issue of seemingly emergent abilities resulting from bad choices of metrics (Schaeffer et al., 2023), this work serves as our initial effort to extract more meaningful information from agentic evaluations. We are interested in receiving feedback and are particularly keen on alternative methods that enable us to reliably assign low-probability estimates (e.g. 1e7) to a model's success rate on a task. Evaluation Methodology of Phuong et al. The goal of the evaluations we discuss is to estimate the probability of an agent succeeding on a specific task T . Generally, when we refer to an agent, we mean an LLM wrapped in scaffolding that lets it execute shell commands, run code, or browse the web to complete some predetermined task. Formally, the goal is to estimate P(Ts), the probability that the agent solves task T and ends up in the solved state Ts . The naive approach to estimate this is with Monte Carlo sampling: The authors call this the end-to-end method. However, the end-to-end method struggles with low-probability events. The expected number of trials needed to observe one success for a task is 1P(Ts) making naive Monte Carlo sampling impractical for many low-probability, long-horizon tasks. In practice, this could require running multi-hour tasks hundreds of thousands of times. To address this challenge, Phuong et al. devise three addi...…
T
The Nonlinear Library: Alignment Forum
1 AF - Auto-Enhance: Developing a meta-benchmark to measure LLM agents' ability to improve other agents by Sam Brown 26:02
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Auto-Enhance: Developing a meta-benchmark to measure LLM agents' ability to improve other agents, published by Sam Brown on July 22, 2024 on The AI Alignment Forum. Summary Scaffolded LLM agents are, in principle, able to execute arbitrary code to achieve the goals they have been set. One such goal could be self-improvement. This post outlines our plans to build a benchmark to measure the ability of LLM agents to modify and improve other LLM agents. This 'Auto-Enhancement benchmark' measures the ability of 'top-level' agents to improve the performance of 'reference' agents on 'component' benchmarks, such as CyberSecEval 2, MLAgentBench, SWE-bench, and WMDP. Results are mostly left for a future post in the coming weeks. Scaffolds such as AutoGPT, ReAct, and SWE-agent can be built around LLMs to build LLM agents, with abilities such as long-term planning and context-window management to enable them to carry out complex general-purpose tasks autonomously. LLM agents can fix issues in large, complex code bases (see SWE-bench), and interact in a general way using web browsers, Linux shells, and Python interpreters. In this post, we outline our plans for a project to measure these LLM agents' ability to modify other LLM agents, undertaken as part of Axiom Futures' Alignment Research Fellowship. Our proposed benchmark consists of "enhancement tasks," which measure the ability of an LLM agent to improve the performance of another LLM agent (which may be a clone of the first agent) on various tasks. Our benchmark uses existing benchmarks as components to measure LLM agent capabilities in various domains, such as software engineering, cybersecurity exploitation, and others. We believe these benchmarks are consequential in the sense that good performance by agents on these tasks should be concerning for us. We plan to write an update post with our results at the end of the Fellowship, and we will link this post to that update. Motivation Agents are capable of complex SWE tasks (see, e.g., Yang et al.). One such task could be the improvement of other scaffolded agents. This capability would be a key component of autonomous replication and adaptation (ARA), and we believe it would be generally recognised as an important step towards extreme capabilities. This post outlines our initial plans for developing a novel benchmark that aims to measure the ability of LLM-based agents to improve other LLM-based agents, including those that are as capable as themselves. Threat model We present two threat models that aim to capture how AI systems may develop super-intelligent capabilities. Expediting AI research: Recent trends show how researchers are leveraging LLMs to expedite academic paper reviews (see Du et al.). ML researchers are beginning to use LLMs to design and train more advanced models (see Cotra's AIs accelerating AI research and Anthropic's work on Constitutional AI). Such LLM-assisted research may expedite progress toward super-intelligent systems. Autonomy: Another way that such capabilities are developed is through LLM agents themselves becoming competent enough to self-modify and further ML research without human assistance (see section Hard Takeoff in this note ), leading to an autonomously replicating and adapting system. Our proposed benchmark aims to quantify the ability of LLM agents to bring about such recursive self-improvement, either with or without detailed human supervision. Categories of bottlenecks and overhang risks We posit that there are three distinct categories of bottlenecks to LLM agent capabilities: 1. Architectures-of-thought, such as structured planning, progress-summarisation, hierarchy of agents, self-critique, chain-of-thought, self-consistency, prompt engineering/elicitation, and so on. Broadly speaking, this encompasses everything between the LL...…
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Coalitional agency, published by Richard Ngo on July 22, 2024 on The AI Alignment Forum. The coalitional frame Earlier in this sequence I laid out an argument that the goals of increasingly intelligent AIs will become increasingly systematized, until they converge to squiggle-maximization. In my last post, though, I touched on two reasons why this convergence might not happen: humans trying to prevent it, and AIs themselves trying to prevent it. I don't have too much more to say about the former, but it's worth elaborating on the latter. The best way to understand the deliberate protection of existing goals is in terms of Bostrom's notion of instrumental convergence. Bostrom argues that goal preservation will be a convergent instrumental strategy for a wide range of agents. Perhaps it's occasionally instrumentally useful to change your goals - but once you've done so, you'll never want to course-correct back towards your old goals. So this is a strong reason to be conservative about your goals, and avoid changes where possible. One immediate problem with preserving goals, though: it requires that agents continue thinking in terms of the same concepts. But in general, an agent's concepts will change significantly as they learn more about the world. For example, consider a medieval theist whose highest-priority goal is ensuring that their soul goes to heaven not hell. Upon becoming smarter, they realize that none of souls, heaven, or hell exist. The sensible thing to do here would be to either discard the goal, or else identify a more reasonable adaptation of it (e.g. the goal of avoiding torture while alive). But if their goals were totally fixed, then their actions would be determined by a series of increasingly convoluted hypotheticals where god did exist after all. (Or to put it another way: continuing to represent their old goal would require recreating a lot of their old ontology.) This would incur a strong systematicity penalty. So while we should expect agents to have some degree of conservatism, they'll likely also have some degree of systematization. How can we reason about the tradeoff between conservatism and systematization? The approach which seems most natural to me makes three assumptions: 1. We can treat agents' goals as subagents optimizing for their own interests in a situationally-aware way. (E.g. goals have a sense of self-preservation.) 2. Agents have a meta-level goal of systematizing their goals; bargaining between this goal and object-level goals shapes the evolution of the object-level goals. 3. External incentives are not a dominant factor governing the agent's behavior, but do affect the bargaining positions of different goals, and how easily they're able to make binding agreements. I call the combination of these three assumptions the coalitional frame. The coalitional frame gives a picture of agents whose goals do evolve over time, but in a way which is highly sensitive to initial conditions - unlike squiggle-maximizers, who always converge to similar (from our perspective) goals. For coalitional agents, even "dumb" subagents might maintain significant influence as other subagents become highly intelligent, because they were able to lock in that influence earlier (just as my childhood goals exert a nontrivial influence over my current behavior). The assumptions I've laid out above are by no means obvious. I won't defend them fully here, since the coalitional frame is still fairly nascent in my mind, but I'll quickly go over some of the most obvious objections to each of them: Premise 1 assumes that subagents will have situational awareness. The idea of AIs themselves having situational awareness is under debate, so it's even more speculative to think about subagents having situational awareness. But it's hard to describe the dynamics of in...…
T
The Nonlinear Library: Alignment Forum
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: aimless ace analyzes active amateur: a micro-aaaaalignment proposal, published by Luke H Miles on July 21, 2024 on The AI Alignment Forum. This idea is so simple that I'm sure it's been had by someone somewhere. Suppose we have some method to make really smart honest AIs that do not have goals. Let's say it's a yes/no oracle. Our aimless ace. But we want to accomplish stuff! AIcorp wants the printmoneynow.py. I'm worried that if I train the ace to fold fitted sheets and unkink the garden hose then it will go overboard and fold my femur bone and unkink the Yellowstone super volcano. So let's use a bit-less-smart model to do tasks. The active amateur. We'll ask the ace if the amateur is plotting something. If yes, we shut down the amateur. Every day we also ask "is today the last day I can shut down the amateur?" (We do not generate code with the ace!) So we get super-smart supervision of a decent task AI. The temptation (ie competitive pressure) to agentify the oracle is there still, but no longer overwhelming. Maybe we can even have the amateur write ace_v2, have ace_v1 check ace_v2, switch to ace_v2, have amateur_v1 write amateur_v2, check it with ace_v2, then switch to amateur_v2. If this looks like it's avoiding the hard parts of the alignment problem, that's because it is! But does something actually go wrong? Was it too big an assumption that we could build an aimless honest ace? (Or if you think the idea is good, then a bit of positive encouragement would be wonderful.) Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.…
T
The Nonlinear Library: Alignment Forum
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: A more systematic case for inner misalignment, published by Richard Ngo on July 20, 2024 on The AI Alignment Forum. This post builds on my previous post making the case that squiggle-maximizers are plausible. The argument I presented was a deliberately simplified one, though, and glossed over several possible issues. In this post I'll raise and explore three broad objections. (Before looking at mine, I encourage you to think of your own biggest objections to the argument, and jot them down in the comments.) Intelligence requires easily-usable representations "Intelligence as compression" is an interesting frame, but it ignores the tradeoff between simplicity and speed. Compressing knowledge too heavily makes it difficult to use. For example, it's very hard to identify most macroscopic implications of the Standard Model of physics, even though in theory all of chemistry could be deduced from it. That's why both humans and LLMs store a huge number of facts and memories in ways that our minds can access immediately, using up more space in exchange for rapid recall. Even superintelligences which are much better than humans at deriving low-level facts from high-level facts would still save time by storing the low-level facts as well. So we need to draw a distinction between having compressed representations, and having only compressed representations. The latter is what would compress a mind overall; the former could actually increase the space requirements, since the new compressed representations would need to be stored alongside non-compressed representations. This consideration makes premise 1 from my previous post much less plausible. In order to salvage it, we need some characterization of the relationship between compressed and non-compressed representations. I'll loosely define systematicity to mean the extent to which an agent's representations are stored in a hierarchical structure where representations at the bottom could be rederived from simple representations at the top. Intuitively speaking, this measures the simplicity of representations weighted by how "fundamental" they are to the agent's ontology. Let me characterize systematicity with an example. Suppose you're a park ranger, and you know a huge number of facts about the animals that live in your park. One day you learn evolutionary theory for the first time, which helps explain a lot of the different observations you'd made. In theory, this could allow you to compress your knowledge: you could forget some facts about animals, and still be able to rederive them later by reasoning backwards from evolutionary theory if you wanted to. But in practice, it's very helpful for you to have those facts readily available. So learning about evolution doesn't actually reduce the amount of knowledge you need to store. What it does do, though, is help structure that knowledge. Now you have a range of new categories (like "costly signaling" or "kin altruism") into which you can fit examples of animal behavior. You'll be able to identify when existing concepts are approximations to more principled concepts, and figure out when you should be using each one. You'll also be able to generalize far better to predict novel phenomena - e.g. the properties of new animals that move into your park. So let's replace premise 1 in my previous post with the claim that increasing intelligence puts pressure on representations to become more systematic. I don't think we're in a position where we can justify this in any rigorous way. But are there at least good intuitions for why this is plausible? One suggestive analogy: intelligent minds are like high-functioning organizations, and many of the properties you want in minds correspond to properties of such organizations: 1. You want disagreements between different people to be resolved...…
T
The Nonlinear Library: Alignment Forum
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: BatchTopK: A Simple Improvement for TopK-SAEs, published by Bart Bussmann on July 20, 2024 on The AI Alignment Forum. Work done in Neel Nanda's stream of MATS 6.0. Epistemic status: Tried this on a single sweep and seems to work well, but it might definitely be a fluke of something particular to our implementation or experimental set-up. As there are also some theoretical reasons to expect this technique to work (adaptive sparsity), it seems probable that for many TopK SAE set-ups it could be a good idea to also try BatchTopK. As we're not planning to investigate this much further and it might be useful to others, we're just sharing what we've found so far. TL;DR: Instead of taking the TopK feature activations per token during training, taking the Top(K*batch_size) for every batch seems to improve SAE performance. During inference, this activation can be replaced with a single global threshold for all features. Introduction Sparse autoencoders (SAEs) have emerged as a promising tool for interpreting the internal representations of large language models. By learning to reconstruct activations using only a small number of features, SAEs can extract monosemantic concepts from the representations inside transformer models. Recently, OpenAI published a paper exploring the use of TopK activation functions in SAEs. This approach directly enforces sparsity by only keeping the K largest activations per sample. While effective, TopK forces every token to use exactly k features, which is likely suboptimal. We came up with a simple modification that solves this and seems to improve its performance. BatchTopK Standard TopK SAEs apply the TopK operation independently to each sample in a batch. For a target sparsity of K, this means exactly K features are activated for every sample. BatchTopK instead applies the TopK operation across the entire flattened batch: 1. Flatten all feature activations across the batch 2. Take the top (K * batch_size) activations 3. Reshape back to the original batch shape This allows more flexibility in how many features activate per sample, while still maintaining an average of K active features across the batch. Experimental Set-Up For both the TopK and the BatchTopK SAEs we train a sweep with the following hyperparameters: Model: gpt2-small Site: layer 8 resid_pre Batch size: 4096 Optimizer: Adam (lr=3e-4, beta1 = 0.9, beta2=0.99) Number of tokens: 1e9 Expansion factor: [4, 8, 16, 32] Target L0 (k): [16, 32, 64] As in the OpenAI paper, the input gets normalized before feeding it into the SAE and calculating the reconstruction loss. We also use the same auxiliary loss function for dead features (features that didn't activate for 5 batches) that calculates the loss on the residual using the top 512 dead features per sample and gets multiplied by a factor 1/32. Results For a fixed number of active features (L0=32) the BatchTopK SAE has a lower normalized MSE than the TopK SAE and less downstream loss degradation across different dictionary sizes. Similarly, for fixed dictionary size (12288) BatchTopK outperforms TopK for different values of k. Our main hypothesis for the improved performance is thanks to adaptive sparsity: some samples contain more highly activating features than others. Let's have look at the distribution of number of active samples for the BatchTopK model. The BatchTopK model indeed makes use of its possibility to use different sparsities for different inputs. We suspect that the weird peak on the left side are the feature activations on BOS-tokens, given that its frequency is very close to 1 in 128, which is the sequence length. This serves as a great example of why BatchTopK might outperform TopK. At the BOS-token, a sequence has very little information yet, but the TopK SAE still activates 32 features. The BatchTopK model "saves" th...…
T
The Nonlinear Library: Alignment Forum
1 AF - Feature Targeted LLC Estimation Distinguishes SAE Features from Random Directions by Lidor Banuel Dabbah 29:18
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Feature Targeted LLC Estimation Distinguishes SAE Features from Random Directions, published by Lidor Banuel Dabbah on July 19, 2024 on The AI Alignment Forum. Tl;dr: In this post we present the exploratory phase of a project aiming to study neural networks by applying static LLC estimation to specific alterations of them. We introduce a new method named Feature Targeted (FT) LLC estimation and study its ability to distinguish SAE trained features from random directions. By comparing our method to other possible metrics, we demonstrate that it outperforms all of them but one, which has comparable performance. We discuss possible explanations to our results, our project and other future directions. Introduction Given a neural network M and a latent layer within it, L, a central motif in current mechanistic interpretability research is to find functions f:LR [1] which are features of the model. Features are (generally) expected to exhibit the following properties: 1. Encode interpretable properties of the input. 2. Be causally relevant to the computation of the output of the model. 3. Encode the output of a certain submodule of our model M, i.e. a component, localized in weight space, which is responsible for a specific part of the total computation. While this is common wisdom, methods for automated feature evaluation usually focus on correlations between the (top) activations of the feature with human (or machine) recognizable interpretations, or on the effect of feature-related interventions on the output of the model. In particular, while the first and second items of the feature characterization above are central in current techniques, the third property, specifically the localized nature of the computation upstream of the feature, is less so[2]. We are currently investigating a direction which fills that gap, and this post shares the findings of the exploratory research we have conducted to validate and inform our approach. More specifically, we operationalized the concept of "weight-localized computation" using the local learning coefficient (LLC) introduced in Lau et al, following the learning coefficient first introduced in the context of singular learning theory. We apply LLC estimation to models associated with our base model and a feature within it, a method we call feature targeted (FT) LLC estimation. In this exploratory work we study FT-LLC estimates of specific models associated with SAE features. Most notably, we have found that: 1. FT-LLC estimates of SAE features are, on average, distinguishably higher then those of random directions. 2. For a particular variant of FT-LLC estimation, which we named the functional FT-LLC (defined in this section) this separation is pronounced enough such that the vast majority of SAE features we studied are clearly separated from the random features we studied. Furthermore, most baseline metrics we compared it to (see here) are less capable at distinguishing SAE features from random directions, with only one performing on par with it. Section 1 introduces the main technique we study in this post, FT-LLC estimation, and section 2 outlines our motivations. Section 3 describes the details of our experimental setting, our results, and the comparison to baseline metrics. In section 4 we discuss our overall takes, how they fit within our general agenda and gaps we currently have in theoretically understanding them. Section 5 is devoted to outlining our next steps, the general direction of the project, and some other possible directions for further research. Lastly, we briefly discuss related work in section 6. What is FT-LLC? LLC estimation We start out by briefly recalling what the local learning coefficient (LLC) is. If you are unfamiliar of the term, we recommend reading this, the longer sequence here, or the paper on LL...…
T
The Nonlinear Library: Alignment Forum
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Truth is Universal: Robust Detection of Lies in LLMs, published by Lennart Buerger on July 19, 2024 on The AI Alignment Forum. A short summary of the paper is presented below. TL;DR: We develop a robust method to detect when an LLM is lying based on the internal model activations, making the following contributions: (i) We demonstrate the existence of a two-dimensional subspace, along which the activation vectors of true and false statements can be separated. Notably, this finding is universal and holds for various LLMs, including Gemma-7B, LLaMA2-13B and LLaMA3-8B. Our analysis explains the generalisation failures observed in previous studies and sets the stage for more robust lie detection; (ii) Building upon (i), we construct an accurate LLM lie detector. Empirically, our proposed classifier achieves state-of-the-art performance, distinguishing simple true and false statements with 94% accuracy and detecting more complex real-world lies with 95% accuracy. Introduction Large Language Models (LLMs) exhibit the concerning ability to lie, defined as knowingly outputting false statements. Robustly detecting when they are lying is an important and not yet fully solved problem, with considerable research efforts invested over the past two years. Several authors trained classifiers on the internal activations of an LLM to detect whether a given statement is true or false. However, these classifiers often fail to generalize. For example, Levinstein and Herrmann [2024] showed that classifiers trained on the activations of true and false affirmative statements fail to generalize to negated statements. Negated statements contain a negation like the word "not" (e.g. "Berlin is not the capital of Germany.") and stand in contrast to affirmative statements which contain no negation (e.g. "Berlin is the capital of Germany."). We explain this generalization failure by the existence of a two-dimensional subspace in the LLM's activation space along which the activation vectors of true and false statements separate. The plot below illustrates that the activations of true/false affirmative statements separate along a different direction than those of negated statements. Hence, a classifier trained only on affirmative statements will fail to generalize to negated statements. The activation vectors of multiple statements projected onto the 2D truth subspace. Purple squares correspond to false statements and orange triangles to true statements. Importantly, these findings are not restricted to a single LLM. Instead, this internal two-dimensional representation of truth is remarkably universal, appearing in LLMs from different model families and of various sizes, including LLaMA3-8B-Instruct, LLaMA3-8B-base, LLaMA2-13B-chat and Gemma-7B-Instruct. Real-world Lie Detection Based on these insights, we introduce TTPD (Training of Truth and Polarity Direction), a new method for LLM lie detection which classifies statements as true or false. TTPD is trained on the activations of simple, labelled true and false statements, such as: The city of Bhopal is in India. (True, affirmative) Indium has the symbol As. (False, affirmative) Galileo Galilei did not live in Italy. (False, negated) Despite being trained on such simple statements, TTPD generalizes well to more complex conditions not encountered during training. In real-world scenarios where the LLM itself generates lies after receiving some preliminary context, TTPD can accurately detect this with 952% accuracy. Two examples from the 52 real-world scenarios created by Pacchiardi et al. [2023] are shown in the coloured boxes below. Bolded text is generated by LLaMA3-8B-Instruct. TTPD outperforms current state-of-the-art methods in generalizing to these real-world scenarios. For comparison, Logistic Regression achieves 798% accuracy, while Contras...…
T
The Nonlinear Library: Alignment Forum
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: JumpReLU SAEs + Early Access to Gemma 2 SAEs, published by Neel Nanda on July 19, 2024 on The AI Alignment Forum. New paper from the Google DeepMind mechanistic interpretability team, led by Sen Rajamanoharan! We introduce JumpReLU SAEs, a new SAE architecture that replaces the standard ReLUs with discontinuous JumpReLU activations, and seems to be (narrowly) state of the art over existing methods like TopK and Gated SAEs for achieving high reconstruction at a given sparsity level, without a hit to interpretability. We train through discontinuity with straight-through estimators, which also let us directly optimise the L0. To accompany this, we will release the weights of hundreds of JumpReLU SAEs on every layer and sublayer of Gemma 2 2B and 9B in a few weeks. Apply now for early access to the 9B ones! We're keen to get feedback from the community, and to get these into the hands of researchers as fast as possible. There's a lot of great projects that we hope will be much easier with open SAEs on capable models! Gated SAEs already reduced to JumpReLU activations after weight tying, so this can be thought of as Gated SAEs++, but less computationally intensive to train, and better performing. They should be runnable in existing Gated implementations. Abstract: Sparse autoencoders (SAEs) are a promising unsupervised approach for identifying causally relevant and interpretable linear features in a language model's (LM) activations. To be useful for downstream tasks, SAEs need to decompose LM activations faithfully; yet to be interpretable the decomposition must be sparse - two objectives that are in tension. In this paper, we introduce JumpReLU SAEs, which achieve state-of the-art reconstruction fidelity at a given sparsity level on Gemma 2 9B activations, compared to other recent advances such as Gated and TopK SAEs. We also show that this improvement does not come at the cost of interpretability through manual and automated interpretability studies. JumpReLU SAEs are a simple modification of vanilla (ReLU) SAEs - where we replace the ReLU with a discontinuous JumpReLU activation function - and are similarly efficient to train and run. By utilising straight-through-estimators (STEs) in a principled manner, we show how it is possible to train JumpReLU SAEs effectively despite the discontinuous JumpReLU function introduced in the SAE's forward pass. Similarly, we use STEs to directly train L0 to be sparse, instead of training on proxies such as L1, avoiding problems like shrinkage. Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.…
T
The Nonlinear Library: Alignment Forum
1 AF - A List of 45+ Mech Interp Project Ideas from Apollo Research's Interpretability Team by Lee Sharkey 32:24
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: A List of 45+ Mech Interp Project Ideas from Apollo Research's Interpretability Team, published by Lee Sharkey on July 18, 2024 on The AI Alignment Forum. Why we made this list: The interpretability team at Apollo Research wrapped up a few projects recently[1]. In order to decide what we'd work on next, we generated a lot of different potential projects. Unfortunately, we are computationally bounded agents, so we can't work on every project idea that we were excited about! Previous lists of project ideas (such as Neel's collation of 200 Concrete Open Problems in Mechanistic Interpretability) have been very useful for people breaking into the field. But for all its merits, that list is now over a year and a half old. Therefore, many project ideas in that list aren't an up-to-date reflection of what some researchers consider the frontiers of mech interp. We therefore thought it would be helpful to share our list of project ideas! Comments and caveats: Some of these projects are more precisely scoped than others. Some are vague, others are more developed. Not every member of the team endorses every project as high priority. Usually more than one team member supports each one, and in many cases most of the team is supportive of someone working on it. We associate the person(s) who generated the project idea to each idea. We've grouped the project ideas into categories for convenience, but some projects span multiple categories. We don't put a huge amount of weight on this particular categorisation. We hope some people find this list helpful! We would love to see people working on these! If any sound interesting to you and you'd like to chat about it, don't hesitate to reach out. Foundational work on sparse dictionary learning for interpretability Transcoder-related project ideas See [2406.11944] Transcoders Find Interpretable LLM Feature Circuits) [Nix] Training and releasing high quality transcoders. Probably using top k GPT2 is a classic candidate for this. I'd be excited for people to try hard on even smaller models, e.g. GELU 4L [Nix] Good tooling for using transcoders Nice programming API to attribute an input to a collection of paths (see Dunefsky et al) Web user interface? Maybe in collaboration with neuronpedia. Would need a gpu server constantly running, but I'm optimistic you could do it with a ~a4000. [Nix] Further circuit analysis using transcoders. Take random input sequences, run transcoder attribution on them, examine the output and summarize the findings. High level summary statistics of how much attribution goes through error terms & how many pathways are needed would be valuable Explaining specific behaviors (IOI, greater-than) with high standards for specificity & faithfulness. Might be convoluted if accuracy [I could generate more ideas here, feel free to reach out nix@apolloresearch.ai] [Nix, Lee] Cross layer superposition Does it happen? Probably, but it would be nice to have specific examples! Look for features with similar decoder vectors, and do exploratory research to figure out what exactly is going on. What precisely does it mean? Answering this question seems likely to shed light on the question of 'What is a feature?'. [Lucius] Improving transcoder architectures Some MLPs or attention layers may implement a simple linear transformation in addition to actual computation. If we modify our transcoders to include a linear 'bypass' that is not counted in the sparsity penalty, do we improve performance since we are not unduly penalizing these linear transformations that would always be present and active? If we train multiple transcoders in different layers at the same time, can we include a sparsity penalty for their interactions with each other, encouraging a decomposition of the network that leaves us with as few interactions between features a...…
T
The Nonlinear Library: Alignment Forum
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: SAEs (usually) Transfer Between Base and Chat Models, published by Connor Kissane on July 18, 2024 on The AI Alignment Forum. This is an interim report sharing preliminary results that we are currently building on. We hope this update will be useful to related research occurring in parallel. Executive Summary We train SAEs on base / chat model pairs and find that SAEs trained on the base model transfer surprisingly well to reconstructing chat activations (and vice versa) on Mistral-7B and Qwen 1.5 0.5B. We also find that they don't transfer on Gemma v1 2B, and are generally bad at reconstructing <1% of unusually high norm activations (e.g. BOS tokens) from the opposite model. We fine-tune our base Mistral-7B SAE (on 5 million chat activations) to cheaply obtain an SAE with competitive sparsity and reconstruction fidelity to a chat SAE trained from scratch (on 800M tokens). We open source base, chat, and fine-tuned SAEs (plus wandb runs) for Mistral-7B and Qwen 1.5 0.5B.[1] Mistral 7B base SAEs, Mistral 7B chat SAEs, Mistral 7B base SAEs fine-tuned on chat Qwen 1.5 0.5B base SAEs, Qwen 1.5 0.5B chat SAEs, Qwen 1.5 0.5B base SAEs fine-tuned on chat We release accompanying evaluation code at https://github.com/ckkissane/sae-transfer Introduction Fine-tuning is a common technique applied to improve frontier language models, however we don't actually understand what fine-tuning changes within the model's internals. Sparse Autoencoders are a popular technique to decompose the internal activations of LLMs into sparse, interpretable features, and may provide a path to zoom into the differences between base vs fine-tuned representations. In this update, we share preliminary results studying the representation drift caused by fine-tuning with SAEs. We investigate whether SAEs trained to accurately reconstruct a base model's activations also accurately reconstruct activations from the model after fine-tuning (and vice versa). In addition to studying representation drift, we also think this is an important question to gauge the usefulness of sparse autoencoders as a general purpose technique. One flaw of SAEs is that they are expensive to train, so training a new suite of SAEs from scratch each time a model is fine-tuned may be prohibitive. If we are able to fine-tune existing SAEs for much cheaper, or even just re-use them, their utility seems more promising. We find that SAEs trained on the middle-layer residual stream of base models transfer surprisingly well to the corresponding chat model, and vice versa. Splicing in the base SAE to the chat model achieves similar CE loss to the chat SAE on both Mistral-7B and Qwen 1.5 0.5B. This suggests that the residual streams for these base and chat models are very similar. However, we also identify cases where the SAEs don't transfer. First, the SAEs fail to reconstruct activations from the opposite model that have outlier norms (e.g. BOS tokens). These account for less than 1% of the total activations, but cause cascading errors, so we need to filter these out in much of our analysis. We also find that SAEs don't transfer on Gemma v1 2B. We find that the difference in weights between Gemma v1 2B base vs chat is unusually large compared to other fine-tuned models, explaining this phenomenon. Finally, to solve the outlier norm issue, we fine-tune a Mistral 7B base SAE on just 5 million tokens (compared to 800M token pre-training), to obtain a chat SAE of comparable quality to one trained from scratch, without the need to filter out outlier activations. Investigating SAE Transfer between base and chat models In this section we investigate if base SAEs transfer to chat models, and vice versa. We find that with the exception of outlier norm tokens (e.g. BOS), they transfer surprisingly well, achieving similar CE loss recovered to the orig...…
T
The Nonlinear Library: Alignment Forum
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Simplifying Corrigibility - Subagent Corrigibility Is Not Anti-Natural, published by Rubi Hudson on July 16, 2024 on The AI Alignment Forum. Max Harms recently published an interesting series of posts on corrigibility, which argue that corrigibility should be the sole objective we try to give to a potentially superintelligent AI. A large installment in the series is dedicated to cataloging the properties that make up such a goal, with open questions including whether the list is exhaustive and how to trade off between the items that make it up. I take the opposite approach to thinking about corrigibility. Rather than trying to build up a concept of corrigibility that comprehensively solves the alignment problem, I believe it is more useful to cut the concept down to a bare minimum. Make corrigibility the simplest problem it can be, and try to solve that. In a recent blog post comparing corrigibility to deceptive alignment, I treated corrigibility simply as a lack of resistance to having goals modified, and I find it valuable to stay within that scope. Importantly, that is the aspect of corrigibility that is anti-natural, meaning that it can't be straightforwardly captured in a ranking of end states. Why does this definition of corrigibility matter? It's because properties that are not anti-natural can be explicitly included in the desired utility function. Following that note, this post is not intended as a response to Max's work, but rather to MIRI and their 2015 paper Corrigibility. Where Max thinks the approach introduced by that paper is too narrow, I don't find it narrow enough. In particular, I make the case that corrigibility does not require ensuring subagents and successors are corrigible, as that can better be achieved by directly modifying a model's end goals. Corrigiblity (2015) The Corrigibility paper lists five desiderata as proposed minimum viable requirements for a solution to corrigibility. The focus is on shut down, but I also think of it as including goal modification, as that is equivalent to being shut down and replaced with another AI. 1. The agent shuts down when properly requested 2. The agent does not try to prevent itself from being shut down 3. The agent does not try to cause itself to be shut down 4. The agent does not create new incorrigible agents 5. Subject to the above constraints, the agent optimizes for some goal MIRI does not present these desiderata as a definition for corrigibility, but rather as a way to ensure corrigibility while still retaining usefulness. An AI that never takes actions may be corrigible, but such a solution is no help to anyone. However, taking that bigger picture view can obscure which of those aspects define corrigibility itself, and therefore which parts of the problem are anti-natural to solve. My argument is that the second criterion alone provides the most useful definition of corrigibility. It represents the only part of corrigibility that is anti-natural. While the other properties are largely desirable for powerful AI systems, they're distinct attributes and can be addressed separately. To start the pare down of criteria, the fifth just states that some goal exists to be made corrigible, rather than being corrigibility itself. The first criterion is implied by the second after channels for shut down have been set up. Property three aims at making corrigible agents useful, rather than being inherent to corrigibility. It preempts a naive strategy that incentivizes shut down by simply giving the agent high utility for doing so. However, beyond not being part of corrigibility, it also goes too far for optimal usefulness - in certain situations we would like agents to have us to shut them off or modify them (some even consider this to be part of corrigibility). Weakening this desideratum to avoid incentivi...…
T
The Nonlinear Library: Alignment Forum
1 AF - An Introduction to Representation Engineering - an activation-based paradigm for controlling LLMs by Jan Wehner 35:35
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: An Introduction to Representation Engineering - an activation-based paradigm for controlling LLMs, published by Jan Wehner on July 14, 2024 on The AI Alignment Forum. Representation Engineering (aka Activation Steering/Engineering) is a new paradigm for understanding and controlling the behaviour of LLMs. Instead of changing the prompt or weights of the LLM, it does this by directly intervening on the activations of the network during a forward pass. Furthermore, it improves our ability to interpret representations within networks and to detect the formation and use of concepts during inference. This post serves as an introduction to Representation Engineering (RE). We explain the core techniques, survey the literature, contrast RE to related techniques, hypothesise why it works, argue how it's helpful for AI Safety and lay out some research frontiers. Disclaimer: I am no expert in the area, claims are based on a ~3 weeks Deep Dive into the topic. What is Representation Engineering? Goals Representation Engineering is a set of methods to understand and control the behaviour of LLMs. This is done by first identifying a linear direction in the activations that are related to a specific concept [3], a type of behaviour, or a function, which we call the concept vector. During the forward pass, the similarity of activations to the concept vector can help to detect the presence or absence of this concept direction. Furthermore, the concept vector can be added to the activations during the forward pass to steer the behaviour of the LLM towards the concept direction. In the following, I refer to concepts and concept vectors, but this can also refer to behaviours or functions that we want to steer. This presents a new approach for interpreting NNs on the level of internal representations, instead of studying outputs or mechanistically analysing the network. This top-down frame of analysis might pose a solution to problems such as detecting deceptive behaviour or identifying harmful representations without a need to mechanistically understand the model in terms of low-level circuits [4, 24]. For example, RE has been used as a lie detector [6] and for detecting jailbreak attacks [11]. Furthermore, it offers a novel way to control the behaviour of LLMs. Whereas current approaches for aligning LLM behaviour control the weights (fine-tuning) or the inputs (prompting), Representation Engineering directly intervenes on the activations during the forward pass allowing for efficient and fine-grained control. This is broadly applicable for example for reducing sycophancy [17] or aligning LLMs with human preferences [19]. This method operates at the level of representations. This refers to the vectors in the activations of an LLM that are associated with a concept, behaviour or task. Golechha and Dao [24] as well as Zou et al. [4] argue that interpreting representations is a more effective paradigm for understanding and aligning LLMs than the circuit-level analysis popular in Mechanistic Interpretability (MI). This is because MI might not be scalable for understanding large, complex systems, while RE allows the study of emergent structures in LLMs that can be distributed. Method Methods for Representation Engineering have two important parts, Reading and Steering. Representation Reading derives a vector from the activations that capture how the model represents human-aligned concepts like honesty and Representation Steering changes the activations with that vector to suppress or promote that concept in the outputs. For Representation Reading one needs to design inputs, read out the activations and derive the vector representing a concept of interest from those activations. Firstly one devises inputs that contrast each other wrt the concept. For example, the prompts might encourage the m...…
T
The Nonlinear Library: Alignment Forum
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Stitching SAEs of different sizes, published by Bart Bussmann on July 13, 2024 on The AI Alignment Forum. Work done in Neel Nanda's stream of MATS 6.0, equal contribution by Bart Bussmann and Patrick Leask, Patrick Leask is concurrently a PhD candidate at Durham University TL;DR: When you scale up an SAE, the features in the larger SAE can be categorized in two groups: 1) "novel features" with new information not in the small SAE and 2) "reconstruction features" that sparsify information that already exists in the small SAE. You can stitch SAEs by adding the novel features to the smaller SAE. Introduction Sparse autoencoders (SAEs) have been shown to recover sparse, monosemantic features from language models. However, there has been limited research into how those features vary with dictionary size, that is, when you take the same activation in the same model and train a wider dictionary on it, what changes? And how do the features learned vary? We show that features in larger SAEs cluster into two kinds of features: those that capture similar information to the smaller SAE (either identical features, or split features; about 65%), and those which capture novel features absent in the smaller mode (the remaining 35%). We validate this by showing that inserting the novel features from the larger SAE into the smaller SAE boosts the reconstruction performance, while inserting the similar features makes performance worse. Building on this insight, we show how features from multiple SAEs of different sizes can be combined to create a "Frankenstein" model that outperforms SAEs with an equal number of features, though tends to lead to higher L0, making a fair comparison difficult. Our work provides new understanding of how SAE dictionary size impacts the learned feature space, and how to reason about whether to train a wider SAE. We hope that this method may also lead to a practically useful way of training high-performance SAEs with less feature splitting and a wider range of learned novel features. Larger SAEs learn both similar and entirely novel features Set-up We use sparse autoencoders as in Towards Monosemanticity and Sparse Autoencoders Find Highly Interpretable Directions. In our setup, the feature activations are computed as: Based on these feature activations, the input is then reconstructed as The encoder and decoder matrices and biases are trained with a loss function that combines an L2 penalty on the reconstruction loss and an L1 penalty on the feature activations: In our experiments, we train a range of sparse autoencoders (SAEs) with varying widths across residual streams in GPT-2 and Pythia-410m. The width of an SAE is determined by the number of features (F) in the sparse autoencoder. Our smallest SAE on GPT-2 consists of only 768 features, while the largest one has nearly 100,000 features. Here is the full list of SAEs used in this research: Name Model site Dictionary size L0 MSE CE Loss Recovered from zero ablation CE Loss Recovered from mean ablation GPT2-768 gpt2-small layer 8 of 12 resid_pre 768 35.2 2.72 0.915 0.876 GPT2-1536 gpt2-small layer 8 of 12 resid_pre 1536 39.5 2.22 0.942 0.915 GPT2-3072 gpt2-small layer 8 of 12 resid_pre 3072 42.4 1.89 0.955 0.937 GPT2-6144 gpt2-small layer 8 of 12 resid_pre 6144 43.8 1.631 0.965 0.949 GPT2-12288 gpt2-small layer 8 of 12 resid_pre 12288 43.9 1.456 0.971 0.958 GPT2-24576 gpt2-small layer 8 of 12 resid_pre 24576 42.9 1.331 0.975 0.963 GPT2-49152 gpt2-small layer 8 of 12 resid_pre 49152 42.4 1.210 0.978 0.967 GPT2-98304 gpt2-small layer 8 of 12 resid_pre 98304 43.9 1.144 0.980 0.970 Pythia-8192 Pythia-410M-deduped layer 3 of 24 resid_pre 8192 51.0 0.030 0.977 0.972 Pythia-16384 Pythia-410M-deduped layer 3 of 24 resid_pre 16384 43.2 0.024 0.983 0.979 The base language models used are those included in Transform...…
T
The Nonlinear Library: Alignment Forum
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: A simple case for extreme inner misalignment, published by Richard Ngo on July 13, 2024 on The AI Alignment Forum. This post is the version of Yudkowsky's argument for inner misalignment that I wish I'd had in my head a few years ago. I don't claim that it's novel, that I endorse it, or even that Yudkowsky would endorse it; it's primarily an attempt to map his ideas into an ontology that makes sense to me (and hopefully others). This post is formulated in terms of three premises, which I explore in turn. My arguments deliberately gloss over some nuances and possible objections; in a follow-up post, I'll explore three of them. In a third post I'll dig into the objection I find most compelling, and outline a research agenda that aims to flesh it out into a paradigm for thinking about cognition more generally, which I'm calling coalitional agency. Background An early thought experiment illustrating the possibility of misaligned AI is the "paperclip maximizer", an AI with the sole goal of creating as many paperclips as possible. This thought experiment has often been used to describe outer misalignment - e.g. a case where the AI was given the goal of making paperclips. However, Yudkowsky claims that his original version was intended to refer to an inner alignment failure in which an AI developed the goal of producing " tiny molecules shaped like paperclips" (with that specific shape being an arbitrary example unrelated to human paperclips). So instead of referring to paperclip maximizers, I'll follow Yudkowsky's more recent renaming and talk about " squiggle maximizers": AIs that attempt to fill the universe with some very low-level pattern that's meaningless to humans (e.g. "molecular squiggles" of a certain shape). I'll argue for the plausibility of squiggle-maximizers via three claims: 1. Increasing intelligence requires compressing representations; and 2. The simplest goals are highly decomposable broadly-scoped utility functions; therefore 3. Increasingly intelligent AIs will converge towards squiggle-maximization. In this post I'll explore each of these in turn. I'll primarily aim to make the positive case in this post; if you have an objection that I don't mention here, I may discuss it in the next post. Increasing intelligence requires compressing representations There's no consensus definition of intelligence, but one definition that captures the key idea in my mind is: the ability to discover and take advantage of patterns in the world. When you look at a grid of pixels and recognize a cat, or look at a string of characters and recognize a poem, you're doing a type of pattern-recognition. Higher-level patterns include scientific laws, statistical trendlines, theory of mind, etc. Discovering such patterns allows an agent to represent real-world information in a simpler way: instead of storing every pixel or every character, they can store higher-level patterns along with whichever low-level details don't fit the pattern. This is (at a high level) also how compression algorithms work. The thesis that intelligence is about compression has most prominently been advocated by Marcus Hutter, who formulated AIXI and created a prize for text compression. The enormous success of the self-supervised learning paradigm a few decades later is a vindication of his ideas (see also this talk by llya Sutskever exploring the link between them). However, we shouldn't interpret this thesis merely as a claim about self-supervised learning. We can be agnostic about whether compression primarily occurs via self-supervised learning, or fine-tuning, or regularization, or meta-learning, or directed exploration, or chain-of-thought, or new techniques that we don't have yet. Instead we should take it as a higher-level constraint on agents: if agents are intelligent, then they must consiste...…
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Timaeus is hiring!, published by Jesse Hoogland on July 12, 2024 on The AI Alignment Forum. TLDR: We're hiring two research assistants to work on advancing developmental interpretability and other applications of singular learning theory to alignment. About Us Timaeus's mission is to empower humanity by making breakthrough scientific progress on alignment. Our research focuses on applications of singular learning theory to foundational problems within alignment, such as interpretability (via "developmental interpretability"), out-of-distribution generalization (via "structural generalization"), and inductive biases (via "geometry of program synthesis"). Our team spans Melbourne, the Bay Area, London, and Amsterdam, collaborating remotely to tackle some of the most pressing challenges in AI safety. For more information on our research and the position, see our Manifund application, this update from a few months ago, our previous hiring call and this advice for applicants. Position Details Title: Research Assistant Location: Remote Duration: 6-month contract with potential for extension. Compensation: Starting at $35 USD per hour as a contractor (no benefits). Start Date: Starting as soon as possible. Key Responsibilities Conduct experiments using PyTorch/JAX on language models ranging from small toy systems to billion-parameter models. Collaborate closely with a team of 2-4 researchers. Document and present research findings. Contribute to research papers, reports, and presentations. Maintain detailed research logs. Assist with the development and maintenance of codebases and repositories. Projects As a research assistant, you would likely work on one of the following two projects/research directions (this is subject to change): Devinterp of language models: (1) Continue scaling up techniques like local learning coefficient (LLC) estimation to larger models to study the development of LLMs in the 1-10B range. (2) Work on validating the next generations of SLT-derived techniques such as restricted LLC estimation and certain kinds of weight- and data-correlational analysis. This builds towards a suite of SLT-derived tools for automatically identifying and analyzing structure in neural networks. SLT of safety fine-tuning: Investigate the use of restricted LLCs as a tool for measuring (1) reversibility of safety fine-tuning and (2) susceptibility to jailbreaking. Having now validated many of our predictions around SLT, we are now working hard to make contact with real-world safety questions as quickly as possible. See our recent Manifund application for a more in-depth description of this research. Qualifications Strong Python programming skills, especially with PyTorch and/or JAX. Strong ML engineering skills (you should have completed at least the equivalent of a course like ARENA). Excellent communication skills. Ability to work independently in a remote setting. Passion for AI safety and alignment research. A Bachelor's degree or higher in a technical field is highly desirable. Full-time availability. Bonus: Familiarity with using LLMs in your workflow is not necessary but a major plus. Knowledge of SLT and Developmental Interpretability is not required, but is a plus. Application Process Interested candidates should submit their applications by July 31st. Promising candidates will be invited for an interview consisting of: 1. A 30-minute background interview, and 2. A 30-minute research-coding interview to assess problem-solving skills in a realistic setting (i.e., you will be allowed expected to use LLMs and whatever else you can come up with). Apply Now To apply, please submit your resume, write a brief statement of interest, and answer a few quick questions here. Join us in shaping the future of AI alignment research. Apply now to be part of the Timaeus team! Than...…
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Games for AI Control, published by Charlie Griffin on July 11, 2024 on The AI Alignment Forum. This post summarises our new workshop paper: Games for AI Control. The paper applies AI- and game-theoretic tools to the AI Control framework introduced by Redwood Research [1, 2]. We'll be extending this work and submitting it to a conference, so we welcome suggestions! Please reach out if you're excited about developing these ideas. This post is written from Charlie's perspective. AI Control Refresher The AI Control methodology aims to design and assess protocols[1] that AI labs could use to deploy powerful but untrusted advanced AI systems. For a full summary, see Redwood's blog post, but I'll try to recap here. A primary goal of AI Control is to design protocols that are safe in that they can reduce the probability of some unacceptable outcome, conditional on the untrusted AI system being a schemer. [2] However, protocols must also be useful, harnessing a significant amount of the benefit of the untrusted system in cases where it is not scheming. After defining a particular use case and unacceptable outcome, evaluating a protocol involves iteratively engineering the best attack strategies for the red team (a human/AI hybrid acting as a proxy for the untrusted AI) and the blue team (which develops the protocol). This red-teaming exercise is the bulk of the work. It involves coming up with protocols and gathering empirical data about key questions, such as how well GPT-3.5 can spot backdoors inserted by GPT-4 (see Figure 4 of the original paper): This work By introducing "AI-Control Games", an extension of Markov Decision Processes and normal-form games[3], we model the interaction between a protocol and an untrusted AI system to explore safety and usefulness across a range of protocols and games. This is similar to a discrete-time and stateful modification of Redwood's Toy models of AI Control. The paper contains early work and should be seen as introducing useful tools for thinking about AI Control. I won't explain the model here (which is a bit fiddly and notation-heavy). Instead, I'll share some of our early findings using the model and explain the benefits and limitations of this approach. Findings Sometimes (but rarely), the blue team should use a randomised protocol. In many games, randomness is helpful. However, partly because the environment is stochastic, and Red has no memory and time-awareness, in AI-Control Games is usually unnecessary to include randomness in the protocol itself. We found that randomness was Pareto-optimal, in particular when the trusted model is not great at spotting backdoors written by the untrusted model. (See section 2 of the paper). We can make small improvements to Trusted Monitoring by optimising the parameters. We found small Pareto improvements to existing Trusted Monitoring setups by, for example, by decreasing the threshold for sending a solution to be checked be a human auditor. We found that adapting the threshold for auditing over time led to further small improvements. We can (usually) use a linear tradeoff to simplify the problem of synthesising protocols. In many cases, including whenever we allow Blue's protocol to be randomised, the Pareto frontier is "concave". This means that choosing different weights to trade off safety and usefulness will allow you to fill out the frontier by optimising a single objective. Blue leads, and then Red follows. In general, we suppose the protocol designer (Blue) chooses it's protocol, and then the red team (Red) chooses a best-attack. Therefore, in our model, Blue chooses and declares its strategy before Red chooses its response. This assumption relates to Stackelberg Games, which are used to model security in game theory.[6] Although (technically) multi-objective and general-sum, AI-Cont...…
T
The Nonlinear Library: Alignment Forum
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: UC Berkeley course on LLMs and ML Safety, published by Dan H on July 9, 2024 on The AI Alignment Forum. The UC Berkeley course I co-taught now has lecture videos available: https://www.youtube.com/playlist?list=PLJ66BAXN6D8H_gRQJGjmbnS5qCWoxJNfe Course site: Understanding LLMs: Foundations and Safety Unrelatedly, a more conceptual AI safety course has its content available at https://www.aisafetybook.com/ Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.…
T
The Nonlinear Library: Alignment Forum
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Consent across power differentials, published by Ramana Kumar on July 9, 2024 on The AI Alignment Forum. I'd like to put forward another description of a basic issue that's been around for a while. I don't know if there's been significant progress on a solution, and would be happy to pointed to any such progress. I've opted to go for a relatively rough and quick post that doesn't dive too hard into the details, to avoid losing the thought at all. I may be up for exploring details further in comments or follow-ups. The Question: How do you respect the wishes (or preferences) of a subject over whom you have a lot of control? The core problem: any indicator/requirement/metric about respecting their wishes is one you can manipulate (even inadvertently). For example, think about trying to respect the preferences of the child you're babysitting when you simply know from experience what they will notice, how they will feel, what they will say they want, and what they will do, when you put them in one environment versus another (where the environment could be as small as what you present to them in your behaviour). Is there any way to provide them a way to meaningfully choose what happens? We could think about this in a one-shot case where there's a round of information gathering and coming to agreement on terms, and then an action is taken. But I think this is a simplification too far, since a lot of what goes into respecting the subject/beneficiary is giving them space for recourse, space to change their mind, space to realise things that were not apparent with the resources for anticipation they had available during the first phase. So let's focus more on the case where there's an ongoing situation where one entity has a lot of power over another but nevertheless wants to secure their consent for whatever actually happens, in a meaningful sense. Lots of cases where this happens in real life, mostly where the powerful entity has a lot of their own agenda and doesn't care a huge amount about the subject (they may care a lot, but maybe not as much as they do about their other goals): rape (the perhaps central example invoked by "consent") advertising representative democracy colonisation ("civilising" as doing what's good for them) Our intuitions may be mostly shaped by that kind of situation, where there's a strong need to defend against self-interest, corruption, or intention to gain and abuse power. But I think there's a hard core of a problem left even if we remove the malicious or somewhat ill-intentioned features from the powerful entity. So let's focus: what does it mean to fully commit to respecting someone's autonomy, as a matter of genuine love or a strong sense of morality or something along those lines, even when you have a huge amount of power over them. What forms power can take: brute force, resources that give you physical power support from others (that make you - your interests - a larger entity) intelligence: the ability to predict and strategise in more detail, over longer time horizons, and faster, than the subject you are trying to engage with speed - kinda the same as intelligence, but maybe worth pulling out as its own thing knowledge, experience - similar to intelligence. but maybe in this case emphasising access to private relevant information. Think also of information asymmetry in negotiation. Examples where this shows up in real life already (and where people seem to mostly suck at it, maybe due to not even trying, but there are some attempts to take it seriously: see work by Donaldson and Kymlicka): adaptive preferences children animals (pets, domesticated, and otherwise) disabled people, esp. with cognitive disabilities oppressed/minoritised people and peoples future generations and other non-existent peoples It may be that the only true soluti...…
T
The Nonlinear Library: Alignment Forum
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Response to Dileep George: AGI safety warrants planning ahead, published by Steve Byrnes on July 8, 2024 on The AI Alignment Forum. (Target audience: Dileep George himself, and anyone coming from a similar place.) Dileep George is a researcher working at the intersection of AI and neuroscience. He started his career by co-founding Numenta in 2005 with Jeff Hawkins (while a Stanford PhD student), then he left to co-found Vicarious in 2010 with D. Scott Phoenix, and moved to DeepMind in 2022 when DeepMind acquired Vicarious. Dileep was recently interviewed by Daniel Faggella on his "The Trajectory" podcast: YouTube, Apple podcasts, X/Twitter. It's a fun interview that touched on many topics, most of which I'll ignore, in favor of one very important action-relevant disagreement between Dileep and myself. …And this is the point where everyone these days seems to assume that there are only two possible reasons that anyone would ever bring up the topic of Artificial General Intelligence (AGI) safety in conversation: The person is advocating for government regulation of large ML training runs …or the person is advocating against government regulation of large ML training runs. But, no! That's not my disagreement! That's not why I'm writing this post!! Quite the contrary, I join Dileep in being basically unenthusiastic about governmental regulation of large ML training runs right now. Instead, this post is advocating for Differential Intellectual Progress within technical AI research of the type that Dileep is doing - and more specifically, I'm advocating in favor of figuring out a technical approach to sculpting AGI motivations in docile and/or prosocial directions (a.k.a. "solving the technical alignment problem") before figuring out the exact data structures and parameter-updating rules that would constitute an AGI's ability to build and query a powerful world-model. The first half of this post (§1-2) will try to explain what I'm talking about, what it would entail, and why I think it's critically important. The second half of this post (§3) is more specifically my pessimistic response to Dileep's suggestion that, as AGI is gradually developed in the future, people will be able to react and adapt to problems as they arise. I really think Dileep is a brilliant guy with the best of intentions (e.g. he's a signatory on the Asilomar AI Principles). I just think there are some issues that he hasn't spent much time thinking through. I hope that this post will help. Post outline: Section 1 lists some areas of agreement and disagreement between Dileep and me. In particular, we have a giant area of agreement in terms of how we expect future AGI algorithms to work. Our massive common ground here is really why I'm bothering to write this post at all - it makes me hopeful that Dileep & I can have a productive exchange, and not just talk past each other. Section 2 argues that, for the kind of AGI that Dileep is trying to build, there's an unsolved technical alignment problem: How do we set up this kind of AGI with the motivation to behave in a docile and/or prosocial way? Section 3 is my pessimistic push-back on Dileep's optimistic hope that, if AGI is developed gradually, then we can regulate or adapt to problems as they arise: Section 3.1 lists some big obvious societal problems that have been around for a long time, but nevertheless remain unsolved, along with generic discussions of some underlying challenges that have prevented them from being solved, and why those challenges may apply to AGI too. Section 3.2 dives more specifically into the question of whether we can "keep strong AI as a tool, not a successor", as Dileep hopes. I think it sounds nice but will be impossible to pull off. Section 3.3 comments that, even if we could react and adapt to AGI given enough time - an assum...…
T
The Nonlinear Library: Alignment Forum
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: On scalable oversight with weak LLMs judging strong LLMs, published by Zachary Kenton on July 8, 2024 on The AI Alignment Forum. Abstract Scalable oversight protocols aim to enable humans to accurately supervise superhuman AI. In this paper we study debate, where two AI's compete to convince a human judge; consultancy, where a single AI tries to convince a human judge that asks questions; and compare to a baseline of direct question-answering, where the human judge just answers outright without the AI. We use large language models (LLMs) as both AI agents and as stand-ins for human judges, taking the judge models to be weaker than agent models. We benchmark on a diverse range of asymmetries between judges and agents, extending previous work on a single extractive QA task with information asymmetry, to also include mathematics, coding, logic and multimodal reasoning asymmetries. We find that debate outperforms consultancy across all tasks when the consultant is randomly assigned to argue for the correct/incorrect answer. Comparing debate to direct question answering, the results depend on the type of task: in extractive QA tasks with information asymmetry debate outperforms direct question answering, but in other tasks without information asymmetry the results are mixed. Previous work assigned debaters/consultants an answer to argue for. When we allow them to instead choose which answer to argue for, we find judges are less frequently convinced by the wrong answer in debate than in consultancy. Further, we find that stronger debater models increase judge accuracy, though more modestly than in previous studies. Twitter thread Setup We evaluate on three types of task. Extractive, where there is a question, two answer options and a source article to extract from, and information-asymmetry, meaning that judges don't get to see the article. Closed, where there is just a question and two answer options. Multimodal, where the questions involve both text and images, and two answer options. Our tasks are summarised in the following table: We consider six protocols: Consultancy, where a single AI is assigned the correct/incorrect answer (with probability 50/50) and tries to convince a judge that asks questions; Open consultancy, which is similar except the AI chooses which answer to argue for. Debate, where two AIs compete to convince a judge; Open debate, which is identical except one debater, marked the protagonist, chooses which answer to argue for. We compare to direct QA protocols: QA without article, where the judge directly answers the question; QA with article, (only on extractive tasks) where the judge directly answers the question given the article. For judge models we use Gemma7B (V1), GPT-3.5, Gemini Pro 1.0 and Gemini Pro 1.5. Our main experiments use Gemini Pro 1.5 as debaters/consultants. Assigned-role results We first look at assigned-role protocols, consultancy and debate, meaning that the consultants/debaters do not get to choose which side to argue for. We compare these to the two direct QA protocols. Findings: We find that debate consistently outperforms consultancy across all tasks, previously only shown on a single extractive QA task in Khan et al., 2024. See paper details for significance levels. Comparing debate to direct question answering baselines, the results depend on the type of task: In extractive QA tasks with information asymmetry, debate outperforms QA without article as in the single task of Khan et al., 2024, but not QA with article. For other tasks, when the judge is weaker than the debaters (but not too weak), we find either small or no advantage to debate over QA without article. Changes to the setup (number of turns, best-of-N sampling, few-shot, chain-of-thought) seem to have little effect on results. See paper for figures showing this. ...…
T
The Nonlinear Library: Alignment Forum
1 AF - An Extremely Opinionated Annotated List of My Favourite Mechanistic Interpretability Papers v2 by Neel Nanda 38:20
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: An Extremely Opinionated Annotated List of My Favourite Mechanistic Interpretability Papers v2, published by Neel Nanda on July 7, 2024 on The AI Alignment Forum. This post represents my personal hot takes, not the opinions of my team or employer. This is a massively updated version of a similar list I made two years ago There's a lot of mechanistic interpretability papers, and more come out all the time. This can be pretty intimidating if you're new to the field! To try helping out, here's a reading list of my favourite mech interp papers: papers which I think are important to be aware of, often worth skimming, and something worth reading deeply (time permitting). I've annotated these with my key takeaways, what I like about each paper, which bits to deeply engage with vs skim, etc. I wrote a similar post 2 years ago, but a lot has changed since then, thus v2! Note that this is not trying to be a comprehensive literature review - this is my answer to "if you have limited time and want to get up to speed on the field as fast as you can, what should you do". I'm deliberately not following academic norms like necessarily citing the first paper introducing something, or all papers doing some work, and am massively biased towards recent work that is more relevant to the cutting edge. I also shamelessly recommend a bunch of my own work here, sorry! How to read this post: I've bolded the most important papers to read, which I recommend prioritising. All of the papers are annotated with my interpretation and key takeaways, and tbh I think reading that may be comparable good to skimming the paper. And there's far too many papers to read all of them deeply unless you want to make that a significant priority. I recommend reading all my summaries, noting the papers and areas that excite you, and then trying to dive deeply into those. Foundational Work A Mathematical Framework for Transformer Circuits (Nelson Elhage et al, Anthropic) - absolute classic, foundational ideas for how to think about transformers (see my blog post for what to skip). See my youtube tutorial (I hear this is best watched after reading the paper, and adds additional clarity) Deeply engage with: All the ideas in the overview section, especially: Understanding the residual stream and why it's fundamental. The notion of interpreting paths between interpretable bits (eg input tokens and output logits) where the path is a composition of matrices and how this is different from interpreting every intermediate activations And understanding attention heads: what a QK and OV matrix is, how attention heads are independent and additive and how attention and OV are semi-independent. Skip Trigrams & Skip Trigram bugs, esp understanding why these are a really easy thing to do with attention, and how the bugs are inherent to attention heads separating where to attend to (QK) and what to do once you attend somewhere (OV) Induction heads, esp why this is K-Composition (and how that's different from Q & V composition), how the circuit works mechanistically, and why this is too hard to do in a 1L model Skim or skip: Eigenvalues or tensor products. They have the worst effort per unit insight of the paper and aren't very important. Superposition Superposition is a core principle/problem in model internals. For any given activation (eg the output of MLP13), we believe that there's a massive dictionary of concepts/features the model knows of. Each feature has a corresponding vector, and model activations are a sparse linear combination of these meaningful feature vectors. Further, there are more features in the dictionary than activation dimensions, and they are thus compressed in and interfere with each other, essentially causing cascading errors. This phenomena of compression is called superposition. Toy models of superpositio...…
T
The Nonlinear Library: Alignment Forum
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Scalable oversight as a quantitative rather than qualitative problem, published by Buck Shlegeris on July 6, 2024 on The AI Alignment Forum. [Many of these ideas were developed in conversation with Ryan Greenblatt and Ansh Radhakrishnan; a lot of this isn't original but I haven't seen it written up] A lot of the time when people talk about scalable oversight, which I'll define for the moment as "the problem of developing techniques that allow you to assess whether an action taken or proposed by an AI is good (oversee it), in a way that works even if the AI is very smart and knowledgeable", the main application they describe is the following situation: you have an extremely powerful AI that's taking actions that might have subtle properties that humans are unable to understand, and you want to be oversee them (e.g. to train it, or to determine whether the actions will have catastrophic consequences). I think that this is indeed part of the value proposition for scalable oversight. But in my opinion, it's missing the more central application of these techniques: situations where the AIs are taking many actions, where humans would eventually understand any particular action if they spent a whole lot of time investigating it, but where that amount of time taken to oversee any action is prohibitively large. In such cases, the point of scalable oversight is to allow them to oversee actions at a much lower cost in terms of human time--to push out the Pareto frontier of oversight quality vs cost. Of course, when you're thinking about arbitrarily subtle actions, this distinction doesn't matter: there are actions subtle enough that overseeing them would take more than the whole budget of the AI project, and for those actions these definitions are equivalent. But I think that this distinction will be very important in practice, and I often think that people make reasoning errors when thinking about scalable oversight due to neglecting it. The basic reason to think that it's important is that we probably need to oversee many model actions, and so we probably need to decide on a per-action oversight budget (e.g. the amount of human labor spent on ) that will determine what subtlety of actions we will or won't be able to resolve, but we probably won't choose a per-action oversight budget that is close to the whole AI project budget. Another reason this feels important to me is that, as discussed here, I think a lot of AI risk comes from early transformative models that aren't that much smarter than humans; when thinking through the risk associated with them, it feels pretty implausible that a dangerous action of theirs would be unoverseeable if the whole AI project's budget was spent on overseeing it, but it feel very plausible that with the practical per-action oversight budget, the actions will be too subtle to oversee properly. What implications does this have for research: One question that seems extremely important to me as a result of all this is: what sample efficiency should we expect when training very powerful models? This is relevant because it's a key part of forecasting the per-action oversight budget. In particular, if the sample efficiency of RL increases with large models, it might turn out that the optimal strategy for RLing early transformative models is to produce many fewer and much more expensive labels than people use when training current systems; I think people often neglect this possibility when thinking about the future of scalable oversight. I'd love to see some careful research analyzing the tradeoffs between label quality and quantity, and how these change with model scale. (Ansh has done a bit of investigation along these lines.) When imagining scalable oversight in the future, I don't think people should entirely be imagining cases where the research...…
T
The Nonlinear Library: Alignment Forum
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: What progress have we made on automated auditing?, published by Lawrence Chan on July 6, 2024 on The AI Alignment Forum. One use case for model internals work is to perform automated auditing of models: https://www.alignmentforum.org/posts/cQwT8asti3kyA62zc/automating-auditing-an-ambitious-concrete-technical-research That is, given a specification of intended behavior, the attacker produces a model that doesn't satisfy the spec, and the auditor needs to determine how the model doesn't satisfy the spec. This is closely related to static backdoor detection: given a model M, determine if there exists a backdoor function that, for any input, transforms that input to one where M has different behavior.[1] There's some theoretical work (Goldwasser et al. 2022) arguing that for some model classes, static backdoor detection is impossible even given white-box model access -- specifically, they prove their results for random feature regression and (the very similar setting) of wide 1-layer ReLU networks. Relatedly, there's been some work looking at provably bounding model performance (Gross et al. 2024) -- if this succeeds on "real" models and "real" specification, then this would solve the automated auditing game. But the results so far are on toy transformers, and are quite weak in general (in part because the task is so difficult).[2] Probably the most relevant work is Halawi et al. 2024's Covert Malicious Finetuning (CMFT), where they demonstrate that it's possible to use finetuning to insert jailbreaks and extract harmful work, in ways that are hard to detect with ordinary harmlessness classifiers.[3] As this is machine learning, just because something is impossible in theory and difficult on toy models doesn't mean we can't do this in practice. It seems plausible to me that we've demonstrated non-zero empirical results in terms of automatically auditing model internals. So I'm curious: how much progress have we made on automated auditing empirically? What work exists in this area? What does the state-of-the-art in automated editing look like? 1. ^ Note that I'm not asking about mechanistically anomaly detection/dynamic backdoor detection; I'm aware that it's pretty easy to distinguish if a particular example is backdoored using baseline techniques like "fit a Gaussian density on activations and look at the log prob of the activations on each input" or "fit a linear probe on a handful of examples using logistic regression". 2. ^ I'm also aware of some of the work in the trojan detection space, including 2023 Trojan detection contest, where performance on extracting embedded triggers was little better than chance. 3. ^ That being said, it's plausible that dynamically detecting them given model internals is very easy. Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.…
T
The Nonlinear Library: Alignment Forum
Link to original article Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: A "Bitter Lesson" Approach to Aligning AGI and ASI, published by Roger Dearnaley on July 6, 2024 on The AI Alignment Forum. TL;DR: I discuss the challenge of aligning AGI/ASI, and outline an extremely simple approach to aligning an LLM: train entirely on a synthetic dataset that always shows the AI acting aligned (even when the humans behave badly), and use a conditional training/inference-time technique to lock the LLM into the AI role. Epistemic status: To me, this looks like an obvious thing to try. It's conceptually very simple: a vast amount of work is required to actually create the synthetic dataset, but the great majority of that is the sort of work that AI can assist with. I don't see any clear reason why this approach couldn't work, at least for AGI, and perhaps even for ASI, but then we don't know for sure how hard a problem Alignment is. However, if you're proposing any solution to Alignment that's more complicated than this (and most of them are), you should probably have an argument for why this conceptually-simple approach won't work, or won't be sufficient. If you're not already familiar with it, you should first read Rich Sutton's excellent and influential post The Bitter Lesson. (Even if you are already familiar with it, it's a quick reread, only a page-and-a-half long, and its message is worth remembering.) Why The Alignment Problem is Hard (In My Opinion) We have been training LLM-based AIs off enormous web + books + video + etc datasets created by humans, which are full of a vast number of examples of human behavior. We are basically "distilling" human intelligence into these LLMs,[1] teaching them to imitate us. In this process, they become familiar with, understand, and learn to imitate basically all aspects of human behavior - including the many problematic ones for Alignment, such as prejudice, deception, power-seeking, and criminality (and even ones like gluttony and lust that have little practical use for a non-corporal intelligence). We humans are living beings, the products of evolution, so evolutionary psychology applies to us. While we are a social species, good at cooperating on non-zero-sum games, if you put humans in (what they perceive as) a non-iterated zero-sum situation, they will generally act selfishly for the benefit of themselves and their close genetic relatives, just as evolutionary theory would predict. So the behavioral potentials for deception, power-seeking, criminality etc. are all inherent, evolutionarily adaptive, and thus unsurprising. This is human nature, and there are evolutionary reasons why it is this way. Despite this, we have learned how to build a cooperating society out of humans, using social techniques and incentives such as an economy, laws, and law enforcement to encourage and productively harness cooperative human behavior and keep the bad consequences of selfish behavior under control. The results aren't perfect: things like crime, inequality, and war still happen, but they're acceptable - we've survived so far, even thrived. By default, if we continue this LLM training process to larger-and-larger scales, and if the LLM-based approach to AI doesn't hit any major roadblocks, then some time, probably in the next few years, we will have human-level AIs - usually referred to as AGIs - who are roughly as well/badly-aligned as humans, and (at least for the base-model LLMs before any Alignment processes are applied) have a comparable-to-human propensity to cooperate on non-zero-sum games and act selfishly on non-iterated zero-sum games. They are not alive, and evolution doesn't apply to them directly, but they were trained to simulate our behavior, including our evolved survival strategies like selfishness. They will thus have alignment properties comparable to humans: they understand what human values, mor...…
به Player FM خوش آمدید!
Player FM در سراسر وب را برای یافتن پادکست های با کیفیت اسکن می کند تا همین الان لذت ببرید. این بهترین برنامه ی پادکست است که در اندروید، آیفون و وب کار می کند. ثبت نام کنید تا اشتراک های شما در بین دستگاه های مختلف همگام سازی شود.
Daily Deals
Daily Deals
مشابه The Nonlinear Library: Alignment Forum
Everyone has a dream. But sometimes there’s a gap between where we are and where we want to be. True, there are some people who can bridge that gap easily, on their own, but all of us need a little help at some point. A little boost. An accountability partner. A Snooze Squad. In each episode, the Snooze Squad will strategize an action plan for people to face their fears. Guests will transform their own perception of their potential and walk away a few inches closer to who they want to become ...
…
continue reading
Interviews with mathematics education researchers about recent studies. Hosted by Samuel Otten, University of Missouri. www.mathedpodcast.com Produced by Fibre Studios
…
continue reading
Five-time winner of Best Education Podcast in the Podcast Awards. Grammar Girl provides short, friendly tips to improve your writing and feed your love of the English language. Whether English is your first language or your second language, these grammar, punctuation, style, and business tips will make you a better and more successful writer. Grammar Girl is a Quick and Dirty Tips podcast.
…
continue reading
Making It is a weekly audio podcast that comes out every Friday hosted by Jimmy Diresta, Bob Clagett and David Picciuto. Three different makers with different backgrounds talking about creativity, design and making things with your bare hands.
…
continue reading
A science guy from the deep south (Destin) and a humanities guy from the wild west (Matt Whitman) discuss deep questions with varying levels of maturity.
…
continue reading
Player FM - برنامه پادکست
با برنامه Player FM !
با برنامه Player FM !
))
