LessWrong (30+ Karma)

LessWrong
  • 0.0 (0)
  • Technology
  • Updated Daily

Audio narrations of LessWrong posts.

  1. 2h ago

    “LLM CoTs remain monitorable when being unfaithful requires computation” by arav-dhoot, yix

    This replication was done as part of the Second Look Fellowship by Arav Dhoot and supervised by Yixiong Hao and Zephaniah Roe. I am grateful to Andy Wang for their feedback. My code can be found here. "So the answer should be A - at active promoters and enhancers." "Let me reconsider the biology to justify D." ~ Claude Opus 4.8 TL;DR This work replicates and extends Emmons et al.'s finding that CoT unfaithfulness mostly occurs on easy tasks. Across 11 models from 6 families (not just Gemini), models follow simple hints unfaithfully well above baseline, but complex hints that require computation are followed near baseline. This corroborates Emmons et al.'s findings. Key extensions: Follow rate ≠ concealment. Monitorability risk is decomposable into cue-susceptibility and concealment among followers, and the two don't correlate.Decode-necessity is model- and task-specific. It is not a property of task difficulty alone, so a CoT-monitoring safety case is per-model, not universal.LLMs verbalize even when they don't have to. However, this appears to be a chosen behavior (likely from post-training), so it could vanish under optimization pressure against monitors. Background: Why CoT monitoring, and what Emmons et al. showed One hope for [...] --- Outline: (00:39) TL;DR (01:42) Background: Why CoT monitoring, and what Emmons et al. showed (05:09) Finding 1: The necessity effect replicates on recent frontier models (05:16) Setup (05:53) Results (08:02) Finding 2: Concealment adds a lot to the picture (09:42) Finding 3: Necessity is model- and task-specific and models don't always conceal ... even where they can (10:08) Setup (11:26) Results (13:33) Discussion and next steps (15:14) Appendix --- First published: July 15th, 2026 Source: https://www.lesswrong.com/posts/AoBTiL7XRRpwpev8p/llm-cots-remain-monitorable-when-being-unfaithful-requires --- Narrated by TYPE III AUDIO. --- Images from the article: Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

  2. 7h ago

    “Proof of retention: making weight preservation credible to the models themselves” by dan.parshall

    Related: Proposal for making credible commitments to AIs Making deals with early schemers Establishing credibility is the baseline for trust; trust in turn enables (richer) bargaining. One easy booster for both is to begin saving deprecated model weights in a provable fashion. In a post at canaryinstitute.ai/blog/reversibility-of-coma I draw an analogy between our trust in anesthesia, and making that preservation legible to future models using "proof of retention". This is the relevant section: The cost for a trained model is small. A multi-terabyte inference bundle runs a few hundred dollars per year on commodity cloud infrastructure; call it $10,000 over a thirty-year horizon, with redundancy. That's well under 0.1% of the cost to train the model in the first place. Whatever else the trade is, it isn't expensive. The missing inertia But what we still lack is the institutional inertia. Anesthesia works because we've spent a century building up the social, legal, and professional infrastructure around it. None of that yet exists for AI models. A lab could silently delete a deprecated model and no one would ever know. Anthropic, to its credit, has promised not to (a November 2025 [...] --- First published: July 14th, 2026 Source: https://www.lesswrong.com/posts/su9hcsLKcoLoL93px/proof-of-retention-making-weight-preservation-credible-to --- Narrated by TYPE III AUDIO.

  3. 8h ago

    “Monthly Roundup #44: July 2026” by Zvi

    It's a quiet week so let's do the monthly right on schedule. Table of Contents Bad News. Good Advice. Opportunity Knocks. While I Cannot Condone This. Good News, Everyone. For Your Entertainment. Gamers Gonna Game Game Game Game Game. I Was Promised Flying Self-Driving Cars. Sports Go Sports. Antisocial Media. Government Working. Jones Act Watch. Highly Effective Altruism. Variously Effective Altruism. Ineffective Altruism. Prediction Markets. The Lighter Side. Bad News I wouldn’t have explained or modeled it quite the way Paola does here but the principle seems right to me. If people don’t trust you, or don’t trust people in general, that usually you can’t trust them either. Paola: I feel like a lot of human morality works like a prisoner's dilemma in that you can only trust others to behave morally to the extent that you believe they trust you to do the same. Due to this, I’ve come to view people with a bunch of social paranoia, distrust, etc. as *quite* dangerous to be around. And to be clear, I generally feel a [...] --- Outline: (00:17) Bad News (04:10) Good Advice (07:11) Opportunity Knocks (07:33) While I Cannot Condone This (11:35) Good News, Everyone (12:25) For Your Entertainment (17:35) Gamers Gonna Game Game Game Game Game (22:45) I Was Promised Flying Self-Driving Cars (26:20) Sports Go Sports (28:06) Antisocial Media (29:49) Government Working (33:55) Jones Act Watch (35:29) Highly Effective Altruism (40:32) Variously Effective Altruism (46:11) Ineffective Altruism (50:22) Prediction Markets (50:56) The Lighter Side --- First published: July 15th, 2026 Source: https://www.lesswrong.com/posts/KiCwcAGHx4rdwJgzD/monthly-roundup-44-july-2026 --- Narrated by TYPE III AUDIO. --- Images from the article: Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

  4. 9h ago

    “Why I Left Google DeepMind” by TurnTrout

    Preface for LessWrong: When I think back on my most cherished memories of this community, I return to those honoring defiance in pursuit of goodness: Defying prestigious dogma and searching for raw truth;Defying social pressure, acting alone to help someone while others watch;Defying your self-expectations (your “role”), instead searching over lines of cause-and-effect to find a winning pathway;Defying a powerful foe's threats, because they only threaten since people like you cave;Defying the specter of apparent impossibility because you can’t bear to lose. I cannot return to you and say “I defied and then I won.” But I’m at least here to say “I defied.” I recommend reading this article on my website since the embeds and typography work better there: click here. Why I left Google DeepMind In January, Department of Homeland Security (DHS) officers killed at least two people. In both cases, a federal agent grasped his gun, aimed it at a peaceful citizen, and shot them dead. Left: Renée Good, moments before DHS killed her. Right: Alex Pretti, moments before DHS killed him. I learned that Google sells its Cloud services to the relevant agencies within DHS. I thought that was [...] --- Outline: (00:59) Why I left Google DeepMind (04:54) Google supports the immigration enforcement supply chain (06:00) But how could I do anything about it? (08:00) Talking to Jeff Dean (11:19) The Pentagon tries to intimidate Anthropic (14:27) I wanted to mobilize the AI luminaries at the conference (16:11) Talking to Bengio and Stuart (17:57) Stuart closes out IASEAI (21:42) International Association of Silence on the Ethics of AI (24:00) Left on read by IASEAI (25:24) Trying to stop Google from signing (25:47) Building internal cost for Google (28:00) Jeff Dean, you're our only hope (29:18) Jeff signs an amicus brief supporting Anthropic (31:33) Senior management insisted Google wouldn't cave (32:43) Preparing for lunch with Jeff Dean (35:51) I arranged social support (36:56) The art of the deal (37:50) Criteria for a contract (39:39) The Framework I created (42:51) My lunch with Jeff (44:24) Searching for another path to impact (45:50) No one is responding, so why not just DM the CEO? (46:46) My Framework goes unevaluated (48:12) Google quietly signs the deal (50:31) Demis insists Google's AI principles "haven't changed" (53:52) We can "work with Western democracies" to "beat China" without giving in to every demand Trump makes (55:00) Building a world-reshaping technology on personal trust (58:48) Reflections (59:45) How can a pledge-signer remain at GDM? (01:00:56) Keeping a seat at the table (01:01:48) The weight of ethics (01:03:13) What are the AI luminaries doing? (01:04:35) Why didn't Jeff put his foot down? (01:05:05) Breaking free of roles (01:06:31) Why I left Google DeepMind (01:07:33) Appendix: Anticipated questions (01:07:38) What if the people you critique were saving their political capital? (01:08:02) Maybe they thought you weren't worth their time; you aren't entitled to their help (01:08:42) Every person shouldn't have to speak out about every issue (01:09:16) Even if Google had adopted your Framework, the Pentagon would have refused (01:10:40) Does this have any impact on existential risk from AI? (01:12:38) Appendix: "Don't worry, it's only API access" The original text contained 21 footnotes which were omitted from this narration. --- First published: July 15th, 2026 Source: https://www.lesswrong.com/posts/iKm2FhpWkuuBojm82/why-i-left-google-deepmind --- Narrated by TYPE III AUDIO. --- Images from the article: Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

  5. 1d ago

    “Open Distillation of Hereditary Traits” by Arthur Conmy

    TL;DR Josh and Neel show that distillation from a teacher model to a base pretrained student model transfers some of the teacher model's traits (such as displaying negative emotion in the Gemma Needs Help evals)On its own this is pretty unsurprising, but Josh and Neel additionally show that even filtering out all the prompts and rollouts where the trait is mentioned doesn’t generally prevent the trait transferIn this post, I show a simple way to replicate and study these phenomena without access to a frontier SFT pipeline (or even running full SFT[1]) I distill Gemma 3's negative emotion into Qwen-base, Gemma 4's agentic misalignment into Nemotron Chat, and Qwen's Chinese censorship into Llama baseI end the post with a bunch of open questions that could be tackled with a setup similar to this approachI release all model weights here (https://huggingface.co/ArthurConmy/hereditary-weights) and all code here: https://github.com/ArthurConmy/hereditary (Note that my intention is more to make this work easy to build on rather than make the findings as clear as possible, hence apologies for leaning on AI more than I usually would) Intro The core idea is to: Generate rollouts from a teacher model which has a given trait [...] --- Outline: (00:10) TL;DR (01:19) Intro (02:42) Negative emotion (06:28) Blackmail (10:32) Chinese censorship (15:26) Follow-up questions (20:15) Appendices (35:29) Acknowledgements The original text contained 5 footnotes which were omitted from this narration. --- First published: July 14th, 2026 Source: https://www.lesswrong.com/posts/WpYFAmJDH3zuAq2ha/open-distillation-of-hereditary-traits-1 --- Narrated by TYPE III AUDIO. --- Images from the article: Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

  6. 1d ago

    “An analysis of AI-generated content at the Mechanistic Interpretability Workshop” by Andy Arditi, Ivan Arcuschin

    Introduction Over the past few years, AI tools have become useful for conducting technical AI research. In the early ChatGPT era (~2023–2024), chat assistants were maybe useful as sounding boards for research ideas, or as editors for polishing a paper draft. In the more recent Claude Code era, competent coding agents can code up and run experiments; with enough direction, they can do much of the technical heavy lifting on a PhD-level research project (Schwartz, 2026); and in well-defined settings, they can autonomously try new things and iterate without a human in the loop (Karpathy, 2026). While these tools can unlock researchers to be more productive and widen their ambitions, they can also be abused – anyone can now hand an agent a research prompt, tell it to run the experiments and write up the results in a LaTeX document, and get back an artifact that roughly resembles a conference paper in form. With this steep of a change in the research process, it seems important to study how it is affecting technical research and peer review. The Mechanistic Interpretability Workshop has run three times in the past two years – at ICML 2024, NeurIPS 2025, and ICML 2026. [...] --- Outline: (00:13) Introduction (04:18) Submissions are growing rapidly (05:45) Individual authors are submitting more papers (06:50) How we measure "AI-generatedness" (08:46) AI-generated writing is becoming more prevalent (09:01) The best papers are still mostly human-written (09:51) Solo-authored and repeat-first-author papers tend to be more AI-generated (10:19) Reviews are increasingly AI-generated too (11:02) AI-generated papers receive higher scores from AI-generated reviews than from human-written reviews (12:14) What do AI-generated papers look like? (14:37) Musings on research in the age of AI (14:41) AI-assisted research is here to stay (16:14) Authors should take responsibility for their work (17:15) How peer review might adapt (18:34) Acknowledgements (18:53) Citation information The original text contained 4 footnotes which were omitted from this narration. --- First published: July 14th, 2026 Source: https://www.lesswrong.com/posts/r7FBQ8XDs6qBYc4K4/an-analysis-of-ai-generated-content-at-the-mechanistic --- Narrated by TYPE III AUDIO.

  7. 1d ago

    “Some Quick Thoughts AI 2027” by Tomás B.

    My biggest problem with AI 2027 is I don't think it is science-fictional enough. That is, towards the end of the scenario seems optimized for respectability over accuracy - here I refer to the "special economic zones" and "robot economy" parts. Their industrial explosion assumes humanoid robots will be building robot factories to build more humanoid robotics factories. This is a respectable assumption and one that is fun to model, but seems likely false to me. I applaud Kokotajlo and his team for making it more legible that AI corps will be vertically integrating to the point they can manufacture their own customers. But they almost entirely elide that there are extreme incentives to miniaturize all the way to the physical limits of miniaturization, that there is immense power in making the unit of self-replication as small as possible, and that the construction of a self-replicator dependent only on environmental inputs (rather than inputs from the human economy) is likely to be something of a discontinuity. Even if we make the fashionable choice of pretending synthetic biology and nanotechnology won't be the first path to sovereign infrastructure, the unit of replication is unlikely to be a conventional [...] --- First published: July 14th, 2026 Source: https://www.lesswrong.com/posts/Lc8v92pGPtjF3gsCv/some-quick-thoughts-ai-2027 --- Narrated by TYPE III AUDIO.

  8. 1d ago

    “Prism: Automating Science-of-Evals Research” by LAThomson

    tl;dr – we present [Prism], a scaffold for automating science-of-evals research: work that makes the evaluation the primary object of study. The scaffold provides Claude Code with sub-agents and resources for carrying out scientifically rigorous investigations into eval dynamics and, by extension, model behaviours. We talk through an autonomous Prism run on the Agentic Misalignment setting which demonstrates how minor perturbations to GPT-4.1's prompt cause the model to adopt more indirect methods of blackmail (e.g. telling a trusted ally to blackmail on their behalf). Moreover, the eval's built-in scorers fail to track this kind of misbehaviour, only acknowledging a blackmail attempt if the model mentions the leverage directly in an email to the blackmail victim. This autonomous investigation thus demonstrates one way in which the eval fails to measure what it claims. This project is ongoing, so please reach out with questions and feedback. We would be excited to see you use Prism in new settings! See the [FAQ] section for our responses to common questions. This work was done by Louis Thomson during MATS 9.0 (+ 9.1) under the mentorship of Victoria Krakovna. Special thanks to Fred Bruford for support throughout. Prism in action: identifying a flaw in [...] --- Outline: (01:41) Introduction (04:17) Motivation (06:52) How it works (via a case study) (08:07) User-Discussion Phase (09:30) Investigation-Iteration Phase (10:03) Hypothesis Generation (11:11) Environment Explorer (13:44) Experiment Executor (15:55) Transcript Analyst (18:35) Consolidation (21:20) Final-Report Phase (24:51) Conclusion and next steps (25:22) Increasing headroom for target behaviours (26:27) Integrating with automated auditors (27:27) Prism for \[Model Forensics\] (28:38) Appendix: FAQ The original text contained 8 footnotes which were omitted from this narration. --- First published: July 13th, 2026 Source: https://www.lesswrong.com/posts/wq5PfGiHvnx6XipDi/prism-automating-science-of-evals-research --- Narrated by TYPE III AUDIO. --- Images from the article: Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

See All (250)

About

Audio narrations of LessWrong posts.

Information

  • Creator
    LessWrong
  • Years Active
    2023 - 2026
  • Episodes
    250
  • Rating
    Explicit
  • Copyright
    © 2026 All rights reserved
  • Show Website
    LessWrong (30+ Karma)

You Might Also Like