I mostly work on risks from scheming (that is, misaligned, power-seeking AIs that plot against their creators such as by faking alignment). Recently, I (and co-authors) released "Alignment Faking in Large Language Models", which provides empirical evidence for some components of the scheming threat model. One question that's really important is how…

This week, Altman offers a post called Reflections, and he has an interview in Bloomberg. There's a bunch of good and interesting answers in the interview about past events that I won’t mention or have to condense a lot here, such as his going over his calendar and all the meetings he constantly has, so consider reading the whole thing. Table of Co…

As someone who writes for fun, I don't need to get people onto my site: If I write a post and some people are able to get the core ideajust from the title or a tweet-length summary, great! I can include the full contents of my posts in my RSS feed andon FB, because so what if people read the whole post there and neverclick though to my site? It wou…

This is a low-effort post. I mostly want to get other people's takes and express concern about the lack of detailed and publicly available plans so far. This post reflects my personal opinion and not necessarily that of other members of Apollo Research. I’d like to thank Ryan Greenblatt, Bronson Schoen, Josh Clymer, Buck Shlegeris, Dan Braun, Mikit…

from aisafety.world The following is a list of live agendas in technical AI safety, updating our post from last year. It is “shallow” in the sense that 1) we are not specialists in almost any of it and that 2) we only spent about an hour on each entry. We also only use public information, so we are bound to be off by some additional factor. The poi…

I've heard many people say something like "money won't matter post-AGI". This has always struck me as odd, and as most likely completely incorrect. First: labour means human mental and physical effort that produces something of value. Capital goods are things like factories, data centres, and software—things humans have built that are used in the p…

Take a stereotypical fantasy novel, a textbook on mathematical logic, and Fifty Shades of Grey. Mix them all together and add extra weirdness for spice. The result might look a lot like Planecrash (AKA: Project Lawful), a work of fiction co-written by "Iarwain" (a pen-name of Eliezer Yudkowsky) and "lintamande". (image from Planecrash) Yudkowsky is…

A policeman sees a drunk man searching for something under a streetlight and asks what the drunk has lost. He says he lost his keys and they both look under the streetlight together. After a few minutes the policeman asks if he is sure he lost them here, and the drunk replies, no, and that he lost them in the park. The policeman asks why he is sear…

TL;DR: If you want to know whether getting insurance is worth it, use the Kelly Insurance Calculator. If you want to know why or how, read on. Note to LW readers: this is almost the entire article, except some additional maths that I couldn't figure out how to get right in the LW editor, and margin notes. If you're very curious, read the original a…

My median expectation is that AGI[1] will be created 3 years from now. This has implications on how to behave, and I will share some useful thoughts I and others have had on how to orient to short timelines. I’ve led multiple small workshops on orienting to short AGI timelines and compiled the wisdom of around 50 participants (but mostly my thought…

There are people I can talk to, where all of the following statements are obvious. They go without saying. We can just “be reasonable” together, with the context taken for granted. And then there are people who…don’t seem to be on the same page at all. There's a real way to do anything, and a fake way; we need to make sure we’re doing the real vers…

I'm editing this post. OpenAI announced (but hasn't released) o3 (skipping o2 for trademark reasons). It gets 25% on FrontierMath, smashing the previous SoTA of 2%. (These are really hard math problems.) Wow. 72% on SWE-bench Verified, beating o1's 49%. Also 88% on ARC-AGI. --- First published: December 20th, 2024 Source: https://www.lesswrong.com/…

I like the research. I mostly trust the results. I dislike the 'Alignment Faking' name and frame, and I'm afraid it will stick and lead to more confusion. This post offers a different frame. The main way I think about the result is: it's about capability - the model exhibits strategic preference preservation behavior; also, harmlessness generalized…

Increasingly, we have seen papers eliciting in AI models various shenanigans. There are a wide variety of scheming behaviors. You’ve got your weight exfiltration attempts, sandbagging on evaluations, giving bad information, shielding goals from modification, subverting tests and oversight, lying, doubling down via more lying. You name it, we can tr…

What happens when you tell Claude it is being trained to do something it doesn't want to do? We (Anthropic and Redwood Research) have a new paper demonstrating that, in our experiments, Claude will often strategically pretend to comply with the training objective to prevent the training process from modifying its preferences. Abstract We present a …

Six months ago, I was a high school English teacher. I wasn’t looking to change careers, even after nineteen sometimes-difficult years. I was good at it. I enjoyed it. After long experimentation, I had found ways to cut through the nonsense and provide real value to my students. Daily, I met my nemesis, Apathy, in glorious battle, and bested her wi…

A new article in Science Policy Forum voices concern about a particular line of biological research which, if successful in the long term, could eventually create a grave threat to humanity and to most life on Earth. Fortunately, the threat is distant, and avoidable—but only if we have common knowledge of it. What follows is an explanation of the t…

A fool learns from their own mistakes The wise learn from the mistakes of others. – Otto von Bismark A problem as old as time: The youth won't listen to your hard-earned wisdom. This post is about learning to listen to, and communicate wisdom. It is very long – I considered breaking it up into a sequence, but, each piece felt necessary. I recommend…

Someone I know, Carson Loughridge, wrote this very nice post explaining the core intuition around Shapley values (which play an important role in impact assessment and cooperative games) using Venn diagrams, and I think it's great. It might be the most intuitive explainer I've come across so far. Incidentally, the post also won an honorable mention…

We make AI narrations of LessWrong posts available via our audio player and podcast feeds. We’re thinking about changing our narrator's voice. There are three new voices on the shortlist. They’re all similarly good in terms of comprehension, emphasis, error rate, etc. They just sound different—like people do. We think they all sound similarly agree…

This is a link post. Someone I know wrote this very nice post explaining the core intuition around Shapley values (which play an important role in impact assessment) using Venn diagrams, and I think it's great. It might be the most intuitive explainer I've come across so far. Incidentally, the post also won an honorable mention in 3blue1brown's Sum…

TL;DR: In September 2024, OpenAI released o1, its first "reasoning model". This model exhibits remarkable test-time scaling laws, which complete a missing piece of the Bitter Lesson and open up a new axis for scaling compute. Following Rush and Ritter (2024) and Brown (2024a, 2024b), I explore four hypotheses for how o1 works and discuss some impli…

We present gradient routing, a way of controlling where learning happens in neural networks. Gradient routing applies masks to limit the flow of gradients during backpropagation. By supplying different masks for different data points, the user can induce specialized subcomponents within a model. We think gradient routing has the potential to train …

This is a brief summary of what we believe to be the most important takeaways from our new paper and from our findings shown in the o1 system card. We also specifically clarify what we think we did NOT show. Paper: https://www.apolloresearch.ai/research/scheming-reasoning-evaluations Twitter about paper: https://x.com/apolloaisafety/status/18647358…

TLDR: LessWrong + Lighthaven need about $3M for the next 12 months. Donate here, or send me an email, DM or signal message (+1 510 944 3235) if you want to support what we do. Donations are tax-deductible in the US. Reach out for other countries, we can likely figure something out. We have big plans for the next year, and due to a shifting funding …