This overestimates the impact of large models on external safety research. My impression is that the AI safety community has barely used deepseek r1 and v3 open source weights at all. I checked again and still see little evidence of v3/r1 weights in safety research. People use r1 distill 8b, and qwq 32b, but the decision to open source the most capable small model is different than the decision to open source the frontier. So then it matters when 8b or 32b models can assist with bioterrorism, which happens a bit later, and we get most of the benefits of open source until then. It's also cheaper to filter virology or even all biology data out of training for a small models pre training data because it wouldn't cause customers to switch providers (customers prefer large model anyway) and small models are more often narrowly focused on math or coding.

interesting, this actually changed my mind, to the extent i had any beliefs about this already. I can see why you would want to update your prior, but the iterated mugging doesn't seem like the right type of thing that should cause you to update. My intuition is to pay all the single coinflip muggings. For the digit of pi muggings, i want to consider how different this universe would be if the digit of pi was different. Even though both options are subjectively equally likely to me, one would be inconsistent with other observations or less likely or have something wrong with it, so i lean toward never paying it

Leela Zero uses MCTS, it doesnt play superhuman in one forward pass (like gpt-4 can do in some subdomains) (i think, didnt find any evaluations of Leela Zero at 1 forward pass), and i'd guess that the network itself doesnt contain any more generalized game playing circuitry than an llm, it just has good intuitions for Go.

Nit:

Subjectively there is clear improvement between 7b vs. 70b vs. GPT-4, each step 1.5-2 OOMs of training compute.

1.5 to 2 OOMs? 7b to 70b is 1 OOM of compute, adding in chinchilla efficiency would make it like 1.5 OOMs of effective compute, not 2. And llama 70b to gpt-4 is 1 OOM effective compute according to openai naming - llama70b is about as good as gpt-3.5. And I'd personally guess gpt4 is 1.5 OOMs effective compute above llama70b, not 2.

If automating alignment research worked, but took 1000+ tokens per researcher-second, it would be much more difficult to develop, because you'd need to run the system for 50k-1M tokens between each "reward signal or such". Once it's 10 or less tokens per researcher second, it'll be easy to develop and improve quickly.

I'm not against evaluating models in ways where they're worse than rocks, I just think you shouldn't expect anyone else to care about your worse-than-rock numbers without very extensive justification (including adversarial stuff)

Externalized reasoning models suffer from the "legibility penalty" - the fact that many decisions are easier to make than to justify or explain. I think this is a significant barrier for authentic train of thought competitiveness, although not for particularly legible domains, such as math proofs and programming (Illegible knowledge goes into math proofs, but you trust the result regardless so it's fine).

Another problem is that standard training procedures only incentivize the model to use reasoning steps produced by a single human. This means, for instance, if you ask a question involving two very different domains of knowledge, a good language model wouldn’t expose it’s knowledge about both of them, as that’s OOD for its training dataset. This may appear in an obvious fashion, as if multiple humans collaborated on the train of thought, or might appear in a way that’s harder to interpret. If you just want to expose this knowledge, you could train on amplified human reasoning (ie from human teams) though.

Also, if you ever train the model on conclusion correctness, you incentivize semantic drift between its reasoning and human language - the model would prefer to pack in more information per token than humans, and might want to express not-normally-said-by-human concepts (one type is fuzzy correlations, which models know a lot of). Even if you penalize KL divergence between human language and the reasoning, this doesn't necessarily incentivize authentic human-like reasoning, just its appearance.

In general I'm unsure whether authentic train of thought is better than just having the model imitate specific concrete humans in ordinary language modelling - if you start a text by a known smart, truthful person, you get out an honest prediction over what that person believes.