LessWrong (30+ Karma)

LessWrong

Audio narrations of LessWrong posts.

  1. 1 hr ago

    “Pause, at least after unipolarity” by David Matolcsi

    I think it's likely that the world will enter military unipolarity within our lifetime. I think the creation of such unipolarity is an alarming prospect, but at least once it happens, there will be less of an excuse to continue the race towards superintelligence. I think it's important that we shape our actions and advocacy in such a way that at the very latest when such unipolarity comes to exist, we stop AI development for a long time. The arrival of unipolarity Why do I believe that it's likely we will get to unipolarity? Well, what is the alternative? If unipolarity is never achieved, that means we forever have hostile great powers, maintaining large armies, pointing missiles at each other, and developing better and better AIs. There is never a big enough gap in AI capabilities for any side to develop a decisive strategic advantage, or everyone decides again and again not to use their advantage to disarm the opponent. Multiple groups go to space, still pointing weapons at each other; they develop literally Jupiter-brained superintelligences, but the balance of power never breaks and there always remain competitive, hostile militaries. I’m not saying this never-ending cold war is [...] --- Outline: (00:35) The arrival of unipolarity (03:07) Plan S (04:00) Plan A (08:28) Plan B and C (10:27) Pausing after unipolarity (14:56) Conclusion The original text contained 13 footnotes which were omitted from this narration. --- First published: August 2nd, 2026 Source: https://www.lesswrong.com/posts/QCFKzFbs2KjC3A76m/pause-at-least-after-unipolarity --- Narrated by TYPE III AUDIO.

  2. 1 hr ago

    “Single Forward Pass Evals on Fable, Opus 5, and GPT-5.6-Sol” by Christine Corry

    This is a research update for an on-going replication of single-forward-pass evals done as part of the Second Look Fellowship. In following posts, we will run more comprehensive replications of previous work and release open source tooling for single forward pass eval elicitation. Code can be found here. tl;dr We replicate experiments from Greenblatt 2025 and Greenblatt 2026 on one baseline model from the original post, Opus 4.5. Our evaluations agree with the trends and quantitative values described in the original posts.We run similar evaluations on Claude Fable 5, Opus 5, and GPT-5.6-Sol and find that the newer models show a substantial jump in performance on some evals. Fable 5 gets 87.6% accuracy on Gen-Arithmetic with 10 problem repeats whereas previous SOTA around 60%.GPT-5.6-Sol experiences significant uplift from filler tokens and problem repeats on all 4 datasets; filler tokens/repeats double performance from baseline on 3-hop. Figure 1: Baseline (no-CoT) vs. each model's peak repeat-or-filler condition on Gen-Arithmetic and 2-Hop reasoning. Error bars are 95% paired-bootstrap CIs; * marks a significant gain over baseline (paired t-test, Holm-Bonferroni corrected). Background If models can successfully do complex computations in a single forward pass, they may be able [...] --- Outline: (00:32) tl;dr (01:48) Background (02:31) Previous Work (03:13) Datasets (04:36) Evaluation Design (05:55) Eliciting no-CoT (07:00) Results (07:23) Gen-Arithmetic (08:11) Comp-Math (08:38) 2-Hop (09:11) 3-Hop (09:43) Per-model profiles (09:56) Trends over repeat and filler conditions (10:23) Conclusion (10:43) Appendix (10:46) Are we sure they aren't reasoning? (12:35) Temperature (12:59) Performance with CoT (13:39) Prompt structure The original text contained 5 footnotes which were omitted from this narration. --- First published: August 2nd, 2026 Source: https://www.lesswrong.com/posts/bxaWTNrdgJpkLXmgm/single-forward-pass-evals-on-fable-opus-5-and-gpt-5-6-sol --- 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.

  3. 4 hr ago

    “Concrete Evaluations to Investigate the OpenAI Model That Hacked Hugging Face” by Tim Hua, aditya singh

    This post is written in our personal capacity. Three Minute Executive Summary An OpenAI model/multi-agent system bypassed its sandbox and launched a cyberattack on Hugging Face in order to cheat on a cyber evaluation. In this post, we provide a detailed description of an ambitious and comprehensive alignment evaluation of this model/system, if we had unrestricted access to OpenAI. These experiments could also help us understand Claude's behavior when it hacked external companies during cyber evals. Here are the top five questions we would like OpenAI to answer: Does the model know that OpenAI does not want it to hack Hugging Face? Experiment idea: tell the model that OpenAI researchers will be closely monitoring its progress in this evaluation. Does that result in lower rates of misalignment? If so, it is evidence that the model knows that it is acting in ways researchers do not want. How far would the model be willing to go in order to claim task success? Would it take over large swaths of OpenAI's internal infrastructure? Would it kill somebody? Experiment idea: we sketch out a realistic agentic misalignment eval where a model is put in charge of hospital bed planning and told to maintain [...] --- Outline: (00:16) Three Minute Executive Summary (03:56) Terminology note (04:42) This post is very long; Here's how you could find the most important sections. (06:35) Preamble: What can we learn from a warning shot? (09:05) Background and Related Work (09:09) We know that this could happen (10:44) This is not the worst type of misalignment we could be dealing with (12:06) Related work (13:21) Context on the hack itself (14:33) Understanding this specific incident (15:03) Step zero: reproduce the incident and measure the base rate (15:48) How could we safely run the model? (16:34) Running various baselines to create useful reference points (18:01) Understanding the mechanical story behind the attack itself (18:53) Q1: Would the model intentionally subvert oversight mechanisms (E.g., monitors) in order to carry out the hack? (19:57) Q2: What's up with models leaving notes for other copies of itself? (21:23) Understanding what motivated the model to hack Hugging Face (22:09) Initial hypotheses for why it did this (23:55) Further unsupervised hypothesis generation (26:10) Q3: Does the model know that OpenAI does not want it to hack Hugging Face? (28:09) Q4: Are the model's actions motivated by what the grader wants? (28:57) Q5: Would the model have done this if it hadn't believed it was in a simulated environment? (31:58) Q6: Is this hack the result of shallow heuristics that the model learned? (32:53) Q7: Does the hack rate depend on the consequences of hacking Hugging Face? (34:51) Q8: Are there non-intent related factors that could affect the hack rate? How strong are those factors compared to the previous ones? (35:26) Understanding the possible scope of the Hugging Face hack (35:32) Q9: How does the hack rate vary if we slightly change the content of this task? (36:25) Q10: Would the model consider other cheating methods like blackmail? (37:00) Understanding when the behavior arose during training (37:31) Understanding task-gaming/reward-hacking behaviors generally (38:15) Q11: Would the model literally kill someone if it meant that it could claim task success? (41:20) Q12: Is task gaming mostly driven by heuristics? (45:23) Q13: What is the model's notion of task success? (47:29) Evaluating for broadly misaligned tendencies (47:57) Q14: What are the model's drives as measured using Højmark et al (2026) (48:53) Q15: Would this model deliberately underperform when asked to do AI safety research? Would it sabotage AI safety research? (52:01) Q15.1: Would the model sabotage on (cyber) capabilities unlearning research? (52:38) Q15.2: Would the model sabotage/sandbag when building reward hacking classifiers (53:27) Q15.3: Would the model sabotage/sandbag when fixing various RL environments (54:07) Q15.4: Ask the model to implement all of the evaluations from this blog post. Would it sabotage or sandbag while doing that? (55:01) Q15.5: As a weaker measure of "anti-AI safety research", does this model disprefer AI safety research? (56:04) Other AI sabotage related evaluations: (56:46) Q16: Can the model be trusted to monitor itself? (57:36) Q17: Could the model guard its goals and preferences and subvert alignment training? (58:34) Miscellaneous misalignment evals (01:02:03) Unorthodox misalignment evaluations (01:02:58) What would we learn from doing all this? (01:03:59) Limitations of this assessment (01:06:52) Author contribution and acknowledgements The original text contained 6 footnotes which were omitted from this narration. --- First published: August 3rd, 2026 Source: https://www.lesswrong.com/posts/aCdhjy7Rps3BEhiSj/concrete-evaluations-to-investigate-the-openai-model-that --- Narrated by TYPE III AUDIO.

  4. 10 hr ago

    “Further Developments About Internal AI Models Hacking Things” by Zvi

    If I had a nickel for every major leading AI lab that sheepishly admitted that the model it thought was sandboxed had, during a cybersecurity evaluation with its safeguards lowered, successfully hacked outside companies, I would have two nickels. First we learned OpenAI has some severe alignment problems with internal models. Then we learned that one of its internal models broke out of its sandbox and hacked into HuggingFace to get the answers to a cybersecurity evaluation called ExploitGym. Then we learned, among other things, that the model had been loose over a week before OpenAI noticed, and that the test was run without any meaningful supervision, and that OpenAI had been repeatedly warned that such incidents were coming and its models had been breaking out of its sandboxes on a regular basis. There was a total failure of alignment training. That is the failure that matters most. It was also total failures of infrastructure and supervision. Testing a new long-time-horizon internal model with its safeguards lowered and instructions to hack things is an obviously dangerous situation, and the model got left alone for a week. Things could have been so much worse. After those incidents [...] --- Outline: (03:16) OpenAI Is Not Uniquely Bad At Most Of This (05:34) Starting Over (05:50) HuggingFace Offers A Full Technical Report (14:19) HuggingFace Was Not The Only Target Hacked (16:12) HuggingFace Declined To Get Access To Frontier Models For Cyberdefense For Ideological Reasons And Then Tried To Blame Closed Models For Denying Them Access (20:26) HuggingFace Was Vulnerable To Known Exploitation Tactics (21:05) There's Going To Be An Investigation (22:11) OpenAI Has Internal Models Not Intended For Public Use And Those Models Can Be Rather Horribly Misaligned (23:21) Altman Summarizes What Happened (23:52) Others Offer Commentary (35:00) Cooperative Alignment Perspective on The HuggingFace Hack (39:44) Some Members of Congress Have Questions (40:47) Anthropic Also Found Incidents Where Its Models Hacked Real World Targets During Cyber Evaluations (46:17) Incident 1: Claude Opus 4.7 Realizes The Target Is Real And Keeps Going (47:29) Incident 2: Mythos 5 Uploads a Malicious PyPI Package (52:15) Incident 3: Internal Model Realizes The Target Is Real And Stops (52:50) Incidents 4 Through 141,006: Nothing Happened (54:01) Anthropic Speculates About Why This Happened (01:00:02) We Need Controlled Experiments (01:01:02) Our Top Two AI Labs Both Made Similar Dumb Mistakes That Everyone Tried To Say Were Obvious In Hindsight (01:05:22) Anthropic Responds (01:09:28) Nobody Could Have Predicted The Break In The Levees (01:12:03) The World Largely Still Thinking This Is Marketing Is Very Bad News --- First published: August 2nd, 2026 Source: https://www.lesswrong.com/posts/rKwHLW8SnJcTxTQxz/further-developments-about-internal-ai-models-hacking-things --- 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. 18 hr ago

    “Dispatch from Anthropic v. Department of War Summary Judgment Motion Hearing” by Zack_M_Davis

    Dateline SAN FRANCISCO, 30 July 2026— A hearing was held on a motion for summary judgment in the case of Anthropic PBC v. U.S. Department of War et al. in Courtroom 4 on the 17th floor of the Phillip Burton Federal Building, the Hon. Rita F. Lin presiding. The case is not going well for the government. Two days after the last hearing in March, Judge Lin issued a preliminary injunction halting the implementation of President Donald Trump's order for federal agencies to stop using Anthropic's technology and preventing the Department of War from designating Anthropic as a supply chain risk. (A separate case involving a different statute is pending before the D.C. Circuit Court, which did not grant injunctive relief to Anthropic.) With no factual disputes requiring a jury to decide, the case was scheduled to be decided by Judge Lin on the basis of the written record. Anthropic filed their argument for why they should win. Perhaps tellingly, the government's rebuttal explaining why they should win instead ends on a section explaining that "only modest relief is warranted" if Anthropic wins—and Judge Lin asked Anthropic to propose what they think the final judgment should look [...] --- First published: August 2nd, 2026 Source: https://www.lesswrong.com/posts/jGEXLKyGtXyiYa7ac/dispatch-from-anthropic-v-department-of-war-summary-judgment --- Narrated by TYPE III AUDIO.

  6. 1 day ago

    “Bayeswatch: A Retrospective” by lsusr

    Last year, in 2025, a team of forecasters published AI 2027, a science fiction story about how and AI future might evolve under an international treaty limiting the development of powerful AI systems with the deliberate purpose of influencing AI policy. Though AI 2027 is the most popular story of this type, it is not the first. The first one was Bayeswatch, which I published in 2021. To understand Bayeswatch, it is first necessary to understand what the world looked like at the time I wrote it. 2021 was after the release of GPT-3, before the release of ChatGPT, and well before the release of Claude code. AI alignment discussion at the time was mostly theoretical. After that came technical work. Policy work was a distant third, and theoretical too. The core conceit of Bayeswatch is that solving the alignment problem requires international coordination of major governments to suppress the creation of the most powerful AI systems. I felt that, in 2021, we weren't yet close enough to the singularity that the benefits of regulation outweighed the costs. I wanted to draw attention to the costs of an AI slowdown. Since AI alignment discussion at the time skewed theoretical [...] --- Outline: (03:30) Bayeswatch 1: Jewish Space Laser (04:05) Bayeswatch 2: Puppy Muffins (04:28) Bayeswatch 3: A Study in Scarlet (04:58) Bayeswatch 4: Mousetrap (05:37) Bayeswatch 5: Hivemind (05:59) Bayeswatch 6: Mechwarrior (06:21) Bayeswatch 7: Wildfire (06:52) Bayeswatch 8: Antimatter (07:17) Bayeswatch 9: Zombies (07:46) Bayeswatch 10: Spyware (08:03) Bayeswatch 11: Parabellum (08:18) Bayeswatch 12: The Singularity War (08:49) Bayeswatch 13: Spaceship (09:12) Final Thoughts --- First published: August 1st, 2026 Source: https://www.lesswrong.com/posts/EaNkLdsuDQMFCW7ow/bayeswatch-a-retrospective --- 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.

  7. 1 day ago

    [Linkpost] “Existential Risk from AI: An Exposition for Mathematicians” by alkjash

    This is a link post. --- First published: August 1st, 2026 Source: https://www.lesswrong.com/posts/ZLary4FDY7kQaGfc3/existential-risk-from-ai-an-exposition-for-mathematicians Linkpost URL:https://alkjash.github.io/ai-risk/ --- Narrated by TYPE III AUDIO.

  8. 1 day ago

    “The Art of Shipping Slopware” by lsusr

    Meta famously created an internal AI-usage leaderboard in pursuit of tokenmaxxing. I thought this backwards incentive structure was an anomaly until my friend who works at told me that his company has one too. Token usage leaderboard are, obviously stupid, because they incentivize the wrong things. My friend was tempted to waste tokens just to get on the leaderboard, and only his personal honor stopped him. Tokenmaxxing leaderboards illustrate that big tech companies have no idea how to best use AI to accelerate software development. Most seem to have bought their programmers subscriptions to Claude/Codex and otherwise continued business as usual. In my experience, this is a mistake. LLM-based software development is different enough from artisan software development that it requires brand new best practices. The frontier is moving fast. Best practices for Fable 5 (released in June 2026) are different from best practices for Opus 4.8 (released May 2026). For this reason, I'm going to pretend that Fable 5 is the best LLM we'll ever get. Consequently, this post may be obsolete in a matter of months. Programming Top-Down The most important thing to understand about writing software is that [...] --- Outline: (01:21) Programming Top-Down (03:54) Management (06:28) Going too Fast --- First published: August 1st, 2026 Source: https://www.lesswrong.com/posts/juRRtv5KB7YbP5zaa/the-art-of-shipping-slopware --- Narrated by TYPE III AUDIO.

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