AXRP - the AI X-risk Research Podcast

Very brief survey: bit.ly/axrpsurvey2023 Store is closing in a week! Link: store.axrp.net/ Patreon: patreon.com/axrpodcast Ko-fi: ko-fi.com/axrpodcast

What can we learn about advanced deep learning systems by understanding how humans learn and form values over their lifetimes? Will superhuman AI look like ruthless coherent utility optimization, or more like a mishmash of contextually activated desires? This episode's guest, Quintin Pope, has been thinking about these questions as a leading researcher in the shard theory community. We talk about what shard theory is, what it says about humans and neural networks, and what the implications are for making AI safe. Patreon: patreon.com/axrpodcast Ko-fi: ko-fi.com/axrpodcast Episode art by Hamish Doodles: hamishdoodles.com   Topics we discuss, and timestamps:  - 0:00:42 - Why understand human value formation?    - 0:19:59 - Why not design methods to align to arbitrary values?  - 0:27:22 - Postulates about human brains    - 0:36:20 - Sufficiency of the postulates    - 0:44:55 - Reinforcement learning as conditional sampling    - 0:48:05 - Compatibility with genetically-influenced behaviour    - 1:03:06 - Why deep learning is basically what the brain does  - 1:25:17 - Shard theory    - 1:38:49 - Shard theory vs expected utility optimizers    - 1:54:45 - What shard theory says about human values  - 2:05:47 - Does shard theory mean we're doomed?    - 2:18:54 - Will nice behaviour generalize?    - 2:33:48 - Does alignment generalize farther than capabilities?  - 2:42:03 - Are we at the end of machine learning history?  - 2:53:09 - Shard theory predictions  - 2:59:47 - The shard theory research community    - 3:13:45 - Why do shard theorists not work on replicating human childhoods?  - 3:25:53 - Following shardy research   The transcript: axrp.net/episode/2023/06/15/episode-22-shard-theory-quintin-pope.html   Shard theorist links:  - Quintin's LessWrong profile: lesswrong.com/users/quintin-pope  - Alex Turner's LessWrong profile: lesswrong.com/users/turntrout  - Shard theory Discord: discord.gg/AqYkK7wqAG  - EleutherAI Discord: discord.gg/eleutherai   Research we discuss:  - The Shard Theory Sequence: lesswrong.com/s/nyEFg3AuJpdAozmoX  - Pretraining Language Models with Human Preferences: arxiv.org/abs/2302.08582  - Inner alignment in salt-starved rats: lesswrong.com/posts/wcNEXDHowiWkRxDNv/inner-alignment-in-salt-starved-rats  - Intro to Brain-like AGI Safety Sequence: lesswrong.com/s/HzcM2dkCq7fwXBej8  - Brains and transformers:    - The neural architecture of language: Integrative modeling converges on predictive processing: pnas.org/doi/10.1073/pnas.2105646118    - Brains and algorithms partially converge in natural language processing: nature.com/articles/s42003-022-03036-1    - Evidence of a predictive coding hierarchy in the human brain listening to speech: nature.com/articles/s41562-022-01516-2  - Singular learning theory explainer: Neural networks generalize because of this one weird trick: lesswrong.com/posts/fovfuFdpuEwQzJu2w/neural-networks-generalize-because-of-this-one-weird-trick  - Singular learning theory links: metauni.org/slt/  - Implicit Regularization via Neural Feature Alignment, aka circles in the parameter-function map: arxiv.org/abs/2008.00938  - The shard theory of human values: lesswrong.com/s/nyEFg3AuJpdAozmoX/p/iCfdcxiyr2Kj8m8mT  - Predicting inductive biases of pre-trained networks: openreview.net/forum?id=mNtmhaDkAr  - Understanding and controlling a maze-solving policy network, aka the cheese vector: lesswrong.com/posts/cAC4AXiNC5ig6jQnc/understanding-and-controlling-a-maze-solving-policy-network  - Quintin's Research agenda: Supervising AIs improving AIs: lesswrong.com/posts/7e5tyFnpzGCdfT4mR/research-agenda-supervising-ais-improving-ais  - Steering GPT-2-XL by adding an activation vector: lesswrong.com/posts/5spBue2z2tw4JuDCx/steering-gpt-2-xl-by-adding-an-activation-vector   Links for the addendum on mesa-optimization skepticism:  - Quintin's response to Yudkowsky arguing against AIs being steerable by gradient descent: lesswrong.com/posts/wAczufCpMdaamF9fy/my-objections-to-we-re-all-gonna-die-with-eliezer-yudkowsky#Yudkowsky_argues_against_AIs_being_steerable_by_gradient_descent_  - Quintin on why evolution is not like AI training: lesswrong.com/posts/wAczufCpMdaamF9fy/my-objections-to-we-re-all-gonna-die-with-eliezer-yudkowsky#Edit__Why_evolution_is_not_like_AI_training  - Evolution provides no evidence for the sharp left turn: lesswrong.com/posts/hvz9qjWyv8cLX9JJR/evolution-provides-no-evidence-for-the-sharp-left-turn  - Let's Agree to Agree: Neural Networks Share Classification Order on Real Datasets: arxiv.org/abs/1905.10854

Lots of people in the field of machine learning study 'interpretability', developing tools that they say give us useful information about neural networks. But how do we know if meaningful progress is actually being made? What should we want out of these tools? In this episode, I speak to Stephen Casper about these questions, as well as about a benchmark he's co-developed to evaluate whether interpretability tools can find 'Trojan horses' hidden inside neural nets. Patreon: patreon.com/axrpodcast Ko-fi: ko-fi.com/axrpodcast   Topics we discuss, and timestamps:  - 00:00:42 - Interpretability for engineers    - 00:00:42 - Why interpretability?    - 00:12:55 - Adversaries and interpretability    - 00:24:30 - Scaling interpretability    - 00:42:29 - Critiques of the AI safety interpretability community    - 00:56:10 - Deceptive alignment and interpretability  - 01:09:48 - Benchmarking Interpretability Tools (for Deep Neural Networks) (Using Trojan Discovery)    - 01:10:40 - Why Trojans?    - 01:14:53 - Which interpretability tools?    - 01:28:40 - Trojan generation    - 01:38:13 - Evaluation  - 01:46:07 - Interpretability for shaping policy  - 01:53:55 - Following Casper's work   The transcript: axrp.net/episode/2023/05/02/episode-21-interpretability-for-engineers-stephen-casper.html   Links for Casper:  - Personal website: stephencasper.com/  - Twitter: twitter.com/StephenLCasper  - Electronic mail: scasper [at] mit [dot] edu   Research we discuss:  - The Engineer's Interpretability Sequence: alignmentforum.org/s/a6ne2ve5uturEEQK7  - Benchmarking Interpretability Tools for Deep Neural Networks: arxiv.org/abs/2302.10894  - Adversarial Policies beat Superhuman Go AIs: goattack.far.ai/  - Adversarial Examples Are Not Bugs, They Are Features: arxiv.org/abs/1905.02175  - Planting Undetectable Backdoors in Machine Learning Models: arxiv.org/abs/2204.06974  - Softmax Linear Units: transformer-circuits.pub/2022/solu/index.html  - Red-Teaming the Stable Diffusion Safety Filter: arxiv.org/abs/2210.04610   Episode art by Hamish Doodles: hamishdoodles.com

How should we scientifically think about the impact of AI on human civilization, and whether or not it will doom us all? In this episode, I speak with Scott Aaronson about his views on how to make progress in AI alignment, as well as his work on watermarking the output of language models, and how he moved from a background in quantum complexity theory to working on AI.   Note: this episode was recorded before this story (vice.com/en/article/pkadgm/man-dies-by-suicide-after-talking-with-ai-chatbot-widow-says) emerged of a man committing suicide after discussions with a language-model-based chatbot, that included discussion of the possibility of him killing himself. Patreon: https://www.patreon.com/axrpodcast Ko-fi: https://ko-fi.com/axrpodcast   Topics we discuss, and timestamps:  - 0:00:36 - 'Reform' AI alignment    - 0:01:52 - Epistemology of AI risk    - 0:20:08 - Immediate problems and existential risk    - 0:24:35 - Aligning deceitful AI    - 0:30:59 - Stories of AI doom    - 0:34:27 - Language models    - 0:43:08 - Democratic governance of AI    - 0:59:35 - What would change Scott's mind  - 1:14:45 - Watermarking language model outputs    - 1:41:41 - Watermark key secrecy and backdoor insertion  - 1:58:05 - Scott's transition to AI research    - 2:03:48 - Theoretical computer science and AI alignment    - 2:14:03 - AI alignment and formalizing philosophy    - 2:22:04 - How Scott finds AI research  - 2:24:53 - Following Scott's research   The transcript: axrp.net/episode/2023/04/11/episode-20-reform-ai-alignment-scott-aaronson.html   Links to Scott's things:  - Personal website: scottaaronson.com  - Book, Quantum Computing Since Democritus: amazon.com/Quantum-Computing-since-Democritus-Aaronson/dp/0521199565/  - Blog, Shtetl-Optimized: scottaaronson.blog   Writings we discuss:  - Reform AI Alignment: scottaaronson.blog/?p=6821  - Planting Undetectable Backdoors in Machine Learning Models: arxiv.org/abs/2204.06974

Store: https://store.axrp.net/ Patreon: https://www.patreon.com/axrpodcast Ko-fi: https://ko-fi.com/axrpodcast Video: https://www.youtube.com/watch?v=kmPFjpEibu0

How good are we at understanding the internal computation of advanced machine learning models, and do we have a hope at getting better? In this episode, Neel Nanda talks about the sub-field of mechanistic interpretability research, as well as papers he's contributed to that explore the basics of transformer circuits, induction heads, and grokking.   Topics we discuss, and timestamps:  - 00:01:05 - What is mechanistic interpretability?  - 00:24:16 - Types of AI cognition  - 00:54:27 - Automating mechanistic interpretability  - 01:11:57 - Summarizing the papers  - 01:24:43 - 'A Mathematical Framework for Transformer Circuits'    - 01:39:31 - How attention works    - 01:49:26 - Composing attention heads    - 01:59:42 - Induction heads  - 02:11:05 - 'In-context Learning and Induction Heads'    - 02:12:55 - The multiplicity of induction heads    - 02:30:10 - Lines of evidence    - 02:38:47 - Evolution in loss-space    - 02:46:19 - Mysteries of in-context learning  - 02:50:57 - 'Progress measures for grokking via mechanistic interpretability'    - 02:50:57 - How neural nets learn modular addition    - 03:11:37 - The suddenness of grokking  - 03:34:16 - Relation to other research  - 03:43:57 - Could mechanistic interpretability possibly work?  - 03:49:28 - Following Neel's research   The transcript: axrp.net/episode/2023/02/04/episode-19-mechanistic-interpretability-neel-nanda.html   Links to Neel's things:  - Neel on Twitter: twitter.com/NeelNanda5  - Neel on the Alignment Forum: alignmentforum.org/users/neel-nanda-1  - Neel's mechanistic interpretability blog: neelnanda.io/mechanistic-interpretability  - TransformerLens: github.com/neelnanda-io/TransformerLens  - Concrete Steps to Get Started in Transformer Mechanistic Interpretability: alignmentforum.org/posts/9ezkEb9oGvEi6WoB3/concrete-steps-to-get-started-in-transformer-mechanistic  - Neel on YouTube: youtube.com/@neelnanda2469  - 200 Concrete Open Problems in Mechanistic Interpretability: alignmentforum.org/s/yivyHaCAmMJ3CqSyj  - Comprehesive mechanistic interpretability explainer: dynalist.io/d/n2ZWtnoYHrU1s4vnFSAQ519J   Writings we discuss:  - A Mathematical Framework for Transformer Circuits: transformer-circuits.pub/2021/framework/index.html  - In-context Learning and Induction Heads: transformer-circuits.pub/2022/in-context-learning-and-induction-heads/index.html  - Progress measures for grokking via mechanistic interpretability: arxiv.org/abs/2301.05217  - Hungry Hungry Hippos: Towards Language Modeling with State Space Models (referred to in this episode as the "S4 paper"): arxiv.org/abs/2212.14052  - interpreting GPT: the logit lens: lesswrong.com/posts/AcKRB8wDpdaN6v6ru/interpreting-gpt-the-logit-lens  - Locating and Editing Factual Associations in GPT (aka the ROME paper): arxiv.org/abs/2202.05262  - Human-level play in the game of Diplomacy by combining language models with strategic reasoning: science.org/doi/10.1126/science.ade9097  - Causal Scrubbing: alignmentforum.org/s/h95ayYYwMebGEYN5y/p/JvZhhzycHu2Yd57RN  - An Interpretability Illusion for BERT: arxiv.org/abs/2104.07143  - Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small: arxiv.org/abs/2211.00593  - Grokking: Generalization Beyond Overfitting on Small Algorithmic Datasets: arxiv.org/abs/2201.02177  - The Effects of Reward Misspecification: Mapping and Mitigating Misaligned Models: arxiv.org/abs/2201.03544  - Collaboration & Credit Principles: colah.github.io/posts/2019-05-Collaboration  - Transformer Feed-Forward Layers Are Key-Value Memories: arxiv.org/abs/2012.14913   - Multi-Component Learning and S-Curves: alignmentforum.org/posts/RKDQCB6smLWgs2Mhr/multi-component-learning-and-s-curves  - The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks: arxiv.org/abs/1803.03635  - Linear Mode Connectivity and the Lottery Ticket Hypothesis: proceedings.mlr.press/v119/frankle20a    

I have a new podcast, where I interview whoever I want about whatever I want. It's called "The Filan Cabinet", and you can find it wherever you listen to podcasts. The first three episodes are about pandemic preparedness, God, and cryptocurrency. For more details, check out the podcast website (thefilancabinet.com), or search "The Filan Cabinet" in your podcast app.

Concept extrapolation is the idea of taking concepts an AI has about the world - say, "mass" or "does this picture contain a hot dog" - and extending them sensibly to situations where things are different - like learning that the world works via special relativity, or seeing a picture of a novel sausage-bread combination. For a while, Stuart Armstrong has been thinking about concept extrapolation and how it relates to AI alignment. In this episode, we discuss where his thoughts are at on this topic, what the relationship to AI alignment is, and what the open questions are.   Topics we discuss, and timestamps:  - 00:00:44 - What is concept extrapolation  - 00:15:25 - When is concept extrapolation possible  - 00:30:44 - A toy formalism  - 00:37:25 - Uniqueness of extrapolations  - 00:48:34 - Unity of concept extrapolation methods  - 00:53:25 - Concept extrapolation and corrigibility  - 00:59:51 - Is concept extrapolation possible?  - 01:37:05 - Misunderstandings of Stuart's approach  - 01:44:13 - Following Stuart's work   The transcript: axrp.net/episode/2022/09/03/episode-18-concept-extrapolation-stuart-armstrong.html   Stuart's startup, Aligned AI: aligned-ai.com   Research we discuss:  - The Concept Extrapolation sequence: alignmentforum.org/s/u9uawicHx7Ng7vwxA  - The HappyFaces benchmark: github.com/alignedai/HappyFaces  - Goal Misgeneralization in Deep Reinforcement Learning: arxiv.org/abs/2105.14111

Sometimes, people talk about making AI systems safe by taking examples where they fail and training them to do well on those. But how can we actually do this well, especially when we can't use a computer program to say what a 'failure' is? In this episode, I speak with Daniel Ziegler about his research group's efforts to try doing this with present-day language models, and what they learned. Listeners beware: this episode contains a spoiler for the Animorphs franchise around minute 41 (in the 'Fanfiction' section of the transcript).   Topics we discuss, and timestamps:  - 00:00:40 - Summary of the paper  - 00:02:23 - Alignment as scalable oversight and catastrophe minimization  - 00:08:06 - Novel contribtions  - 00:14:20 - Evaluating adversarial robustness  - 00:20:26 - Adversary construction  - 00:35:14 - The task  - 00:38:23 - Fanfiction  - 00:42:15 - Estimators to reduce labelling burden  - 00:45:39 - Future work  - 00:50:12 - About Redwood Research   The transcript: axrp.net/episode/2022/08/21/episode-17-training-for-very-high-reliability-daniel-ziegler.html   Daniel Ziegler on Google Scholar: scholar.google.com/citations?user=YzfbfDgAAAAJ   Research we discuss:  - Daniel's paper, Adversarial Training for High-Stakes Reliability: arxiv.org/abs/2205.01663  - Low-stakes alignment: alignmentforum.org/posts/TPan9sQFuPP6jgEJo/low-stakes-alignment  - Red Teaming Language Models with Language Models: arxiv.org/abs/2202.03286  - Uncertainty Estimation for Language Reward Models: arxiv.org/abs/2203.07472  - Eliciting Latent Knowledge: docs.google.com/document/d/1WwsnJQstPq91_Yh-Ch2XRL8H_EpsnjrC1dwZXR37PC8/edit

Many people in the AI alignment space have heard of AI safety via debate - check out AXRP episode 6 (axrp.net/episode/2021/04/08/episode-6-debate-beth-barnes.html) if you need a primer. But how do we get language models to the stage where they can usefully implement debate? In this episode, I talk to Geoffrey Irving about the role of language models in AI safety, as well as three projects he's done that get us closer to making debate happen: using language models to find flaws in themselves, getting language models to back up claims they make with citations, and figuring out how uncertain language models should be about the quality of various answers.   Topics we discuss, and timestamps:  - 00:00:48 - Status update on AI safety via debate  - 00:10:24 - Language models and AI safety  - 00:19:34 - Red teaming language models with language models  - 00:35:31 - GopherCite  - 00:49:10 - Uncertainty Estimation for Language Reward Models  - 01:00:26 - Following Geoffrey's work, and working with him   The transcript: axrp.net/episode/2022/07/01/episode-16-preparing-for-debate-ai-geoffrey-irving.html   Geoffrey's twitter: twitter.com/geoffreyirving   Research we discuss:  - Red Teaming Language Models With Language Models: arxiv.org/abs/2202.03286  - Teaching Language Models to Support Answers with Verified Quotes, aka GopherCite: arxiv.org/abs/2203.11147  - Uncertainty Estimation for Language Reward Models: arxiv.org/abs/2203.07472  - AI Safety via Debate: arxiv.org/abs/1805.00899  - Writeup: progress on AI safety via debate: lesswrong.com/posts/Br4xDbYu4Frwrb64a/writeup-progress-on-ai-safety-via-debate-1  - Eliciting Latent Knowledge: ai-alignment.com/eliciting-latent-knowledge-f977478608fc  - Training Compute-Optimal Large Language Models, aka Chinchilla: arxiv.org/abs/2203.15556