LessWrong (30+ Karma)

LessWrong

Audio narrations of LessWrong posts.

  1. 1 TIME SIDEN

    “We’re actually running out of benchmarks to upper bound AI capabilities” by LawrenceC

    Written quickly as part of the Inkhaven Residency. Opinions are my own and do not represent METR's official opinion. In early 2025, the situation for upper-bounding[1] model capabilities using fixed benchmarks was already somewhat challenging. As part of the trend where benchmarks were being saturated at an ever increasing rate, benchmarks that were incredibly challenging for AI in early 2024 such as GPQA were being saturated scarcely a year later. An oft-cited screenshot from Our World In Data (including in our time horizon blog post!), showing the ever increasing pace of saturation for AI benchmarks. Thankfully, we saw a wave of alternative approaches to measure AI agent capabilities: for example, at METR, we released both the Time Horizon methodology as well as a preliminary uplift study that found no significant productivity uplift from AI. As part of their frontier AI safety policies, AI developers such as Anthropic and OpenAI built newer, more extensive evaluations to demonstrate that their AIs did not reach dangerous capability thresholds, such as BrowseComp and GDPval. Many research teams, both in academia and in industry, stepped up and created newer, ever more challenging agentic benchmarks, including τ2 -Bench, MCP-Atlas, terminal-bench [...] The original text contained 1 footnote which was omitted from this narration. --- First published: April 6th, 2026 Source: https://www.lesswrong.com/posts/gfkJp8Mr9sBm83Rcz/we-re-actually-running-out-of-benchmarks-to-upper-bound-ai --- 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 min.

  2. 2 TIMER SIDEN

    “Don’t write for LLMs, just record everything” by RobertM

    Some people have argued the advent of LLMs has dramatically increased the value of having a public writing footprint. The first reason given is that this might help secure a meaningful form of immortality. The second reason given is that this might make future LLMs trained on public writing corpora more useful to you, personally, in mundane ways[1]. I think that the first one doesn't check out, the second one is possible but a long-shot, but you can get a lot of the anticipated benefits of the second one by dropping the "public" bit and doing something a little unorthodox. Contra Immortality I don't know if gwern believes in this specific story: two years ago, he wrote a comment which contained the sentence: This seems like a bad move to me on net: you are erasing yourself (facts, values, preferences, goals, identity) from the future, by which I mean, LLMs. That sounds a bit like the immortality/value-propagation argument! But that paragraph ends with: For the trifling cost of some writing, all the worlds' LLM providers are competing to make their LLMs ever more like, and useful to, me. That seems like a "mundane utility" consideration, so who can say [...] --- Outline: (00:43) Contra Immortality (03:44) Contra Mundane Utility from Pretraining (06:48) Pro Panopticon The original text contained 11 footnotes which were omitted from this narration. --- First published: April 6th, 2026 Source: https://www.lesswrong.com/posts/aFo8M9KPRJAnY3csR/don-t-write-for-llms-just-record-everything --- Narrated by TYPE III AUDIO.

    10 min.

  3. 4 TIMER SIDEN

    “Contra Nina Panickssery on advice for children” by Sean Herrington

    I recently read this post by Nina Panickssery on advice for children. I felt that several of the recommendations are actively harmful the children they are aimed at. I am going to assume that this advice is targeted at children who are significantly more intelligent than average and maybe 7-12 years of age? It may be worth reading the original post beforehand, or maybe having it open in another tab while you read through this one. We'll go through the points one by one: "Don't be a sheep". There's a difference between noticing when other people are wrong and actively assuming everyone else is dumb. This leans towards the second. There is a huge amount of evolutionary pressure that has gone into designing kids' behaviour; survival past childhood is pretty important if you want to have kids of your own. Simple examples of childhood behaviour, such as playing, are good illustrations of this design, but even seemingly idiotic activities like eating dirt can help you develop the antibodies you need as you grow up. As a kid, even a smart one, I would basically expect following the crowd to produce better results than figuring everything [...] The original text contained 1 footnote which was omitted from this narration. --- First published: April 6th, 2026 Source: https://www.lesswrong.com/posts/B6PFgvy2hz9wKfq94/contra-nina-panickssery-on-advice-for-children --- Narrated by TYPE III AUDIO.

    6 min.

  4. 5 TIMER SIDEN

    “By Strong Default, ASI Will End Liberal Democracy” by MichaelDickens

    Cross-posted from my website. The existence of liberal democracy—with rule of law, constraints on government power, and enfranchised citizens—relies on a balance of power where individual bad actors can't do too much damage. Artificial superintelligence (ASI), even if it's aligned, would end that balance by default. It is not a question of who develops ASI. Whether the first ASI is developed by a totalitarian state or a democracy, the end result will—by strong default—be a de facto global dictatorship. The central problem is that whoever controls ASI can defeat any opposition. Imagine a scenario where (say) DARPA develops the first superintelligence [1] , and the head of the ASI training program decides to seize power. What can anyone do about it? If the president orders the military to capture DARPA's data centers, the ASI can defeat the military. [2] If Congress issues a mandate that DARPA must turn over control of the ASI, DARPA can refuse, and Congress has even less recourse than the president. If liberal democracy continues to exist, it will only be by the grace of whoever controls ASI. There are two plausible scenarios that have some chance [...] --- Outline: (01:38) What if AI capabilities progress slowly? (02:45) What if the ASI itself protects liberal democracy? (04:46) Liberal democracy is not the true target (05:25) Maybe we can avoid totalitarianism, but there is no clear path The original text contained 2 footnotes which were omitted from this narration. --- First published: April 6th, 2026 Source: https://www.lesswrong.com/posts/gmYTwEyvEsCyhESwh/by-strong-default-asi-will-end-liberal-democracy --- Narrated by TYPE III AUDIO.

    6 min.

  5. 17 TIMER SIDEN

    “AIs can now often do massive easy-to-verify SWE tasks and I’ve updated towards shorter timelines” by ryan_greenblatt

    I've recently updated towards substantially shorter AI timelines and much faster progress in some areas. [1] The largest updates I've made are (1) an almost 2x higher probability of full AI R&D automation by EOY 2028 (I'm now a bit below 30% [2] while I was previously expecting around 15%; my guesses are pretty reflectively unstable) and (2) I expect much stronger short-term performance on massive and pretty difficult but easy-and-cheap-to-verify software engineering (SWE) tasks that don't require that much novel ideation [3] . For instance, I expect that by EOY 2026, AIs will have a 50%-reliability [4] time horizon of years to decades on reasonably difficult easy-and-cheap-to-verify SWE tasks that don't require much ideation (while the high reliability—for instance, 90%—time horizon will be much lower, more like hours or days than months, though this will be very sensitive to the task distribution). In this post, I'll explain why I've made these updates, what I now expect, and implications of this update. I'll refer to "Easy-and-cheap-to-verify SWE tasks" as ES tasks and to "ES tasks that don't require much ideation (as in, don't require 'new' ideas)" as ESNI tasks for brevity. Here are the main drivers of [...] --- Outline: (04:58) Whats going on with these easy-and-cheap-to-verify tasks? (08:17) Some evidence against shorter timelines Ive gotten in the same period (10:46) Why does high performance on ESNI tasks shorten my timelines? (13:15) How much does extremely high performance on ESNI tasks help with AI R&D? (18:22) My experience trying to automate safety research with current models (19:58) My experience seeing if my setup can automate massive ES tasks (21:08) SWE tasks (23:29) AI R&D task (24:20) Cyber (24:41) Appendix: Somewhat more detailed updated timelines The original text contained 13 footnotes which were omitted from this narration. --- First published: April 6th, 2026 Source: https://www.lesswrong.com/posts/dKpC6wHFqDrGZwnah/ais-can-now-often-do-massive-easy-to-verify-swe-tasks-and-i --- 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.

    30 min.

  6. 20 TIMER SIDEN

    “Paper close reading: “Why Language Models Hallucinate”” by LawrenceC

    People often talk about paper reading as a skill, but there aren’t that many examples of people walking through how they do it. Part of this is a problem of supply: it's expensive to document one's thought process for any significant length of time, and there's the additional cost of probably looking quite foolish when doing so. Part of this is simply a question of demand: far more people will read a short paragraph or tweet thread summarizing a paper and offering some pithy comments, than a thousand-word post of someone's train of thought as they look through a paper. Thankfully, I’m willing to risk looking a bit foolish, and I’m pretty unresponsive to demand at this present moment, so I’ll try and write down my thought processes as I read through as much of a a paper I can in 1-2 hours. Standard disclaimers apply: this is unlikely to be fully faithful for numerous reasons, including the fact that I read and think substantially faster than I can type or talk.[1] Specifically, I tried to do this for a paper from last year: “Why Language Models Hallucinate”, by Kalai et al at OpenAI.[2] Due to time [...] --- Outline: (01:25) The Abstract (08:37) The Introduction (08:49) A quick sanity check of examples in the introduction (12:45) A digression on computational learning theory (14:38) Key result #1: relating generation error to binary classification error (15:29) Key result #2: The original text contained 6 footnotes which were omitted from this narration. --- First published: April 5th, 2026 Source: https://www.lesswrong.com/posts/rAjtnXx5qLgubsrGQ/paper-close-reading-why-language-models-hallucinate --- 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.

    18 min.

  7. 1 DAG SIDEN

    “Ten different ways of thinking about Gradual Disempowerment” by David Scott Krueger (formerly: capybaralet)

    About a year ago, we wrote a paper that coined the term “Gradual Disempowerment.” It proved to be a great success, which is terrific. A friend and colleague told me that it was the most discussed paper at DeepMind last year (selection bias, grain of salt, etc.) It spawned articles in the Economist and the Guardian. Most importantly, it entered the lexicon. It's not commonplace for people in AI safety circles and even outside of them to use the term, often in contrast with misalignment or rogue AI. Gradual Disempowerment tends to resonate more than Rogue AI with people outside AI safety circles. But there's still a lot of confusion about what it really is and what it really means. I think it's a very intuitive concept, but also I still feel like I don’t have everything clear in my mind. For instance, I think our paper both introduces the concept and presents a structured argument that it could occur and be catastrophic. But these things seem somewhat jumbled together both in my mind and the discourse.. So for reasons including all of the above, I plan to write a few posts on the topic, starting with [...] The original text contained 4 footnotes which were omitted from this narration. --- First published: April 4th, 2026 Source: https://www.lesswrong.com/posts/W9XQ9CcMTbZQa33eP/ten-different-ways-of-thinking-about-gradual-disempowerment --- Narrated by TYPE III AUDIO.

    10 min.

  8. 1 DAG SIDEN

    “11 pieces of advice for children” by Nina Panickssery

    I came up with these principles when I was a child myself. Don’t be a sheep 🐑. Avoid mindlessly copying others. Resist the urge towards conformity. Think for yourself whether something is worth doing and useful for your goals. If appearing to conform is useful for your goals, think about ways to do the bare minimum. Others are making very many mistakes you don’t want to make, and things can be done much better and more effectively than most people do them. (Be extra aware of this point if you are a girl, girls are naturally drawn towards conformity. Girls must practice not conforming, standing out, being weird, so that they are comfortable with not following the herd when it comes to important matters.)Don’t delude yourself. Sometimes it's useful to pretend to belief a falsehood, but don’t go as far as to start actually believing itself yourself.Related—think freely. Never be afraid to think a thought in the privacy of your own head. All thoughts are thinkable, no matter how scared you might be to express them.Be realistic about your (and others’) natural/genetic qualities. If you are much smarter than others, keep that in mind. If you are [...] The original text contained 2 footnotes which were omitted from this narration. --- First published: April 5th, 2026 Source: https://www.lesswrong.com/posts/cqriGHCwmKZa2v862/11-pieces-of-advice-for-children --- Narrated by TYPE III AUDIO.

    6 min.
Se alle (250)

Om

Audio narrations of LessWrong posts.

Oplysninger

  • Skaber
    LessWrong
  • Aktive år
    2023 - 2026
  • Episoder
    250
  • Vurdering
    Ikke anstødeligt
  • Copyright
    © 2026 All rights reserved
  • Websted for serie
    LessWrong (30+ Karma)

Måske vil du også synes om