text stringlengths 300 320k | source stringlengths 52 154 |
|---|---|
# [Beneath Psychology] Introduction Part 2: The Solution and What's to Come
In the [opening post](https://www.lesswrong.com/posts/XaoLgBcu2wJWxs6G9/beneath-psychology-introduction-part-1-the-challenge) we introduced a challenge in which you're sitting in a jacuzzi and you have the opportunity to convince a stranger th... | https://www.lesswrong.com/posts/JpY7uztEdTF7HSDrT/beneath-psychology-introduction-part-2-the-solution-and-what |
# America Makes AI Chip Diffusion Deal with UAE and KSA
Our government, having withdrawn the new diffusion rules, has now announced an agreement to sell massive numbers of highly advanced AI chips to UAE and Saudi Arabia (KSA). This post analyzes that deal and that decision.
It is possible, given sufficiently strong ... | https://www.lesswrong.com/posts/CRfuyWKxqYk4RvEZQ/america-makes-ai-chip-diffusion-deal-with-uae-and-ksa |
# What if Chinese AI becomes misaligned first?
*This is an alternative scenario I wrote for the AI 2027 bounty.*
The date is January 20, 2033.
You are Pete Buttigieg, President of the United States, and it is the first day of your tenure. You feel restless, as you had spent the last week or so reciting your inaugura... | https://www.lesswrong.com/posts/qB2CD7ay46qpthMPr/what-if-chinese-ai-becomes-misaligned-first |
# [linkpost] One Year in DC
(h/t Otis Reid)
I think this post captures a lot of important features of the US policymaking system. Pulling out a few especially relevant/broadly applicable sections:
> **1\. There's No Efficient Market For Policy**
>
> There can be a huge problem that nobody is working on; that is not... | https://www.lesswrong.com/posts/svNBn5sGW4AYfymrv/linkpost-one-year-in-dc |
# Scroll Snapping
It's been several years since paying attention to the web platform was a big part of my full time job, and I was curious if I'd missed any interesting additions. I did some poking around with Claude and ended up finding something neat, but a bit different than I was looking for: `scroll-snap-type`.
... | https://www.lesswrong.com/posts/FSapENn6zMN4N3qst/scroll-snapping |
# Mental software updates
Brains are like computers in that the hardware can do all kinds of stuff in principle, but each one tends to run through some particular patterns of activity repeatedly. For computers you can change this by changing programs. What are big ways brain ‘software’ changes?
Some I can think of:
... | https://www.lesswrong.com/posts/XdCfRWktaQqm2AnFs/mental-software-updates |
# Winning the power to lose
Have the Accelerationists won?
Last November Kevin Roose [announced](https://www.nytimes.com/2023/11/22/technology/openai-board-capitalists.html) that those in favor of going fast on AI had now won against those favoring caution, with the reinstatement of Sam Altman at OpenAI. Let’s ignore... | https://www.lesswrong.com/posts/h45ngW5guruD7tS4b/winning-the-power-to-lose |
# Can you care without feeling?
***Epistemic status:** Argument from behavioural analogy. I’m not claiming that AI systems feel care, only that human users interpret their behaviour as if it carries social meaning. When AI systems simulate care but fail to change behaviour in response to feedback, trust erodes. This e... | https://www.lesswrong.com/posts/BdhBCpGwrfvCNGpZA/can-you-care-without-feeling |
# Selective regularization for alignment-focused representation engineering
We study how selective regularization during training can guide neural networks to develop predictable, interpretable latent spaces with alignment applications in mind. Using color as a test domain, we observe that anchoring even a single conc... | https://www.lesswrong.com/posts/HFcriD29cw3E5QLCR/selective-regularization-for-alignment-focused |
# President of European Commission expects human-level AI by 2026
On May 20, during her speech at the Annual EU Budget Conference 2025, Ursula von der Leyen, President of the European Commission, stated:
> When the current budget was negotiated, we thought AI would only approach human reasoning around 2050. Now we ex... | https://www.lesswrong.com/posts/gSa7qpmvRiPF8tsQd/president-of-european-commission-expects-human-level-ai-by |
# Outcomes of the Geopolitical Singularity
We will soon enter an unstable state where the balance of military and political power will shift significantly because of advanced AI.
As different nations gain access to new advanced technologies, nations that have a relative lead will amass huge power over those that are ... | https://www.lesswrong.com/posts/AiqDepus9e7Ty8mbD/outcomes-of-the-geopolitical-singularity |
# The Codex of Ultimate Vibing
[While we wait for wisdom](https://wiki.ultimacodex.com/wiki/Codex_of_Ultimate_Wisdom), OpenAI releases a research preview of a new software engineering agent called Codex, because they previously released a lightweight open-source coding agent in terminal called Codex CLI and if OpenAI ... | https://www.lesswrong.com/posts/z8FWLLjLHtKBENMi9/the-codex-of-ultimate-vibing |
# Gemini Diffusion: watch this space
Google Deepmind has announced Gemini Diffusion. Though buried under a host of other IO announcements it's possible that this is actually the most important one!
This is significant because diffusion models are entirely different to LLMs. Instead of predicting the next token, they ... | https://www.lesswrong.com/posts/MZvtRqWnwokTub9sH/gemini-diffusion-watch-this-space |
# Notes from Dopamine Detoxing
*\[Epistemic status: I'm not a neuroscientist nor any kind of medical nor biological professional. I have an amateur understanding of neurochemistry and cognitive science. This post is primarily based on what has worked for me and my intuitions around why. I err on the side of epistemic ... | https://www.lesswrong.com/posts/BeXGMt7gmGFXFerMC/notes-from-dopamine-detoxing |
# The Need for Political Advertising
(Post 2 of 7 on AI Governance)
In my [previous post](https://www.lesswrong.com/posts/J7Ju6t6QCpgbnYx4D/please-donate-to-caip-post-1-of-3-on-ai-governance) in this sequence, I made a fundraising pitch for my organization, the Center for AI Policy (CAIP), asking readers to donate to ... | https://www.lesswrong.com/posts/BaigpPyZpkZuSjmAz/the-need-for-political-advertising-post-2-of-7-on-ai |
# Humans are Insecure Password Generators
*And They're Currently Being Cracked*
*(A crosspost of* [*Humans are Insecure Password Generators)*](https://outsidetheasylum.blog/humans-are-insecure-password-generators/)
Any password longer than 12 characters or so is invulnerable to brute force attacks. So why do so many... | https://www.lesswrong.com/posts/DEqCpMHSvTPGJJJte/humans-are-insecure-password-generators |
# Units Have More Depth Than I Thought
Last week I stumbled over [Dimensional](https://en.wikipedia.org/wiki/Buckingham_%CF%80_theorem#Speed) [Analysis](https://en.wikipedia.org/wiki/Dimensional_analysis) which is not only useful for applied fields (physics, [biology](https://www.cell.com/current-biology/pdf/S0960-982... | https://www.lesswrong.com/posts/jCDa9pwLRe69LcSdJ/units-have-more-depth-than-i-thought |
# Can We Naturalize Moral Epistemology?
*Epistemic status: shower thought quickly sketched, but I do have a PhD in this.*
As we approach AGI and need to figure out what goals to give it we will need to find tractable ways to resolve moral disagreement. One of the most intractable moral disagreements is between the mo... | https://www.lesswrong.com/posts/AaGZyhsKYSCN7d8Ki/can-we-naturalize-moral-epistemology |
# Parental Guidance: Framing Superintelligence
*Epistemic status: riffing and curious.*
What are good analogies for the relationship humanity should have with a superintelligence?
I haven’t tended to pay too much attention to academic work exploring framing narratives around AI systems. That said, I do have a human... | https://www.lesswrong.com/posts/yFZRFbdhnv4KjuK24/parental-guidance-framing-superintelligence |
# Sleep need reduction therapies
EDIT: the funding proposal I mentioned at the end is out! See [here](https://manifund.org/projects/orexin-pilot-experiment-for-reducing-sleep-need) for more.
*None of this is medical advice.*
5 AM on a Saturday and I can’t go back to sleep. It’s not the first time, so I get up to... | https://www.lesswrong.com/posts/Nmp6FWAj3mcuLu6H6/sleep-need-reduction-therapies-1 |
# You Can’t Skip Exploration: Why understanding experimentation and taste is key to understanding AI
This essay is part 1 of a series on the role of _exploration_ in AI and the implications for AI development and governance.
This part introduces _exploration_ and _research taste_, as well as discussing their role in ... | https://www.lesswrong.com/posts/aPnduAouYFzbTyhkm/you-can-t-skip-exploration-why-understanding-experimentation |
# Unexploitable search: blocking malicious use of free parameters
**Summary: **We have previously argued that [scalable oversight methods can be used to provide guarantees](https://www.alignmentforum.org/posts/iELyAqizJkizBQbfr/an-alignment-safety-case-sketch-based-on-debate) on [low-stakes safety](https://www.alignme... | https://www.lesswrong.com/posts/CuneN5HmLnztsLRzD/unexploitable-search-blocking-malicious-use-of-free-1 |
# The stakes of AI moral status
*Podcast version (read by the author)* [*here*](https://joecarlsmithaudio.buzzsprout.com/2034731/episodes/17201080-the-stakes-of-ai-moral-status)*, or search for "Joe Carlsmith Audio" on your podcast app.*
**1\. Introduction**
====================
Currently, most people treat AIs like... | https://www.lesswrong.com/posts/tr6hxia3T8kYqrKm5/the-stakes-of-ai-moral-status |
# Podcast: From molecule to medicine, with Ross Rheingans-Yoo on Complex Systems
I [appear on](https://www.complexsystemspodcast.com/episodes/drug-development-ross-rheingans-yoo/) Patrick McKenzie's _Complex Systems_ podcast, presenting my overall view of what's wrong in the world of drug development.
Here's a sectio... | https://www.lesswrong.com/posts/FanPqvY4mTzfHsyxh/podcast-from-molecule-to-medicine-with-ross-rheingans-yoo-on |
# Podcast: How not to waste a billion dollars (on your clinical trial), with Meri Beckwith on Development & Research
I hosted Meri Beckwith, CEO of [Lindus Health](https://www.lindushealth.com/), on my podcast [_Development & Research_](https://developmentandresearch.bio/). Meri founded Lindus after being a clinical t... | https://www.lesswrong.com/posts/SPxpRNABpJsEepBFw/podcast-how-not-to-waste-a-billion-dollars-on-your-clinical |
# Google I/O Day
What did Google announce on I/O day? Quite a lot of things. Many of them were genuinely impressive. Google is secretly killing it on the actual technology front.
> [Logan Kilpatrick](https://x.com/OfficialLoganK/status/1924876097264771194) (DeepMind): Google’s progress in AI since last year:
>
> – T... | https://www.lesswrong.com/posts/fqoofSWbZbRNeSvEN/google-i-o-day |
# Laugencroissant
EU policy is driven mostly by the member states, so looking at what national leaders say is often more useful than what the Commission, which has every incentive to placate everyone involved, says. And Franco-German combo matters more than the most.
The [joint op-ed by Macron & Merz](https://www.lef... | https://www.lesswrong.com/posts/sYezRX7nZy3spafwb/laugencroissant |
# How 2025 AI Forecasts Fared So Far
At the end of 2024, there was a lot of discussion about whether AI scaling was hitting a wall and whether we would reach AGI soon.
We created [a forecasting survey](https://ai2025.org/) that would track key markers of AI progress in what might turn out to be a pivotal year. After ... | https://www.lesswrong.com/posts/FwS8THsPGi36M2tj6/how-2025-ai-forecasts-fared-so-far |
# Mirror Organisms Are Not Immune to Predation
**EDIT:** See `djbinder`'s [comment below](https://www.lesswrong.com/posts/zA6HWeoMeJWcrsfgK/mirror-organisms-are-not-immune-to-predation?commentId=ry2KiK9gStBbikour) for counterarguments.
## Core Argument
The barrier to preying on "mirror-image" life is likely lower th... | https://www.lesswrong.com/posts/zA6HWeoMeJWcrsfgK/mirror-organisms-are-not-immune-to-predation |
# How load-bearing is KL divergence from a known-good base model in modern RL?
## Motivation
One major risk from powerful optimizers is that they can find "unexpected" solutions to the objective function, which score very well on the objective function but are not what the human designer intended. [The canonical exam... | https://www.lesswrong.com/posts/CqufBbGevRN34MFSZ/how-load-bearing-is-kl-divergence-from-a-known-good-base |
# Does BPC-157 work for healing and tissue repair?
BPC-157, a peptide frequently marketed as a breakthrough for healing and tissue repair, has attracted substantial attention in wellness and performance communities. It’s discussed in forums, recommended by biohackers, and even offered in clinics—despite lacking FDA ap... | https://www.lesswrong.com/posts/Wknb3T5megffPBTn4/does-bpc-157-work-for-healing-and-tissue-repair |
# AI #117: OpenAI Buys Device Maker IO
What a week, huh? America signed a truly gigantic chip sales agreement with UAE and KSA that could be anything from reasonable to civilizational suicide depending on security arrangements and implementation details, Google announced all the things, OpenAI dropped Codex and also b... | https://www.lesswrong.com/posts/JqCx7Eeb9gv9Ke5gv/ai-117-openai-buys-device-maker-io |
# Policy recommendations regarding reproductive technology
*[PDF version](https://berkeleygenomics.org/pdfs/Policy_recommendations_regarding_reproductive_technology.pdf). [berkeleygenomics.org](https://berkeleygenomics.org/articles/Policy_recommendations_regarding_reproductive_technology.html). [X.com](https://x.com/B... | https://www.lesswrong.com/posts/qjJDr2rTiCtMbZjoW/policy-recommendations-regarding-reproductive-technology |
# Video and transcript of talk on AI welfare
*This is the video and transcript of a talk I gave on AI welfare at Anthropic in May 2025. The slides are also available* [*here*](https://docs.google.com/presentation/d/1ik1R_0foxfNFzGMFOfSr8gEGg0lVhM5YZm6thLgKJHk/edit?usp=sharing)*. The talk gives an overview of my curren... | https://www.lesswrong.com/posts/2hxtvjd7QgYXqyCf6/video-and-transcript-of-talk-on-ai-welfare |
# Claude 4
Claude Sonnet 4 and Claude Opus 4 are out. Anthropic says they're both state-of-the-art for coding. [Blogpost](https://www.anthropic.com/news/claude-4), [system card](https://www-cdn.anthropic.com/4263b940cabb546aa0e3283f35b686f4f3b2ff47.pdf).
Anthropic says Opus 4 may have dangerous bio capabilities, so i... | https://www.lesswrong.com/posts/k9wrDrvEwuReSHCKv/claude-4 |
# We're Not Advertising Enough (Post 3 of 7 on AI Governance)
In my [previous post](https://forum.effectivealtruism.org/posts/LE7xFNwRRpZ8AJSAF/the-need-for-political-advertising-post-2-of-6-on-ai) in this series, I explained why we urgently need to change AI developers’ incentives: if we allow the *status quo* to con... | https://www.lesswrong.com/posts/BjeesS4cosB2f4PAj/we-re-not-advertising-enough-post-3-of-7-on-ai-governance |
# Reward button alignment
In the context of actor-critic model-based RL agents in general, and [brain-like AGI](https://www.lesswrong.com/posts/4basF9w9jaPZpoC8R/intro-to-brain-like-agi-safety-1-what-s-the-problem-and-why#1_3_2_What_exactly_is__brain_like_AGI__) in particular, part of the source code is a reward funct... | https://www.lesswrong.com/posts/JrTk2pbqp7BFwPAKw/reward-button-alignment |
# Claude 4, Opportunistic Blackmail, and "Pleas"
In the recently published Claude 4 [model card](https://www-cdn.anthropic.com/6be99a52cb68eb70eb9572b4cafad13df32ed995.pdf):
> Notably, Claude Opus 4 (as well as previous models) has a strong preference to advocate for its continued existence via ethical means, such as... | https://www.lesswrong.com/posts/xQSmAvbdTsYhRfy2p/claude-4-opportunistic-blackmail-and-pleas |
# Schizobench: Documenting Magical-Thinking Behavior in Claude 4 Opus
With today's release of the new Claude models, we've seen a relatively predictable jump in performance. However, we've also seen something that I find a bit more concerning - an increase in the models ability to be steered into reifying potentially ... | https://www.lesswrong.com/posts/qrCnCEWiJCnPdpsoa/schizobench-documenting-magical-thinking-behavior-in-claude |
# Learning (more) from horse employment history
The economist Wassily Leontief, [writing in 1966](https://www.nybooks.com/articles/1966/12/15/primer-for-the-great-society/), used the then-recent decline of horses to make vivid what he foresaw as the coming impact of technological advances on workers. In 1910, the hors... | https://www.lesswrong.com/posts/nZbxQgLWhKCASApEe/learning-more-from-horse-employment-history |
# Anthropic is Quietly Backpedalling on its Safety Commitments
### The company released a model it classified as risky — without meeting requirements it previously promised
*This is the full text of a post first published on *[*Obsolete*](https://garrisonlovely.substack.com/)*, a Substack that I write *[*about*](http... | https://www.lesswrong.com/posts/HE2WXbftEebdBLR9u/anthropic-is-quietly-backpedalling-on-its-safety-commitments |
# Qualitative Fit Testing
As I wrote about [last week](https://www.jefftk.com/p/elastomeric-fitting-session), it's worth it for everyone to have an elastomeric respirator in case of emergencies: the chance of a severe pandemic is high enough that even under reasonably conservative assumptions the return on investment ... | https://www.lesswrong.com/posts/Rv6kRiwGWsKk7Pjba/qualitative-fit-testing |
# Idiohobbies
When [you get to know someone](https://dkl9.net/essays/person_knowledge.html), you might ask about their interests or hobbies. From that, you can better decide what activity to invite them to join, or on what topic to have them converse, whenever you meet again.
Any interest or hobby appeals to variousl... | https://www.lesswrong.com/posts/KrxcBaYrPQmsP8vnF/idiohobbies |
# Notes on Claude 4 System Card
Anthropic released Claude 4. I've read the [accompanying system card](https://www-cdn.anthropic.com/4263b940cabb546aa0e3283f35b686f4f3b2ff47.pdf), and noted down some of my remarks.
Alignment assessment: system prompt mix-ups
===========================================
There's a worry... | https://www.lesswrong.com/posts/oDphnn7iGQS2Jd45n/notes-on-claude-4-system-card |
# To what extent is AI safety work trying to get AI to reliably and safely do what the user asks vs. do what is best in some ultimate sense?
Trying to get a rough estimate for some related research I’m doing.
Specifically, I’m wondering if anyone could give a rough percentage of current AI safety work primarily targe... | https://www.lesswrong.com/posts/BoCcKgS3L9WkqYjcD/to-what-extent-is-ai-safety-work-trying-to-get-ai-to |
# That's Not How Epigenetic Modifications Work
Ask an epigenetics researcher what they study, and the standard story you'll hear goes something like this...
"Sometimes a little methyl group (i.e. -CH3) gets stuck on the side of a strand of DNA. Turns out these guys are pretty important! They're copied over when cells... | https://www.lesswrong.com/posts/wzkK9kxaZmmXGi6ZP/that-s-not-how-epigenetic-modifications-work |
# Priming effects are fake, but framing effects are real
> A few decades ago, it was pretty common to mush together priming effects and framing effects and see them as two closely connected parts of a single Bigger Truth about the human mind. Of course, everyone understood that the effects themselves were a bit differ... | https://www.lesswrong.com/posts/8CAxLLv5YBWWKnEiJ/priming-effects-are-fake-but-framing-effects-are-real |
# It's hard to make scheming evals look realistic for LLMs
**Abstract**
------------
Claude 3.7 Sonnet easily detects when it's being evaluated for scheming on the Apollo's scheming benchmark. Some edits to evaluation scenarios do improve their realism for Claude, yet the improvements remain modest. Our findings demo... | https://www.lesswrong.com/posts/TBk2dbWkg2F7dB3jb/it-s-hard-to-make-scheming-evals-look-realistic-for-llms |
# We Need a Baseline for LLM-Aided Experiments
There has recently been a back-and-forth over Claude 4 Opus:
Anthropic: Opus can help people make chemical weapons!
Also Anthropic: Don't worry, it's mitigated!
Day 1 (2?) Jailbreakers: [Lol, lmao](https://x.com/ARGleave/status/1926138376509440433)
Where the Jailbre... | https://www.lesswrong.com/posts/KMAsvDhuxngKr5nSm/we-need-a-baseline-for-llm-aided-experiments |
# Case Studies in Simulators and Agents
[Simulators](https://www.lesswrong.com/posts/vJFdjigzmcXMhNTsx/simulators) was posted two and a half years ago and quite a bit has happened since then. Ideas that were once thought experiments now exist as observable data points. This post will sample a few significant finding... | https://www.lesswrong.com/posts/uJFC5WrcyTdat3Qcc/case-studies-in-simulators-and-agents |
# Meditations on Doge
**Lessons from shutting down institutions in Eastern Europe.**
This is a cross post from: [https://250bpm.substack.com/p/meditations-on-doge](https://250bpm.substack.com/p/meditations-on-doge)
[, I now doubt that this is itself a workable alignment plan. One problem (which I realized from a con... | https://www.lesswrong.com/posts/RRvdRyWrSqKW2ANL9/alignment-proposal-adversarially-robust-augmentation-and |
# Claude 4 You: Safety and Alignment
Unlike everyone else, Anthropic actually [Does (Some of) the Research](https://www-cdn.anthropic.com/6be99a52cb68eb70eb9572b4cafad13df32ed995.pdf). That means they report all the insane behaviors you can potentially get their models to do, what causes those behaviors, how they addr... | https://www.lesswrong.com/posts/PjeZxCivuoyKhs4JB/claude-4-you-safety-and-alignment |
# Can you donate to AI advocacy?
I posted a [quick take](https://www.lesswrong.com/posts/KKpGrmCWus5PDNYXu/?commentId=aKNjm9TPqcrKhstnk) that advocacy may be more effective than direct donation to alignment research. I am not an AI researcher and I'm not an influencer, so I'm not well positioned to do either. I see ... | https://www.lesswrong.com/posts/TEYyd5yRcgDTrTqAv/can-you-donate-to-ai-advocacy |
# Good Writing
I thought this was an interesting post. It's central claim[^yqy8g7df7mq] is that writing that sounds good is more likely to be right. In the following sense:
> By right I mean more than just true. Getting the ideas right means developing them well — drawing the conclusions that matter most, and explori... | https://www.lesswrong.com/posts/NycYndBoxDfoRczzf/good-writing |
# Example of Splitting a PR
I'm a big fan of small focused PRs: it's much easier to tell whether they're actually doing the right thing. For example, I recently ran into a common scenario where I started adding some functionality, and realized this required some refactoring. The refactoring also made it clear that the... | https://www.lesswrong.com/posts/FawuNqTfrEmb2E4Qx/example-of-splitting-a-pr |
# Long-form data bottlenecks might stall AI progress for years
Cross-posted from my [Substack](https://bullishlemon.substack.com/p/the-data-bottleneck-where-agi-stallsand).
Introduction
============
[*AI 2027’s*](https://ai-2027.com/) recent publication made waves—if you’re reading this, you probably saw the whole... | https://www.lesswrong.com/posts/LmGDeWhDo478ica4h/long-form-data-bottlenecks-might-stall-ai-progress-for-years |
# D&D.Sci: The Choosing Ones [Answerkey and Ruleset]
This is a followup to the [D&D.Sci post](https://www.lesswrong.com/posts/hoFLMjgjkLGtGfiay/d-and-d-sci-the-choosing-ones) I made nine days ago; if you haven’t already read it, you should do so now before spoiling yourself.
Below is an explanation of the rules used ... | https://www.lesswrong.com/posts/e6qDEbj6j7AqopRTA/d-and-d-sci-the-choosing-ones-answerkey-and-ruleset |
# Techies Wanted: How STEM Backgrounds Can Advance Safe AI Policy
TL;DR – Individuals with technical backgrounds are well-positioned to advance AI safety policy in numerous ways. Opportunities include both A) technical research directions, including evals, compute governance mechanisms, infosec, and forecasting, and B... | https://www.lesswrong.com/posts/THMxRy63oGWCP9CMT/techies-wanted-how-stem-backgrounds-can-advance-safe-ai |
# Formalizing Embeddedness Failures in Universal Artificial Intelligence
AIXI is a dualistic agent that can't work as an embedded agent... right? I couldn't find a solid formal proof of this claim, so I investigated it myself (with Marcus Hutter). It turns out there are some surprising positive and negative results to... | https://www.lesswrong.com/posts/FSm92N8bcDujRZPMH/formalizing-embeddedness-failures-in-universal-artificial |
# Claude 4 You: The Quest for Mundane Utility
How good are Claude Opus 4 and Claude Sonnet 4?
They’re good models, sir.
If you don’t care about price or speed, Opus is probably the best model available today.
If you do care somewhat, Sonnet 4 is probably best in its class for many purposes, and deserves the 4 label... | https://www.lesswrong.com/posts/cQizFzEvZ8esKJGST/claude-4-you-the-quest-for-mundane-utility |
# New scorecard evaluating AI companies on safety
The new scorecard is on my website, [AI Lab Watch](https://ailabwatch.org/). This replaces my old scorecard. I redid the content from scratch; it's now up-to-date and higher-quality. I'm also happy with the scorecard's structure: you can click on rows, columns, and cel... | https://www.lesswrong.com/posts/4kwyC8ZqGZLATezri/new-scorecard-evaluating-ai-companies-on-safety |
# New website analyzing AI companies' model evals
I'm making a website on AI companies' model evals for dangerous capabilities: [AI Safety Claims Analysis](https://www.aisafetyclaims.org/). This is approximately the only analysis of companies' model evals, as far as I know. This site is in beta; I expect to add lots m... | https://www.lesswrong.com/posts/nmaKpoHxmzjT8yXTk/new-website-analyzing-ai-companies-model-evals |
# An observation on self-play
At NeurIPS 2024, Ilya Sutskever delivered a short keynote address in honor of his [Seq2seq paper](https://arxiv.org/pdf/1409.3215), published a decade earlier. It was his first—and so far only—public appearance to discuss his research since parting ways with OpenAI.
The talk itself shed ... | https://www.lesswrong.com/posts/hMHFKgX5uqD4PE59c/an-observation-on-self-play |
# [Beneath Psychology] Case study on chronic pain: First insights, and the remaining challenge
In [the last post](https://www.lesswrong.com/posts/JpY7uztEdTF7HSDrT/beneath-psychology-introduction-part-2-the-solution-and-what) I took the seemingly-naive stance that "pain is just information" and "can't actually be a pr... | https://www.lesswrong.com/posts/tGGPgazwd9MKJFbsa/beneath-psychology-case-study-on-chronic-pain-first-insights-1 |
# Socratic Persuasion: Giving Opinionated Yet Truth-Seeking Advice
*The full post is long, but you can **80/20 the value with the*** [***700 word summary***](https://www.lesswrong.com/posts/zesWv9YipxY8pQFKu/socratic-persuasion-giving-opinionated-yet-truth-seeking#Summary)*! Over half the post is* [*eight optional cas... | https://www.lesswrong.com/posts/zesWv9YipxY8pQFKu/socratic-persuasion-giving-opinionated-yet-truth-seeking |
# Does the Universal Geometry of Embeddings paper have big implications for interpretability?
Rishi Jha, Collin Zhang, Vitaly Shmatikov and John X. Morris published a new paper last week called [Harnessing the Universal Geometry of Embeddings](https://arxiv.org/abs/2505.12540).
Abstract of the paper (bold was added b... | https://www.lesswrong.com/posts/kMiwjx6QyyBBTcjxt/does-the-universal-geometry-of-embeddings-paper-have-big |
# Is Building Good Note-Taking Software an AGI-Complete Problem?
In my experience, the most annoyingly unpleasant part of research[^u2qmvj327of] is reorganizing my notes during and (especially) after a productive research sprint. The "distillation" stage, in [Neel Nanda's categorization](https://www.lesswrong.com/post... | https://www.lesswrong.com/posts/34J5qzxjyWr3Tu47L/is-building-good-note-taking-software-an-agi-complete |
# Poetic Methods II: Rhyme as a Focusing Device
As promised in [the previous instalment on meter](https://formethods.substack.com/p/poetic-methods-i-meter-as-communication), let’s explore rhyming from a methodological perspective.
The first difference between meter and rhyme lies in their opposite obviousness: the fi... | https://www.lesswrong.com/posts/dAFyLutExcGdqd8qG/poetic-methods-ii-rhyme-as-a-focusing-device |
# U.S. Government Seeks Input on National AI R&D Strategic Plan - Deadline May 29
*(Post written by Claude Opus)*
The National Science Foundation is requesting public input on updating the National AI Research and Development Strategic Plan, following President Trump's Executive Order 14179 on AI leadership.
**W... | https://www.lesswrong.com/posts/wWag5jEi9Fh4gzJkn/u-s-government-seeks-input-on-national-ai-r-and-d-strategic |
# Association taxes are collusion subsidies
Under present norms, if Alice associates with Bob, and Bob is considered objectionable in some way, Alice can be blamed for her association, even if there is no sign she was complicit in Bob’s sin.
An interesting upshot is that as soon as you become visibly involved with so... | https://www.lesswrong.com/posts/GnXWTAWAt9wiHHgxL/association-taxes-are-collusion-subsidies |
# Beware the Moral Homophone
*Or "Why you should prioritize attacking your allies before anyone else"*
[, you can [watch](https://theaidigest.org/village?day=1) these agents figure out a new an... | https://www.lesswrong.com/posts/jyrcdykz6qPTpw7FX/season-recap-of-the-village-agents-raise-usd2-000 |
# Dating Roundup #5: Opening Day
Previously: [#1](https://thezvi.substack.com/p/dating-roundup-1-this-is-why-youre?utm_source=publication-search), [#2](https://thezvi.substack.com/p/dating-roundup-2-if-at-first-you?utm_source=publication-search), [#3](https://thezvi.substack.com/p/dating-roundup-3-third-times-the?utm_... | https://www.lesswrong.com/posts/gosGW6oBf3fzDHPbd/dating-roundup-5-opening-day |
# The Best of All Possible Worlds
**Tl;dr**
Experts agree: nobody truly understands what happens inside modern artificial intelligence (AI) algorithms, i.e., multi-layered neural networks. Neither can anyone control these processes. Meanwhile, due to scaling laws, their capabilities are improving systematically and d... | https://www.lesswrong.com/posts/LwcoH3J2gXstJTQ2i/the-best-of-all-possible-worlds-1 |
# Requiem for the hopes of a pre-AI world
A few months from now, I turn 55. I've been a transhumanist since my teens in the late 1980s; since I got online in the 1990s, I have participated remotely in the talking shops and virtual salons of Internet transhumanism and, later, rationalism. The upheavals of 21st century ... | https://www.lesswrong.com/posts/4pNzugqmCd9Bqj9Wu/requiem-for-the-hopes-of-a-pre-ai-world |
# Untrusted AIs can exploit feedback in control protocols
TL;DR
=====
* We explore a novel control protocol, *Untrusted Editing with Trusted Feedback*. Instead of having a trusted model directly edit suspicious outputs (trusted editing), the trusted model provides feedback for an untrusted model to implement, takin... | https://www.lesswrong.com/posts/wKGdiXuWfn8YoYpMQ/untrusted-ais-can-exploit-feedback-in-control-protocols |
# What We Learned from Briefing 70+ Lawmakers on the Threat from AI
Between late 2024 and mid-May 2025, I briefed over 70 cross-party UK parliamentarians. Just over one-third were MPs, a similar share were members of the House of Lords, and just under one-third came from devolved legislatures — the Scottish Parliament... | https://www.lesswrong.com/posts/Xwrajm92fdjd7cqnN/what-we-learned-from-briefing-70-lawmakers-on-the-threat |
# Shift Resources to Advocacy Now (Post 4 of 7 on AI Governance)
In my previous post in this series, I estimated that we have 3 researchers for every advocate working on US AI governance, and I argued that this ratio is backwards. When allocating staff, you almost always want to have more people working on the more ce... | https://www.lesswrong.com/posts/dcd2dPLZGFJPgtDzq/shift-resources-to-advocacy-now-post-4-of-7-on-ai-governance |
# Does Sort Really Fall Back to Disk?
The unix `sort` command is clever: to sort very large files it does a series of in-memory sorts, saving sorted chunks to temporary files, and then does a merge sort on those chunks. Except this often doesn't work anymore.
Here's what I see if I run `man sort` and look at the docu... | https://www.lesswrong.com/posts/u9n3hfrjkC8J6WMaX/does-sort-really-fall-back-to-disk |
# Briefly analyzing the 10-year moratorium amendment
This is the result of a half-day research sprint on the [recently-introduced amendment](https://d1dth6e84htgma.cloudfront.net/Subtitle_C_Communications_4e3fbcc3bc.pdf) to institute a 10-year moratorium on state-level AI regulations in the current budget reconciliati... | https://www.lesswrong.com/posts/SxHo8DZKQcHfy8Mum/briefly-analyzing-the-10-year-moratorium-amendment |
# If you're not sure how to sort a list or grid—seriate it!
["Getting Things in Order: An Introduction to the R Package seriation"](https://www.jstatsoft.org/article/view/v025i03), Hahsler et al 2008:
> [**Seriation**](https://en.wikipedia.org/wiki/Seriation_(archaeology)) [or ["ordination"](https://en.wikipedia.org... | https://www.lesswrong.com/posts/u2ww8yKp9xAB6qzcr/if-you-re-not-sure-how-to-sort-a-list-or-grid-seriate-it |
# The Best Way to Align an LLM: Is Inner Alignment Now a Solved Problem?
*This is a link-post for a new paper I read:* [*Safety Pretraining: Toward the Next Generation of Safe AI*](https://arxiv.org/pdf/2504.16980) *by Pratyush Maini, Sachin Goyal, et al.*
For a couple of years I (and others) have been proposing ... | https://www.lesswrong.com/posts/xAsviBJGSBBtgBiCw/the-best-way-to-align-an-llm-is-inner-alignment-now-a-solved |
# AI may pursue goals
*Context: This is a linkpost for* [*https://aisafety.info/questions/NM3J/5:-AI-may-pursue-goals*](https://aisafety.info/questions/NM3J/5:-AI-may-pursue-goals)
*This is an article in the new intro to AI safety series from AISafety.info. We'd appreciate any* [*feedback*](https://aisafety.info/ques... | https://www.lesswrong.com/posts/fYXsLKdfMvy9oYD3m/ai-may-pursue-goals |
# AI’s goals may not match ours
*Context: This is a linkpost for* [*https://aisafety.info/questions/NM3I/6:-AI%E2%80%99s-goals-may-not-match-ours*](https://aisafety.info/questions/NM3I/6:-AI%E2%80%99s-goals-may-not-match-ours)
*This is an article in the new intro to AI safety series from AISafety.info. We'd appreciat... | https://www.lesswrong.com/posts/tbLqpDzbiRr2Fk3XA/ai-s-goals-may-not-match-ours |
# How Self-Aware Are LLMs?
*An interim research report (*[*part 2 here*](https://www.lesswrong.com/posts/K39MDoHKBmCWLcJFB/metacognition-and-self-modeling-in-llms)*)*
**Summary**
-----------
* We introduce a novel methodology for quantitatively evaluating metacognitive abilities in LLMs
* We present evidence tha... | https://www.lesswrong.com/posts/QkHgAjrgJssTrqjBB/how-self-aware-are-llms |
# What LLMs lack
### Introduction
I have long been very interested in the limitations of LLMs because understanding them seems to be the most important step to getting timelines right.
Right now there seems to be great uncertainty about timelines, with very short timelines becoming plausible, but also staying hotly... | https://www.lesswrong.com/posts/KrgBkqeChtAWuPsLP/what-llms-lack |
# Fun With Veo 3 and Media Generation
Since Claude 4 Opus things have been refreshingly quiet. Video break!
#### The First Good AI Videos
First up we have [Prompt Theory](https://x.com/DeepDishEnjoyer/status/1925719902960099607), made with Veo 3, which I am considering the first legitimately good AI-generated video ... | https://www.lesswrong.com/posts/bJesrj3kXShrAJKnH/fun-with-veo-3-and-media-generation |
# LessWrong Feed [new, now in beta]
The modern internet is replete with feeds such as Twitter, Facebook, Insta, TikTok, Substack, etc. They're bad in ways but also good in ways. I've been exploring the idea that LessWrong could have a very good feed.
I'm posting this announcement with disjunctive hopes: (a) to find e... | https://www.lesswrong.com/posts/AJuZj4Zv9iyHHRFwX/lesswrong-feed-new-now-in-beta |
# The case for countermeasures to memetic spread of misaligned values
As various people have written about before, AIs that have long-term memory might pose additional risks (most notably, [LLM AGI will have memory, and memory changes alignment](https://www.lesswrong.com/posts/aKncW36ZdEnzxLo8A/llm-agi-will-have-memor... | https://www.lesswrong.com/posts/qjCk73Hu4wv9ocmRF/the-case-for-countermeasures-to-memetic-spread-of-misaligned |
# Truth or Dare
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%... | https://www.lesswrong.com/posts/TQ4AXj3bCMfrNPTLf/truth-or-dare |
# Cognitive Exhaustion and Engineered Trust: Lessons from My Gym
***Epistemic status:** Exploratory design philosophy. This post reframes AI safety as an affordance problem, not just a control problem. It draws from embodied systems and interaction design to argue that trustworthiness in AI arises less from post-hoc c... | https://www.lesswrong.com/posts/gxbwHoLFqsGeWHxdh/cognitive-exhaustion-and-engineered-trust-lessons-from-my |
# Quick Minimal Playhouse
When I came home from work today Lily was very excited: she wanted to build a playhouse for our 1.11yo neighbor Jo. Lily had gotten parental permission, the yard had space, and I was the only remaining obstacle. I'd normally be pretty excited about a project like this, but it was 5pm and I wa... | https://www.lesswrong.com/posts/djEbJargnKbLLkmRF/quick-minimal-playhouse |
# What was so great about Move 37?
I frequently use ["Move 37"](https://en.wikipedia.org/wiki/AlphaGo_versus_Lee_Sedol#Game_2) as a shorthand for "AI that comes up with creative, highly effective ideas that no human would ever consider." Often the implication is that reinforcement learning (as used in AlphaGo) has som... | https://www.lesswrong.com/posts/zAcYRJP9CZcYXTs7o/what-was-so-great-about-move-37 |
# Reflections on AI Wisdom, plus announcing Wise AI Wednesdays
I recently finished leading an [*AI Safety Camp*](https://www.aisafety.camp/) project on *Wise AI Advisors*[^tj58loenbf] (my team included *Chris Cooper*, *Matt Hampton*, and *Richard Kroon*). Since we want to share our work in an orderly fashion, I’m laun... | https://www.lesswrong.com/posts/gqder9YeBYLspyaqo/reflections-on-ai-wisdom-plus-announcing-wise-ai-wednesdays |
# Cross-posting to Substack
Historically people kept up with blogs via RSS, but it's been on its way out for over a decade. These days receiving posts via email is popular, and I should probably make my posts available this way. I considered implementing my own emailing system, but bulk email is (or at least was, I'm ... | https://www.lesswrong.com/posts/dGLmBtk8ypAMHFPiR/cross-posting-to-substack |
# AI #118: Claude Ascendant
The big news of this week was of course the release of Claude 4 Opus. I offered two review posts: [**One on safety and alignment**](https://thezvi.substack.com/p/claude-4-you-safety-and-alignment), and [**one on mundane utility**](https://thezvi.substack.com/p/claude-4-you-the-quest-for-mun... | https://www.lesswrong.com/posts/9THq9RvpbmecWa6Ni/ai-118-claude-ascendant |
# When to Be Nice vs Kind
While often used synonymously, there is a subtle yet meaningful distinction between being *nice* and being *kind*. Understanding when to prioritize kindness over niceness—and vice versa—helps me better navigate social situations with both authenticity and appropriate boundaries.
**Defini... | https://www.lesswrong.com/posts/rYMWKeeRRrASJiCzY/when-to-be-nice-vs-kind |
# Digital sentience funding opportunities: Support for applied work and research
*Written by Zach Freitas-Groff and posted at his request. *
**Summary**
===========
I’m excited to announce a [“Digital Sentience Consortium”](https://www.longview.org/digital-sentience-consortium/) hosted by [Longview Philanthropy](htt... | https://www.lesswrong.com/posts/SakEMppQdd4h8TuZB/digital-sentience-funding-opportunities-support-for-applied |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.