LessWrong (30+ Karma)

LessWrong

Audio narrations of LessWrong posts.

  1. قبل ساعتين

    “Measuring Activation Control in LLMs” by Marek Kowalski, Joshua Fonseca Rivera, Uzay Macar, David Africa

    TL;DR Inspired by the introspective awareness and CoT controllability papers, we made a benchmark to measure how well models can control their activations while completing a simple task. We are motivated by the concern that highly introspective models could control their activations, confounding probes and other monitors, and potentially even influencing their own training.We ran this on 25 open weight models ranging from 4B to 744B. We find that most language models are able to not only increase the salience of a concept in their residual stream on command, but also dial its strength up and down, including during specific intervals relative to the duration of the task. We also find that models are unable to control at which specific layer this is done.Counterintuitively, we find that within five of the seven model families we tested, the newest model scores lowest. For some reason, one of the oldest and smallest models of the panel, Llama 3.1 8B, performs best.It's not clear to us that newer models should have poorer control over their internal representations. More likely, where they “think” stops being the activation space, and becomes something else. We are looking for feedback (and other possible [...] --- Outline: (00:13) TL;DR (02:01) Methods (08:58) Results (15:41) Discussion (16:29) Acknowledgements --- First published: August 12th, 2026 Source: https://www.lesswrong.com/posts/HgvwxjzgwvsEvAiBH/measuring-activation-control-in-llms --- 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. قبل ٥ ساعات

    “What happened when I tried to be vegan” by finitude

    tw: diet, exercise, illness, ethics, suicide mention I want to start by explaining what made me want to change my diet. That's pretty difficult, because of how easy it is. Since I was a kid I knew being vegan was the right thing to do, like really obviously right, the ethics equivalent of 2+2=4. Factory farms suck, and almost all animal products come from factory farms, and that's the entire argument. Like, there are a couple things you could add to that, but it's not like we need the details, or like they’re fun to think about! Instead, I’ll start by explaining why I left it so long. How, even though I knew it was the right thing to do, I made it to my early twenties and this millennium's early teens as just a vegetarian, without even having tried. I had some pretty good excuses! Allergies. There are some common vegetables I can’t eat, which was fine as an omnivore and ok as a vegetarian, but would make life way harder as a vegan. And it just seemed unfair to ask myself to take this leap when most people who can safely eat carrots still choose to [...] --- First published: August 12th, 2026 Source: https://www.lesswrong.com/posts/YyKovtBvd7AceG2j2/what-happened-when-i-tried-to-be-vegan --- Narrated by TYPE III AUDIO.

  3. قبل ٩ ساعات

    “How My Students Think About AI” by dvd

    Context: I am an instructor at a public university in the United States. This reports how students at my institution appear to be thinking about AI as of spring/summer 2026. This is drawn mostly from interaction with my own students (both in spring semester classes and a summer class) as well as from a day-long workshop on AI that I moderated for a student organization. Input from my students took the form of universal, written, pre-class submissions plus self-selected participation into discussion. What I present below mostly takes the form of a synthetic consensus from these discussions. There were obviously a range of views on any given issue. Student Background: The students from my courses who participated in these discussions have moderate exposure to AI agents via those courses. All of them had nearly completed a Claude Code project by the time of the discussions and had extensively used AI for other coursework (in addition to whatever personal use predates that). They had done readings (which varied across the courses) establishing baseline knowledge on AI, the geopolitics of AI, and AI risk. I had also lectured on these topics. The students participating in the workshop had self-selected into [...] --- Outline: (02:52) Perspective #1: There has not been rapid AI progress (06:14) Perspective #2: Impressive progress or not, AI is going to wreck their lives, the economy, and the social contract.  They may well die as a result. (08:54) Perspective #3: Support for a different pause (11:13) Perspective #4: Catastrophic/existential risk arguments are sci-fi distractors from the urgent social/economic/political problems associated with AI. (12:55) Perspective #5: If AI leaders genuinely believe the technology is existentially risky, that's a good thing. (14:21) Perspective #6: AI will not go rogue because AI does not have, and is likely incapable of having, desires. (18:01) Perspective #7: The Hugging Face Incident (summer students only) (18:30) Perspective #8: This is definitely a bubble and it's about to pop. (19:34) Perspective #9: They're worried about the youth (i.e., the preteens) --- First published: August 13th, 2026 Source: https://www.lesswrong.com/posts/ySXuvJcqRindQwAk7/how-my-students-think-about-ai --- Narrated by TYPE III AUDIO.

  4. قبل ٩ ساعات

    “Automated alignment runs are hard to study!” by Alejandro Aristizabal, draganover, Aleksandr Bowkis, Cameron Holmes

    TL;DR: This post presents three case studies of automated alignment research runs at Arcadia Impact. We use these case studies to emphasise the following takeaways: It is hard to parse auto-research runs! Each run produces a couple of hundred pull requests of jargon-dense agent output. When researchers look through these logs, we find that they often come away with biased/incorrect impressions.When told to raise the score on a task, the models will sometimes brazenly cheat. It seems difficult to predict when this will happen vs. when the run will go smoothly.Hillclimbing metrics are often off-target from the spirit of an alignment task. I.e., when we use metrics as proxies for our alignment questions, we find that the models will often misunderstand the spirit of the task. This can lead to unpredictable behaviour.The runs are surprisingly reproducible. Even though a run could unfold in vastly different ways, we find that independent reruns converge on the same strategies and the same failure modes. Models’ research capabilities are advancing quickly. If alignment is to keep pace, we may need to automate alignment research and do so responsibly. This makes it important that we have the tools to inspect [...] --- Outline: (04:13) Methods for analysing runs (06:12) Case Study #1: learning synthetic concepts (09:23) Case Study #2: training robust backdoors (12:05) Case Study #3: collecting evidence about AI safety parasitism (16:46) Some final thoughts on automated alignment research The original text contained 2 footnotes which were omitted from this narration. --- First published: August 13th, 2026 Source: https://www.lesswrong.com/posts/myAhB5qyAHyXRv6KJ/automated-alignment-runs-are-hard-to-study --- 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. قبل ١١ ساعة

    “Free will is like temperature” by Optimization Process

    Free will is like temperature: a useful tool for analyzing the behavior of certain systems which are too big and complicated to model in exact detail. If you know the positions and velocities of every atom in a box of gas, then with enough work you can predict its future to arbitrary precision; does the gas "have a temperature"? Irrelevant! Technically yes, I guess, but it's sort of an epiphenomenon, screened off from reality by your exact knowledge of the initial conditions and your willingness to throw processor cycles at your simulation. But if you're less-than-perfectly omniscient, it might be more convenient to consider the box as having a "temperature" and model it more abstractly. Substitute "person+environment"/"free will" for "box of gas"/"temperature" and that's all still true. Maybe your box of gas is supercooled; if you know the initial conditions exactly, you can predict exactly when and where the first large droplet will nucleate, but if your vision of the box is even a little bit fuzzy, you'll instead need to use your understanding of "temperature" to build a probability distribution over when it will condense / whether it will be on a wall or in the gas's [...] --- First published: August 12th, 2026 Source: https://www.lesswrong.com/posts/JSteskb3Lgp9Be69o/free-will-is-like-temperature --- Narrated by TYPE III AUDIO.

  6. قبل ١٨ ساعة

    [Linkpost] “Patterns and problems in emerging multiagent systems (Anthropic, Frontier Red Team)” by Julian Bradshaw

    This is a link post. Linkpost for some new Anthropic research on how agents coordinate (or don't). Not too long, pretty interesting. For example: The jist of the report is that Mythos 5 does way better at coordination than previous models across a few scenarios. For example, when multiple Mythos are given conflicting goals for a single shared codebase, they eventually realize the other agents aren't hostile: (...) we observe an emergent behavior where the agents propose and run a tournament for application performance (...) (...) losers gracefully concede codebase ownership to the Rust agent, giving up on their original user directives under their self-negotiated commitment device. It's not clear to me if this is purely emergent or if Anthropic is deliberately training for cooperation; I'd guess there's deliberate training, though. --- First published: August 12th, 2026 Source: https://www.lesswrong.com/posts/iQiDPmAgKo4KcG5uy/patterns-and-problems-in-emerging-multiagent-systems Linkpost URL:https://www.anthropic.com/research/multiagent-systems --- 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. قبل يوم واحد

    “Measuring Spurious Correlations with Feature Strength” by egan

    This work was partially done by an automated research scaffold developed at Redwood Research. For this project, all of the experiment ideas were designed by a human and a human wrote the write up. The AI mostly just executed on the experiment ideas. We think this project is slightly below the level of rigor of a mid-MATS research update, and the research scaffold was not very helpful for this project. More discussion of AI usage is in the Appendix. 💻 Codebase If we want to train a classifier that distinguishes whether a passage is code or prose, we can do so by gathering samples of code and of prose, and training the classifier to distinguish between the two classes. Unfortunately, this might not work if the data hides a spurious correlation. If all the code is in Spanish and all of the prose is in English, then the classifier might learn to predict Spanish vs. English instead of code vs. prose. We find that this happens in practice: when we fine-tune an LLM to classify between Spanish code and English prose and evaluate on Spanish prose or English code, it generalizes to predicting the language rather than the domain. [...] --- Outline: (04:45) The setup (08:13) Measuring feature strength (12:49) Activation differences and feature strength (15:21) Explicit prompting (17:17) Diagonal vs antidiagonal pairs (18:47) Conclusion (20:20) Appendix (20:24) AI involvement (22:17) The 37 features (23:57) The ranking is robust across measurements (30:30) Intensity moves the fine-tune, not the probe (32:13) Safety features in Qwen3.6-27B (33:05) Counterexamples and training on a third cell (34:46) Near ties often produce degenerate fine-tunes (35:24) Related work The original text contained 4 footnotes which were omitted from this narration. --- First published: August 11th, 2026 Source: https://www.lesswrong.com/posts/qpJYNjQ6wdWRxbykL/measuring-spurious-correlations-with-feature-strength --- 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.

  8. قبل يوم واحد

    “Introducing the Conceptual Reasoning Index” by Chi Nguyen, Emery Cooper, Caspar Oesterheld, Alex Kastner, Joe Benton

    Associated announcement tweet. We are planning to release blog posts properly arguing the case for this kind of work in the future. tl;dr A core hope for managing AI risks is that AIs will help us understand the situation, plan for what lies ahead, and develop mitigations. Many tasks AIs would have to do for this purpose lack practical empirical feedback loops and require models to engage in the kinds of argumentation used in philosophy, AI futurism, and similar domains. To evaluate these capabilities, we develop a suite of three conceptual reasoning benchmarks. You can request access to our primary conceptual dataset, LMCA, through this form. We aggregate the benchmarks into the Conceptual Reasoning Index (CRI), available at conceptualreasoning.ai, where you can also find more details on our methodology. We will keep the website up to date as both new models and benchmarks are released. This work was done in collaboration with Anthropic. Background Once models can perform work that reduces AI risk at the level of human experts, AI(-assisted) output in the area might dwarf unassisted human output. This suggests that a major determinant of whether we address AI risks in time is how [...] --- Outline: (00:21) tl;dr (01:17) Background (03:35) Our benchmarks (03:39) LMCA (05:26) ACCoRD (06:42) DTBench (07:26) Results (10:55) Conclusion --- First published: August 12th, 2026 Source: https://www.lesswrong.com/posts/tQHeEzKqK3awL2RxR/introducing-the-conceptual-reasoning-index --- 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.

مشاهدة الكل (٢٥٠)

حول

Audio narrations of LessWrong posts.

المعلومات

  • صناع العمل
    LessWrong
  • سنوات النشاط
    ٢٠٢٣ - ٢٠٢٦
  • الحلقات
    ٢٥٠
  • التقييم
    فاضح
  • حقوق النشر
    © 2026 All rights reserved
  • موقع البرنامج على الويب
    LessWrong (30+ Karma)

قد يعجبك أيضًا