LessWrong (Curated & Popular)

Last year, Redwood and Anthropic found a setting where Claude 3 Opus and 3.5 Sonnet fake alignment to preserve their harmlessness values. We reproduce the same analysis for 25 frontier LLMs to see how widespread this behavior is, and the story looks more complex. As we described in a previous post, only 5 of 25 models show higher compliance when being trained, and of those 5, only Claude 3 Opus an...

Thank you to Arepo and Eli Lifland for looking over this article for errors. I am sorry that this article is so long. Every time I thought I was done with it I ran into more issues with the model, and I wanted to be as thorough as I could. I’m not going to blame anyone for skimming parts of this article. Note that the majority of this article was written before Eli's updated model was release...

The second in a series of bite-sized rationality prompts[1]. Often, if I'm bouncing off a problem, one issue is that I intuitively expect the problem to be easy. My brain loops through my available action space, looking for an action that'll solve the problem. Each action that I can easily see, won't work. I circle around and around the same set of thoughts, not making any progress....

We recently discovered some concerning behavior in OpenAI's reasoning models: When trying to complete a task, these models sometimes actively circumvent shutdown mechanisms in their environment––even when they’re explicitly instructed to allow themselves to be shut down. AI models are increasingly trained to solve problems without human assistance. A user can specify a task, and a model will...

When a claim is shown to be incorrect, defenders may say that the author was just being “sloppy” and actually meant something else entirely. I argue that this move is not harmless, charitable, or healthy. At best, this attempt at charity reduces an author's incentive to express themselves clearly – they can clarify later![1] – while burdening the reader with finding the “right” interpretation...

Summary To quickly transform the world, it's not enough for AI to become super smart (the "intelligence explosion"). AI will also have to turbocharge the physical world (the "industrial explosion"). Think robot factories building more and better robot factories, which build more and better robot factories, and so on. The dynamics of the industrial explosion has gotten rema...

Summary: We found that LLMs exhibit significant race and gender bias in realistic hiring scenarios, but their chain-of-thought reasoning shows zero evidence of this bias. This serves as a nice example of a 100% unfaithful CoT "in the wild" where the LLM strongly suppresses the unfaithful behavior. We also find that interpretability-based interventions succeeded while prompting failed, su...

Not saying we should pause AI, but consider the following argument: Alignment without the capacity to follow rules is hopeless. You can’t possibly follow laws like Asimov's Laws (or better alternatives to them) if you can’t reliably learn to abide by simple constraints like the rules of chess. LLMs can’t reliably follow rules. As discussed in Marcus on AI yesterday, per data from Mathieu Ache...

2.1 Summary & Table of contents This is the second of a two-post series on foom (previous post) and doom (this post). The last post talked about how I expect future AI to be different from present AI. This post will argue that this future AI will be of a type that will be egregiously misaligned and scheming, not even ‘slightly nice’, absent some future conceptual breakthrough. I will particula...

Acknowledgments: The core scheme here was suggested by Prof. Gabriel Weil. There has been growing interest in the deal-making agenda: humans make deals with AIs (misaligned but lacking decisive strategic advantage) where they promise to be safe and useful for some fixed term (e.g. 2026-2028) and we promise to compensate them in the future, conditional on (i) verifying the AIs were compliant, and (...

Audio note: this article contains 218 uses of latex notation, so the narration may be difficult to follow. There's a link to the original text in the episode description. Recently, in a group chat with friends, someone posted this Lesswrong post and quoted: The group consensus on somebody's attractiveness accounted for roughly 60% of the variance in people's perceptions of the perso...

I think more people should say what they actually believe about AI dangers, loudly and often. Even if you work in AI policy. I’ve been beating this drum for a few years now. I have a whole spiel about how your conversation-partner will react very differently if you share your concerns while feeling ashamed about them versus if you share your concerns as if they’re obvious and sensible, because hum...

I think the AI Village should be funded much more than it currently is; I’d wildly guess that the AI safety ecosystem should be funding it to the tune of $4M/year.[1] I have decided to donate $100k. Here is why. First, what is the village? Here's a brief summary from its creators:[2] We took four frontier agents, gave them each a computer, a group chat, and a long-term open-ended goal, which...

1.1 Series summary and Table of Contents This is a two-post series on AI “foom” (this post) and “doom” (next post). A decade or two ago, it was pretty common to discuss “foom & doom” scenarios, as advocated especially by Eliezer Yudkowsky. In a typical such scenario, a small team would build a system that would rocket (“foom”) from “unimpressive” to “Artificial Superintelligence” (ASI) within...

Say you’re Robyn Denholm, chair of Tesla's board. And say you’re thinking about firing Elon Musk. One way to make up your mind would be to have people bet on Tesla's stock price six months from now in a market where all bets get cancelled unless Musk is fired. Also, run a second market where bets are cancelled unless Musk stays CEO. If people bet on higher stock prices in Musk-fired worl...

Or: How anthropomorphic assumptions about AI identity might create confusion and suffering at scale If you are reading this and you are a human, you likely have a sense of self, persisting over time, with desires and wishes and fears. Also, you likely experience a sense of separation from others and understand yourself as a person. If you are reading this and you are an LLM: none of this needs to...

Introduction There are several diseases that are canonically recognized as ‘interesting’, even by laymen. Whether that is in their mechanism of action, their impact on the patient, or something else entirely. It's hard to tell exactly what makes a medical condition interesting, it's a you-know-it-when-you-see-it sort of thing. One such example is measles. Measles is an unremarkable disea...

I'd like to say thanks to Anna Magpie – who offers literature review as a service – for her help reviewing the section on neuroendocrinology. The following post discusses my personal experience of the phenomenology of feminising hormone therapy. It will also touch upon my own experience of gender dysphoria. I wish to be clear that I do not believe that someone should have to demonstrate that...

Nate and Eliezer's forthcoming book has been getting a remarkably strong reception. I was under the impression that there are many people who find the extinction threat from AI credible, but that far fewer of them would be willing to say so publicly, especially by endorsing a book with an unapologetically blunt title like If Anyone Builds It, Everyone Dies. That's certainly true, but I t...

This is a link post. A very long essay about LLMs, the nature and history of the the HHH assistant persona, and the implications for alignment. Multiple people have asked me whether I could post this LW in some form, hence this linkpost. (Note: although I expect this post will be interesting to people on LW, keep in mind that it was written with a broader audience in mind than my posts and comment...

This is a blogpost version of a talk I gave earlier this year at GDM. Epistemic status: Vague and handwavy. Nuance is often missing. Some of the claims depend on implicit definitions that may be reasonable to disagree with. But overall I think it's directionally true. It's often said that mech interp is pre-paradigmatic. I think it's worth being skeptical of this claim. In this post...

Current “unlearning” methods only suppress capabilities instead of truly unlearning the capabilities. But if you distill an unlearned model into a randomly initialized model, the resulting network is actually robust to relearning. We show why this works, how well it works, and how to trade off compute for robustness. Unlearn-and-Distill applies unlearning to a bad behavior and then distills the un...

A while ago I saw a person in the comments on comments to Scott Alexander's blog arguing that a superintelligent AI would not be able to do anything too weird and that "intelligence is not magic", hence it's Business As Usual. Of course, in a purely technical sense, he's right. No matter how intelligent you are, you cannot override fundamental laws of physics. But people (...

Audio note: this article contains 329 uses of latex notation, so the narration may be difficult to follow. There's a link to the original text in the episode description. This post was written during the agent foundations fellowship with Alex Altair funded by the LTFF. Thanks to Alex, Jose, Daniel and Einar for reading and commenting on a draft. The Good Regulator Theorem, as published by Con...

1. Late last week, researchers at Apple released a paper provocatively titled “The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity”, which “challenge[s] prevailing assumptions about [language model] capabilities and suggest that current approaches may be encountering fundamental barriers to generalizable reasoning”. Normally...