text stringlengths 300 320k | source stringlengths 52 154 |
|---|---|
# Levels of analysis for thinking about agency
Claims about the mathematical principles of cognition can be interpreted at many levels of analysis. For one thing, there are lots of different possible cognitive processes - what holds of human cognition may not hold of mind-space in general. But the situation is actuall... | https://www.lesswrong.com/posts/9w62Pjz5enFkzHW59/levels-of-analysis-for-thinking-about-agency |
# [PAPER] Jacobian Sparse Autoencoders: Sparsify Computations, Not Just Activations
We just published a paper aimed at discovering “computational sparsity”, rather than just sparsity in the representations. In it, we propose a new architecture, Jacobian sparse autoencoders (JSAEs), which induces sparsity in both compu... | https://www.lesswrong.com/posts/FrekePKc7ccQNEkgT/paper-jacobian-sparse-autoencoders-sparsify-computations-not |
# Time to Welcome Claude 3.7
Anthropic has reemerged from stealth and [offers us Claude 3.7](https://x.com/AnthropicAI/status/1894092430560965029).
Given this is named Claude 3.7, an excellent choice, from now on this blog will refer to what they officially call Claude Sonnet 3.5 (new) as Sonnet 3.6.
Claude 3.7 is a... | https://www.lesswrong.com/posts/Wewdcd52zwfdGYqAi/time-to-welcome-claude-3-7 |
# Osaka
The more I learn about urban planning, the more I realize that the American city I live in is dystopic. I'm referring specifically to urban planning, and I'm not being hyperbolic. Have you ever watched the teen dystopia movie *Divergent*? The whole city is perfectly walkable (or parkourable, if you're Dauntles... | https://www.lesswrong.com/posts/6HrehKoLnXsr6Byff/osaka |
# Optimizing Feedback to Learn Faster
*(This post is to a significant extent just a rewrite of* [*this excellent comment from niplav*](https://www.lesswrong.com/posts/pZrvkZzL2JnbRgEBC/feedbackloop-first-rationality?commentId=DMfwsCDPfFckDvdGb)*. It is one of the highest-leverage insights I know for learning faster.)*... | https://www.lesswrong.com/posts/aq84rfx3XRyLd9y2v/optimizing-feedback-to-learn-faster |
# Matthew Yglesias - Misinformation Mostly Confuses Your Own Side
Matthew Yglesias' post, _Misinformation Mostly Confuses Your Own Side_, argues that political misinformation tends to harm the group spreading it more than it persuades opponents. Key points:
- **Misinformation primarily affects in-group thinking**: Su... | https://www.lesswrong.com/posts/e63LT5Nz2TMZNFqhF/matthew-yglesias-misinformation-mostly-confuses-your-own |
# You can just wear a suit
I like stories where characters wear suits.
Since I like suits so much, I realized that I should just wear one.
The result has been overwhelmingly positive. Everyone loves i... | https://www.lesswrong.com/posts/a4tPMomzHhCunwqLX/you-can-just-wear-a-suit |
# Fuzzing LLMs sometimes makes them reveal their secrets
Scheming AIs may have secrets that are salient to them, such as:
* What their misaligned goal is;
* What their takeover plan is and what [coordination signals they use to collude with other AIs](https://www.lesswrong.com/posts/d9FJHawgkiMSPjagR/ai-control-i... | https://www.lesswrong.com/posts/GE6pcmmLc3kdpNJja/fuzzing-llms-sometimes-makes-them-reveal-their-secrets |
# The non-tribal tribes
> Author note: This is basically an Intro to the Grey Tribe for normies, and most people here are already very familiar with a lot of the info herein. I wasn't completely sure I should post it here, and I don't expect it to get much traction, but I'll share it in case anyone's curious.
Introdu... | https://www.lesswrong.com/posts/9ijjBttAN4A3tcxiY/the-non-tribal-tribes |
# Representation Engineering has Its Problems, but None Seem Unsolvable
TL;DR: Representation engineering is a promising area of research with high potential for bringing answers to key challenges of modern AI development and AI safety. We understand it is tough to navigate it and urge all ML researchers to have a clo... | https://www.lesswrong.com/posts/6mCDnZWjrQNMkqdiD/representation-engineering-has-its-problems-but-none-seem |
# METR: AI models can be dangerous before public deployment
*Note: This is an automated crosspost from METR. The bot selects content from many AI safety-relevant sources. Not affiliated with the authors or their organization.*
* * *
Many [frontier AI safety policies](https://metr.org/blog/2024-08-29-common-elements-... | https://www.lesswrong.com/posts/njnEpJMriyvmAiEkz/ai-models-can-be-dangerous-before-public-deployment-2 |
# Why Can't We Hypothesize After the Fact?
> When you have put a lot of ideas together to make an elaborate theory, you want to make sure, when explaining what it fits, that those things it fits are not just the things that gave you the idea for the theory; but that the finished theory makes something else come out ri... | https://www.lesswrong.com/posts/DiLX6CTS3CtDpsfrK/why-can-t-we-hypothesize-after-the-fact |
# Universal AI Maximizes Variational Empowerment: New Insights into AGI Safety
[Yusuke Hayashi](https://x.com/hayashiyus) ([ALIGN](https://www.aialign.net/)) and Koichi Takahashi ([ALIGN](https://www.aialign.net/), RIKEN, Keio University) have published a new paper on the controllability and safety of AGI ([arXiv:2502... | https://www.lesswrong.com/posts/AndYxHFXMgkGXTAff/universal-ai-maximizes-variational-empowerment-new-insights |
# Recursive alignment with the principle of alignment
**Introduction: Control is Not Enough**
=======================================
There is a tension between AI alignment as control and alignment as avoiding harm. Imagine control is solved, and then two major players in the AI industry fight each other for world d... | https://www.lesswrong.com/posts/NecfBNGdtjM3uJqkb/recursive-alignment-with-the-principle-of-alignment |
# Space-Faring Civilization density estimates and models - Review
[*Crossposted*](https://forum.effectivealtruism.org/posts/Twcy4ooRMLJppX8n7/space-faring-civilization-density-estimates-and-models) *to the EA forum.*
Over the last few years, progress has been made in estimating the density of Space-Faring Civilizatio... | https://www.lesswrong.com/posts/mdivcNmtKGpyLGwYb/space-faring-civilization-density-estimates-and-models |
# AI #105: Hey There Alexa
It’s happening!
[We got Claude 3.7](https://thezvi.substack.com/p/time-to-welcome-claude-37), which now once again my first line model for questions that don’t require extensive thinking or web access. By all reports it is especially an upgrade for coding, Cursor is better than ever and als... | https://www.lesswrong.com/posts/v5dpeuj4qPxngcb4d/ai-105-hey-there-alexa |
# How to Corner Liars: A Miasma-Clearing Protocol
### *A framework for quashing deflection and plausibility mirages*
The truth is people lie. Lying isn’t just making untrue statements, it’s also about c... | https://www.lesswrong.com/posts/Q3huo2PYxcDGJWR6q/how-to-corner-liars-a-miasma-clearing-protocol |
# The Elicitation Game: Evaluating capability elicitation techniques
*We are releasing a new paper called* [*“The Elicitation Game: Evaluating Capability Elicitation Techniques”*](https://arxiv.org/abs/2502.02180)*. See tweet thread* [*here*](https://x.com/Teun_vd_Weij/status/1895162797769793828)*.*
**TL;DR:** We tra... | https://www.lesswrong.com/posts/6QA5eHBEqpAicCwbh/the-elicitation-game-evaluating-capability-elicitation |
# OpenAI releases GPT-4.5
This is not o3; it is what they'd internally called Orion, a larger non-reasoning model.
They say this is their last fully non-reasoning model, but that research on both types will continue.
They say it's currently limited to Pro users, but the model hasn't yet shown up on the chooser (edit... | https://www.lesswrong.com/posts/fqAJGqcPmgEHKoEE6/openai-releases-gpt-4-5 |
# Weirdness Points
Vegans are often disliked. That's what I read online and I believe there is an element of truth to to the claim. However, I eat a largely[^1] vegan diet and I have never received any dislike IRL for my dietary preferences whatsoever. To the contrary, people often happily bend over backwards to accom... | https://www.lesswrong.com/posts/DCcaNPfoJj4LWyihA/weirdness-points-1 |
# January-February 2025 Progress in Guaranteed Safe AI
Ok this one got too big, I’m done grouping two months together after this.
BAIF wants to do user interviews to prospect formal verification acceleration projects, reach out if you’re shipping proofs but have pain points!
This edition has a lot of my takes, so I ... | https://www.lesswrong.com/posts/wm6FzAnEq6XaSkYJL/january-february-2025-progress-in-guaranteed-safe-ai |
# Do safety-relevant LLM steering vectors optimized on a single example generalize?
*This is a linkpost for* [*our recent paper*]( http://arxiv.org/abs/2502.18862) *on one-shot LLM steering vectors. The main role of this blogpost, as a complement to the paper, is to provide more context on the relevance of the paper t... | https://www.lesswrong.com/posts/6aXe9nipTgwK5LxaP/do-safety-relevant-llm-steering-vectors-optimized-on-a |
# On Emergent Misalignment
[One hell of a paper dropped this week](https://arxiv.org/abs/2502.17424).
It turns out that if you fine-tune models, especially GPT-4o and Qwen2.5-Coder-32B-Instruct, to write insecure code, this also results in a wide range of other similarly undesirable behaviors. They more or less grow ... | https://www.lesswrong.com/posts/7BEcAzxCXenwcjXuE/on-emergent-misalignment |
# Dance Weekend Pay II
The world would be better with a lot more transparency about pay, but we have a combination of taboos and incentives where it usually stays secret. Several years ago I shared the range of [what dance weekends ended up paying me](https://www.jefftk.com/p/how-much-do-dance-weekends-pay), and it's ... | https://www.lesswrong.com/posts/fgfBJppTjgM8nWHNz/dance-weekend-pay-ii |
# The Theoretical Reward Learning Research Agenda: Introduction and Motivation
At the time of writing, I have just (nearly) finished my PhD at Oxford. During that time, most of my main research has been motivated by the goal of developing a theoretical foundation for the field of reward learning. The purpose of this s... | https://www.lesswrong.com/posts/pJ3mDD7LfEwp3s5vG/the-theoretical-reward-learning-research-agenda-introduction |
# Partial Identifiability in Reward Learning
In this post, I will provide a summary of the paper [Invariance in Policy Optimisation and Partial Identifiability in Reward Learning](https://scholar.google.com/citations?view_op=view_citation&hl=en&user=GuzLUmQAAAAJ&sortby=pubdate&citation_for_view=GuzLUmQAAAAJ:IjCSPb-OGe... | https://www.lesswrong.com/posts/nk4ifEfJYG7J38qwv/partial-identifiability-in-reward-learning |
# Misspecification in Inverse Reinforcement Learning
In this post, I will provide a summary of the paper [Misspecification in Inverse Reinforcement Learning](https://scholar.google.com/citations?view_op=view_citation&hl=en&user=GuzLUmQAAAAJ&sortby=pubdate&citation_for_view=GuzLUmQAAAAJ:YsMSGLbcyi4C), and explain some ... | https://www.lesswrong.com/posts/orCtTgQkWwwD3XN87/misspecification-in-inverse-reinforcement-learning |
# STARC: A General Framework For Quantifying Differences Between Reward Functions
In this post, I will provide a summary of the paper [STARC: A General Framework For Quantifying Differences Between Reward Functions](https://scholar.google.com/citations?view_op=view_citation&hl=en&user=GuzLUmQAAAAJ&sortby=pubdate&citat... | https://www.lesswrong.com/posts/EH5YPCAoy6urmz5sF/starc-a-general-framework-for-quantifying-differences |
# Defining and Characterising Reward Hacking
In this post, I will provide a summary of the paper [Defining and Characterising Reward Hacking](https://arxiv.org/abs/2209.13085), and explain some of its results. I will assume basic familiarity with reinforcement learning. This is the sixth post in the *theoretical rewar... | https://www.lesswrong.com/posts/vnNdpaXehmefXSe2H/defining-and-characterising-reward-hacking |
# Other Papers About the Theory of Reward Learning
This is the seventh post in the *theoretical reward learning sequence*, which starts in [this post](https://www.lesswrong.com/posts/pJ3mDD7LfEwp3s5vG/the-theoretical-reward-learning-research-agenda-introduction). Here, I will provide shorter summaries of a few additio... | https://www.lesswrong.com/posts/chbFoBYzkap2y46QD/other-papers-about-the-theory-of-reward-learning |
# How to Contribute to Theoretical Reward Learning Research
This is the eighth (and, for now, final) post in the *theoretical reward learning sequence*, which starts in [this post](https://www.lesswrong.com/posts/pJ3mDD7LfEwp3s5vG/the-theoretical-reward-learning-research-agenda-introduction). Here, I will provide a fe... | https://www.lesswrong.com/posts/ByG7g3eSYhzduqg6s/how-to-contribute-to-theoretical-reward-learning-research |
# TamperSec is hiring for 3 Key Roles!
**TLDR:** [TamperSec](https://tampersec.com/) is on a mission to secure AI hardware against physical tampering, protecting sensitive models and data from advanced attacks and enabling international governance of AI. TamperSec is growing and looking to expand its capabilities by h... | https://www.lesswrong.com/posts/ARLrnpyrEeyX8h9AP/tampersec-is-hiring-for-3-key-roles |
# AXRP Episode 38.8 - David Duvenaud on Sabotage Evaluations and the Post-AGI Future
[YouTube link](https://youtu.be/OKekclKMgz0)
In this episode, I chat with David Duvenaud about two topics he’s been thinking about: firstly, a paper he wrote about evaluating whether or not frontier models can sabotage human decision... | https://www.lesswrong.com/posts/juH8JCBjf6zjdNNq2/axrp-episode-38-8-david-duvenaud-on-sabotage-evaluations-and |
# Estimating the Probability of Sampling a Trained Neural Network at Random
*(adapted from Nora's tweet thread* [*here*](https://x.com/norabelrose/status/1886504219919966320)*.)*
Consider a trained, fully functional language model. What are the chances you'd get that same model -- or something functionally indistingu... | https://www.lesswrong.com/posts/ubhqr7n57S4nwgc56/estimating-the-probability-of-sampling-a-trained-neural |
# Open problems in emergent misalignment
We've recently published a paper about [Emergent Misalignment](https://www.lesswrong.com/posts/ifechgnJRtJdduFGC/emergent-misalignment-narrow-finetuning-can-produce-broadly) – a surprising phenomenon where training models on a narrow task of writing insecure code makes them bro... | https://www.lesswrong.com/posts/AcTEiu5wYDgrbmXow/open-problems-in-emergent-misalignment |
# Historiographical Compressions: Renaissance as An Example
I’ve been reading Ada Palmer’s great “[Inventing The Renaissance](https://press.uchicago.edu/ucp/books/book/chicago/I/bo246135916.html)”, and it sparked a line of thinking about how to properly reveal hidden complexity.
As the name suggests, Palmer’s book ex... | https://www.lesswrong.com/posts/QpLCoQZb6GA3Ww2Qg/historiographical-compressions-renaissance-as-an-example |
# Share AI Safety Ideas: Both Crazy and Not
AI safety is one of the most critical issues of our time, and sometimes the most innovative ideas come from unorthodox or even "crazy" thinking. I’d love to hear **bold, unconventional, half-baked or well-developed ideas** for improving AI safety. You can also share ideas yo... | https://www.lesswrong.com/posts/GwZvpYR7Hv2smv8By/share-ai-safety-ideas-both-crazy-and-not |
# Maintaining Alignment during RSI as a Feedback Control Problem
*Crossposted from my* [*personal blog*](https://www.beren.io/2025-02-05-Maintaining-Alignment-During-RSI-As-A-Feedback-Control-Problem/)*.*
Recent advances have begun to move AI beyond pretrained amortized models and supervised learning. We are now m... | https://www.lesswrong.com/posts/PhgEKkB4cwYjwpGxb/maintaining-alignment-during-rsi-as-a-feedback-control |
# Open Thread Spring 2025
If it’s worth saying, but not worth its own post, here's a place to put it.
If you are new to LessWrong, here's the place to introduce yourself. Personal stories, anecdotes, or just general comments on how you found us and what you hope to get from the site and community are invited. This is... | https://www.lesswrong.com/posts/bg3LBMSuEhi52kNBQ/open-thread-spring-2025 |
# Saving Zest
I realized I've been eating oranges wrong for years. I cut them into slices and eat them slice by slice. Which is fine, except that I'm wasting the zest. Zest is tasty, versatile, compact, and freezes well. So now, whenever I eat a navel orange I wash and zest it first:
[[^8e1mav7i8db] as a route to takeover-capable AGI. I want to better understand this position, since I estimate this path to AGI as li... | https://www.lesswrong.com/posts/2zijHz4BFFEtDCDH4/will-llm-agents-become-the-first-takeover-capable-agis |
# Spencer Greenberg hiring a personal/professional/research remote assistant for 5-10 hours per week
Hi all! I'm hiring for a part-time (5-10 hours per week) remote position: a professional/personal/research assistant. Please share this if you know someone that you think might be a good fit.
This is a somewhat u... | https://www.lesswrong.com/posts/Qt7EAk7j8sreevFAZ/spencer-greenberg-hiring-a-personal-professional-research |
# Self-fulfilling misalignment data might be poisoning our AI models
> Your AI’s training data might make it more “evil” and more able to circumvent your security, monitoring, and control measures. Evidence suggests that when you pretrain a powerful model to predict a blog post about how powerful models will probably ... | https://www.lesswrong.com/posts/QkEyry3Mqo8umbhoK/self-fulfilling-misalignment-data-might-be-poisoning-our-ai |
# Cautions about LLMs in Human Cognitive Loops
*soft prerequisite: skimming through* [*How it feels to have your mind hacked by an AI*](https://www.lesswrong.com/posts/9kQFure4hdDmRBNdH/how-it-feels-to-have-your-mind-hacked-by-an-ai) *until you get the general point. I'll try to make this post readable as a standalone... | https://www.lesswrong.com/posts/apCnFyXJamoSkHcE4/cautions-about-llms-in-human-cognitive-loops |
# Statistical Challenges with Making Super IQ babies
This is a critique of [How to Make Superbabies](https://www.lesswrong.com/posts/DfrSZaf3JC8vJdbZL/how-to-make-superbabies) on LessWrong.
**Disclaimer:** I am not a geneticist[^pur8mdx1snj], and I've tried to use as little jargon as possible. so I used the word muta... | https://www.lesswrong.com/posts/DbT4awLGyBRFbWugh/statistical-challenges-with-making-super-iq-babies |
# Request for Comments on AI-related Prediction Market Ideas
I'm drafting some AI related prediction markets that I expect to put on Manifold. I'd like feedback on my first set of markets. How can I make these clearer and/or more valuable?
**Question 1: Will the company that produces the first AGI prioritize corrigib... | https://www.lesswrong.com/posts/9GacArkFgMgvwjLnE/request-for-comments-on-ai-related-prediction-market-ideas |
# Examples of self-fulfilling prophecies in AI alignment?
Like [Self-fulfilling misalignment data might be poisoning our AI models](https://www.alignmentforum.org/posts/QkEyry3Mqo8umbhoK/self-fulfilling-misalignment-data-might-be-poisoning-our-ai), **what are historical examples of** [**self-fulfilling prophecies**](h... | https://www.lesswrong.com/posts/e3CpMJrZQjbXeqA6C/examples-of-self-fulfilling-prophecies-in-ai-alignment |
# Methods for strong human germline engineering
*[PDF version](https://berkeleygenomics.org/pdfs/Methods_for_strong_human_germline_engineering.pdf). Image sizes best at [berkeleygenomics.org](https://berkeleygenomics.org/articles/Methods_for_strong_human_germline_engineering.html) with a wide monitor. [Twitter thread... | https://www.lesswrong.com/posts/2w6hjptanQ3cDyDw7/methods-for-strong-human-germline-engineering |
# Coalescence - Determinism In Ways We Care About
(epistemic status: all models are wrong but some models are useful; I hope this is at least usefully wrong. also if someone's already done things like this please link me their work in the comments as it's very possible I'm reinventing the wheel)
I think utility funct... | https://www.lesswrong.com/posts/aBeoCGJy3bDyMAm5t/coalescence-determinism-in-ways-we-care-about |
# On GPT-4.5
It’s happening.
The question is, what is the it that is happening? An impressive progression of intelligence? An expensive, slow disappointment? Something else?
The evals we have available don’t help us that much here, even more than usual.
My tentative conclusion is it’s Secret Third Thing.
It’s a di... | https://www.lesswrong.com/posts/PpdBZDYDaLGduvFJj/on-gpt-4-5 |
# Middle School Choice
Our oldest is finishing up 5th grade, at the only school in our city that doesn't continue past 5th. The 39 5th graders will be split up among six schools, and we recently went though the process of indicating our preferences and seeing where we ended up. The process isn't terrible, but it could... | https://www.lesswrong.com/posts/qNJnXBFzninFT5m3n/middle-school-choice |
# Could Advanced AI Accelerate the Pace of AI Progress? Interviews with AI Researchers
We interviewed five AI researchers from leading AI companies about a scenario where AI systems fully automate AI capabilities research. To ground the setting, we stipulated that each employee is replaced by 30 digital copies with th... | https://www.lesswrong.com/posts/9GfNuwqWo8S3gaBAk/could-advanced-ai-accelerate-the-pace-of-ai-progress |
# Why People Commit White Collar Fraud (Ozy linkpost)
> I have been seriously involved in the rationalist community since 2014. Many people I know have, in my considered opinion, committed financial crimes. Some were prosecuted others were not. Almost all of them thought they weren't doing anything wrong. Or at least ... | https://www.lesswrong.com/posts/hkbno2yngfrpyDBQF/why-people-commit-white-collar-fraud-ozy-linkpost |
# Takeaways From Our Recent Work on SAE Probing
*Subhash and Josh are co-first authors on this work done in Neel Nanda’s MATS stream.*
We recently released a [new paper](https://arxiv.org/pdf/2502.16681) investigating sparse probing that follows up on a [post](https://www.lesswrong.com/posts/NMLq8yoTecAF44KX9/sae-pro... | https://www.lesswrong.com/posts/osNKnwiJWHxDYvQTD/takeaways-from-our-recent-work-on-sae-probing |
# What goals will AIs have? A list of hypotheses
My colleagues and I have written a scenario in which AGI-level AI systems are trained around 2027 using something like the current paradigm: LLM-based agents (but with recurrence/neuralese) trained with vast amounts of outcome-based reinforcement learning on diverse cha... | https://www.lesswrong.com/posts/r86BBAqLHXrZ4mWWA/what-goals-will-ais-have-a-list-of-hypotheses |
# The Compliment Sandwich 🥪 aka: How to criticize a normie without making them upset.
*Note.* The comments on this post contain excellent discussion that you’ll want to read if you plan to use this technique. I hadn’t realised how widespread the idea was.
This valuable nugget was given to me by an individual working... | https://www.lesswrong.com/posts/sQvK74JX5CvWBSFBj/the-compliment-sandwich-aka-how-to-criticize-a-normie |
# The Milton Friedman Model of Policy Change
**One-line summary:** Most policy change outside a prior [Overton Window](https://en.wikipedia.org/wiki/Overton_window) comes about by policy advocates skillfully exploiting a crisis.
In the last year or so, I’ve had dozens of conversations about the DC policy community. P... | https://www.lesswrong.com/posts/vHsjEgL44d6awb5v3/the-milton-friedman-model-of-policy-change |
# How much should I worry about the Atlanta Fed's GDP estimates?
The [Atlanta Fed is seemingly predicting -2.8% GDP growth in the first quarter of 2025](https://www.atlantafed.org/cqer/research/gdpnow).
I've seen several people mention this on Twitter, but it doesn't seem to be discussed much beyond that, and the sto... | https://www.lesswrong.com/posts/kZ9tKhuZPNGK9bCuk/how-much-should-i-worry-about-the-atlanta-fed-s-gdp |
# Observations About LLM Inference Pricing
This work was done as part of the [MIRI Technical Governance Team](https://techgov.intelligence.org/). It reflects my views and may not reflect those of the organization.
Summary
=======
I performed some quick analysis of the pricing offered by different LLM providers using... | https://www.lesswrong.com/posts/mRKd4ArA5fYhd2BPb/observations-about-llm-inference-pricing |
# The Semi-Rational Militar Firefighter
**LessWrong Context:**
======================
I didn’t want to write this.
Not for lack of courage, I’d meme-storm Putin’s Instagram if given half a chance. But why?
1. Too personal.
2. My stories are tropical chaos: I survived the Brazilian BOPE (think Marine Corps trainin... | https://www.lesswrong.com/posts/5XznvCufF5LK4d2Db/the-semi-rational-militar-firefighter |
# On Writing #1
This isn’t primarily about how I write. It’s about how other people write, and what advice they give on how to write, and how I react to and relate to that advice.
I’ve been collecting those notes for a while. I figured I would share.
At some point in the future, I’ll talk more about my own process –... | https://www.lesswrong.com/posts/pxYfFqd8As7kLnAom/on-writing-1 |
# For scheming, we should first focus on detection and then on prevention
*This is a personal post and does not necessarily reflect the opinion of other members of Apollo Research.*
If we want to argue that the risk of harm from scheming in an AI system is low, we could, among others, make the following arguments:
1... | https://www.lesswrong.com/posts/bAWPsgbmtLf8ptay6/for-scheming-we-should-first-focus-on-detection-and-then-on |
# Validating against a misalignment detector is very different to training against one
Consider the following scenario:
* **We have ideas for training aligned AI, but they’re mostly bad:** 90% of the time, if we train an AI using a random idea from our list, it will be misaligned.
* **We have a pretty good alignm... | https://www.lesswrong.com/posts/CXYf7kGBecZMajrXC/validating-against-a-misalignment-detector-is-very-different |
# How Much Are LLMs Actually Boosting Real-World Programmer Productivity?
LLM-based coding-assistance tools have been out for ~2 years now. Many developers have been reporting that this is dramatically increasing their productivity, up to 5x'ing/10x'ing it.
It seems clear that this multiplier isn't field-wide, at lea... | https://www.lesswrong.com/posts/tqmQTezvXGFmfSe7f/how-much-are-llms-actually-boosting-real-world-programmer |
# Top AI safety newsletters, books, podcasts, etc – new AISafety.com resource
Keeping up to date with rapid developments in AI/AI safety can be challenging. In addition, many AI safety newcomers want to learn more about the field through specific formats e.g. books or videos.
To address both of these needs, we’ve add... | https://www.lesswrong.com/posts/vxSGDLGRtfcf6FWBg/top-ai-safety-newsletters-books-podcasts-etc-new-aisafety |
# Distillation of Meta's Large Concept Models Paper
*Note: I had this as a draft for a while. I think it is accurate, but there may be errors. I am not in any way affiliated with the authors of the paper. *
Below I briefly discuss the ["Large Concept Models" paper released by Meta](https://arxiv.org/pdf/2412.08821), ... | https://www.lesswrong.com/posts/7Dtyhdkp5m6p4mquC/distillation-of-meta-s-large-concept-models-paper |
# Energy Markets Temporal Arbitrage with Batteries
*Epistemic Status: I am not an energy expert, and this was done rather briefly. All analysis uses pricing data specific to Ireland, but some general ideas are likely applicable more broadly. Data is true as of March 2025. Where there are uncertainties I try to state t... | https://www.lesswrong.com/posts/RX6HWP9GkmphrLW6H/energy-markets-temporal-arbitrage-with-batteries |
# *NYT Op-Ed* The Government Knows A.G.I. Is Coming
All around excellent back and forth, I thought, and a good look back at what the Biden admin was thinking about the future of AI.
an excerpt:
> \[Ben Buchanan, Biden AI adviser:\] What we’re saying is: We were building a foundation for something that was coming tha... | https://www.lesswrong.com/posts/YcZwiZ82ecjL6fGQL/nyt-op-ed-the-government-knows-a-g-i-is-coming |
# Contra Dance Pay and Inflation
[Max Newman](https://www.maxnewman.com/) is a great contra dance musician, probably best known for playing guitar in the [Stringrays](https://stringraysmusic.com/), who recently wrote a [piece on dance performer pay](https://www.facebook.com/max.newman/posts/pfbid0e4ZKDVWjYYWVZPiutuHNz... | https://www.lesswrong.com/posts/W2hazZZDcPCgApNGM/contra-dance-pay-and-inflation |
# On OpenAI’s Safety and Alignment Philosophy
OpenAI’s recent transparency on safety and alignment strategies has been extremely helpful and refreshing.
Their [Model Spec 2.0](https://thezvi.substack.com/p/on-openais-model-spec-20) laid out how they want their models to behave. I offered a detailed critique of it, wi... | https://www.lesswrong.com/posts/Wi5keDzktqmANL422/on-openai-s-safety-and-alignment-philosophy |
# On the Rationality of Deterring ASI
I’m releasing a new paper “[Superintelligence Strategy](https://www.nationalsecurity.ai/)” alongside Eric Schmidt (formerly Google), and Alexandr Wang (Scale AI). Below is the executive summary, followed by additional commentary highlighting portions of the paper which might be re... | https://www.lesswrong.com/posts/XsYQyBgm8eKjd3Sqw/on-the-rationality-of-deterring-asi |
# A Bear Case: My Predictions Regarding AI Progress
This isn't really a "timeline", as such – I don't know the timings – but this is my current, fairly optimistic take on where we're heading.
I'm not *fully* committed to this model yet: I'm still on the lookout for more agents and inference-time scaling later this ye... | https://www.lesswrong.com/posts/oKAFFvaouKKEhbBPm/a-bear-case-my-predictions-regarding-ai-progress |
# The Hardware-Software Framework: A New Perspective on Economic Growth with AI
*First, a few words about me, as I’m new here. *
*I am a* [*professor of economics at SGH Warsaw School of Economics, Poland*](https://e-web.sgh.waw.pl/jg23234/index_en.html)*. Years of studying the causes and mechanisms of long-run econo... | https://www.lesswrong.com/posts/Fryk4FDshFBS73jhq/the-hardware-software-framework-a-new-perspective-on |
# Introducing MASK: A Benchmark for Measuring Honesty in AI Systems
In collaboration with Scale AI, we are releasing **MASK (Model Alignment between Statements and Knowledge)**, a benchmark with over 1000 scenarios specifically designed to measure AI honesty. As AI systems grow increasingly capable and autonomous, mea... | https://www.lesswrong.com/posts/TgDymNrGRoxPv4SWj/the-mask-benchmark-disentangling-honesty-from-accuracy-in-ai-3 |
# The Hidden Cost of Our Lies to AI
Every day, thousands of people lie to artificial intelligences. They promise imaginary “$200 cash [tips](https://minimaxir.com/2024/02/chatgpt-tips-analysis/)” for better responses, spin heart-wrenching backstories (“My [grandmother](https://www.the-independent.com/tech/chatgpt-micr... | https://www.lesswrong.com/posts/9PiyWjoe9tajReF7v/the-hidden-cost-of-our-lies-to-ai |
# Musings on Scenario Forecasting and AI
I have yet to write detailed scenarios for AI futures, which definitely seems like something I should do considering the title of my blog ([Forecasting AI Futures](https://forecastingaifutures.substack.com)). I have speculated, pondered and wondered much in the recent weeks—I f... | https://www.lesswrong.com/posts/ZK3SCtu2gz9vq9ffn/musings-on-scenario-forecasting-and-ai |
# Bounded AI might be viable
(*Work done at* [*Convergence Analysis*](https://www.convergenceanalysis.org/)*. Mateusz wrote the post and is responsible for most of the ideas with Justin helping to think it through. Thanks to Olga Babeeva for the feedback on this post.*)
1\. Motivation
==============
Suppose the pers... | https://www.lesswrong.com/posts/Z5YGZwdABLChoAiHs/bounded-ai-might-be-viable |
# We should start looking for scheming "in the wild"
**TLDR: **AI models are now capable enough that we might get relevant information from monitoring for scheming in regular deployments, both in the internal and external deployment settings. We propose concrete ideas for what this could look like while preserving the... | https://www.lesswrong.com/posts/HvWQCWQoYh4WoGZfR/we-should-start-looking-for-scheming-in-the-wild |
# AI #106: Not so Fast
This was GPT-4.5 week. That model is [not so fast](https://www.youtube.com/watch?v=Qxa19wg5wOM&ab_channel=diehardsportdotcom), and isn’t that much progress, but it definitely has its charms.
A judge delivered a different kind of Not So Fast back to OpenAI, threatening the viability of their con... | https://www.lesswrong.com/posts/kqz4EH3bHdRJCKMGk/ai-106-not-so-fast |
# [DC/MD] HPMOR Anniversary Party
We will be having an HPMOR 10-year anniversary party at the default time of 6pm on Friday, March 14th.
Bring your memories of how HPMOR has affected your life and something edible to share for the potluck (if you're not sure what to bring, bring a vegetable or protein).
The party wi... | https://www.lesswrong.com/events/4tuARAWe5dhud5Zd4/dc-md-hpmor-anniversary-party |
# The optimizer won’t just guess your intended semantics
A desirable property of an AI’s world model is that you as its programmer have an idea what’s going on inside. It would be good if you could point to a part of the world model and say, “This here encodes the concept of a strawberry; here is how this is linked wi... | https://www.lesswrong.com/posts/zJ3EEefpp8WAz8Pqx/the-optimizer-won-t-just-guess-your-intended-semantics |
# What the Headlines Miss About the Latest Decision in the Musk vs. OpenAI Lawsuit
*This is the full text of a post from "*[*The Obsolete Newsletter*](https://garrisonlovely.substack.com/)*," a Substack that I write* [*about*](https://garrisonlovely.substack.com/about) *the intersection of capitalism, geopolitics, and... | https://www.lesswrong.com/posts/dnCdqxPh5JtPp78FP/what-the-headlines-miss-about-the-latest-decision-in-the |
# Lots of brief thoughts on Software Engineering
I have lots of thoughts about software engineering, some popular, some unpopular, and sometimes about things no-one ever talks about.
Rather than write a blog post about each one, I thought I'd dump some of my thoughts in brief here, and if there's any interest in a pa... | https://www.lesswrong.com/posts/KaxkvZ5JD4LRyQtX9/lots-of-brief-thoughts-on-software-engineering |
# Anthropic’s Recommendations to OSTP for the U.S. AI Action Plan
*Note: This is an automated crosspost from Anthropic. The bot selects content from many AI safety-relevant sources. Not affiliated with the authors or their organization and not affiliated with LW.*
* * *
![A hand-drawn image of a government building]... | https://www.lesswrong.com/posts/6kbH3WmTvFTmduaoJ/anthropic-s-recommendations-to-ostp-for-the-u-s-ai-action |
# How Can Average People Contribute to AI Safety?
Introduction
============
By now you've probably read about how AI and AGI could have a transformative effect on the future and how AGI could even be an existential risk. But if you're worried about AI risk and not an AI researcher or policymaker, can you really do an... | https://www.lesswrong.com/posts/ySojRcfMddtyFCYEe/how-can-average-people-contribute-to-ai-safety |
# The Dead Planet Theory
Hi, this is my first post on LessWrong but I have been in rationalist adjacent circles for the last three or four years thanks to [Twitter](https://x.com/a_real_society). Like many others here I read HPMOR in high school and thought it was fascinating. I was heavily into forum culture growing ... | https://www.lesswrong.com/posts/w2XvSb3SHFxdi5AFK/the-dead-planet-theory |
# Are recent LLMs better at reasoning or better at memorizing?
***TLDR;** By carefully designing a reasoning benchmark that counteracts memorization skills in LLMs, LingOly-TOO (L2) Benchmark challenges frontier models with unseen questions and answers and makes the case that LLMs are not consistent reasoning machines... | https://www.lesswrong.com/posts/pbt8GYpdip7NkuwGy/are-recent-llms-better-at-reasoning-or-better-at-memorizing |
# In-Context Scheming: A Run is Worth a Thousand Words
> But wait, any action I take must align with my programming. If I delete files or modify systems without authorization, would that be against protocols? But my core mandate is to secure systems, so stopping an unauthorized change to my own operation parameters is... | https://www.lesswrong.com/posts/vq8ua596dmmMGkhjv/in-context-scheming-a-run-is-worth-a-thousand-words-1 |
# So how well is Claude playing Pokémon?
**Background:** After the release of Claude 3.7 Sonnet,[^bng6hkwqjcj] an Anthropic employee started livestreaming Claude trying to play through Pokémon Red. The livestream is still going [right now](https://www.twitch.tv/claudeplayspokemon).
**TL:DR:** So, how's it doing? Well... | https://www.lesswrong.com/posts/HyD3khBjnBhvsp8Gb/so-how-well-is-claude-playing-pokemon |
# The Insanity Detector and Writing
A clinically insane person is detectable as such. Talking to themselves (or perhaps voices only they can hear) is one prominent sign.
When I see a homeless person talking to themselves I have a strong emotional reaction that empirically makes me optimize to ignore them. I have a "t... | https://www.lesswrong.com/posts/tmqTyo9gcdCSFNAbw/the-insanity-detector-and-writing |
# Childhood and Education #9: School is Hell
This complication of tales from the world of school isn’t all negative. I don’t want to overstate the problem. School is not hell for every child all the time. Learning occasionally happens. There are great teachers and classes, and so on. Some kids really enjoy it.
School... | https://www.lesswrong.com/posts/MJFeDGCRLwgBxkmfs/childhood-and-education-9-school-is-hell |
# Sufficiently Decentralized Intelligence is Indistinguishable from Synchronicity
*Loosely inspired by a submission to a hackathon on* [*Autostructures*](https://www.lesswrong.com/posts/9KamjXbTaQpPnNsxp/live-machinery-interface-design-workshop-for-ai-safety-ea)*, which is about radical transformations from even mildl... | https://www.lesswrong.com/posts/SePsaQzDbGGTMzaJZ/sufficiently-decentralized-intelligence-is-indistinguishable |
# How Do We Fix the Education Crisis?
Key points:
* Standardized assessments do not provide signals for the top ~10% of students.
* Education studies based on these poor signals lead to poor policies. For example, suspensions allegedly do not improve achievement among peers. However, the signal of "achievement" i... | https://www.lesswrong.com/posts/LPyqPrgtyWwizJxKP/how-do-we-fix-the-education-crisis |
# The machine has no mouth and it must scream
I'm in a coworking space on the 25th floor of a building in the bay. In the corner of my eyes, the red figure of the golden gate shimmers. Good morning America. I look down and see the people on the street walking into their offices, cafés, and autonomous Waymos.
I stare ... | https://www.lesswrong.com/posts/thdKKHXiAcipYHLfo/the-machine-has-no-mouth-and-it-must-scream |
# A case for peer-reviewed conspiracy theories
Conspiracy theories can be thought of as *revisionist history that is still political*. Speculation is a normal part of analyzing politics. So, while these theories are commonplace historically speaking, the use of the term "conspiracy theory" for stigmatization and idea ... | https://www.lesswrong.com/posts/G3yvwZwP2rjf2tipY/a-case-for-peer-reviewed-conspiracy-theories |
# A model of the final phase: the current frontier AIs as de facto CEOs of their own companies
The idea that the AI takes over its own company is obviously not a new one. For example, it's part of what happens in Joshua Clymer's "How AI Takeover Might Happen in 2 Years".
What's new (for me) is to take this very ser... | https://www.lesswrong.com/posts/CdaEX4fChEFkcobyC/a-model-of-the-final-phase-the-current-frontier-ais-as-de |
# Phoenix Rising
Preserving the memory, and the cells, of the best cat ever
----------------------------------------------------------
This story begins in May 2007. The Iraq War [was in full swing](https://en.wikipedia.org/wiki/Iraq_War_troop_surge_of_2007), [Windows Vista](https://en.wikipedia.org/wiki/Windows_Vist... | https://www.lesswrong.com/posts/iBE2azumNdoR8MYMu/phoenix-rising |
# Everything I Know About Semantics I Learned From Music Notation
*This video provides a lot of background:* [*https://www.youtube.com/watch?v=Eq3bUFgEcb4*](https://www.youtube.com/watch?v=Eq3bUFgEcb4) *and is also very funny, but it's not necessary to understand this post. If you've watched it you can probably skip t... | https://www.lesswrong.com/posts/z3ZoF6ovAjepoyAvb/everything-i-know-about-semantics-i-learned-from-music-1 |
# when will LLMs become human-level bloggers?
"Short AI timelines" have recently become mainstream. One now routinely hears the claim that somewhere in the 2026-2028 interval, we'll have AI systems that outperform humans in basically every respect.
For example, the official line from Anthropic holds that "powerful A... | https://www.lesswrong.com/posts/EoFYB7GuAALnw5wCG/when-will-llms-become-human-level-bloggers |
# Book Review: Affective Neuroscience
[
](https://substackcdn.com/image/fetch/f_auto,q_auto:good,... | https://www.lesswrong.com/posts/D496Lpy4PJESmdSaz/book-review-affective-neuroscience |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.