Today, we're releasing Proving the Coding Interview on arxiv and huggingface. We believe FVAPPS is currently the largest formal verification benchmark, consisting of leetcode-style problems in @leanprover.

RT @AISecurityInst: White box methods can be a useful complement to black box monitoring, especially if:. ❌Chains of thought do not reflect….

one of the coolest things my math prof did was print out wikipedia pages and make sure we knew how to read them. His justification was like "learn this language and you can join the community".

i love it when . me: i'm worried this initiative is a bad idea.coworkers: thanks for your interest in this initiative! i'll loop you into discussions going forward. it's like peak "complaining is volunteering to fix it". seems like a valuable thing in an organization culture.

one project i'd like to do if i was funemployed and not worried about AI takeover is an agentdev library in Lean or Haskell. I already have a bunch of ideas from an agent codebase I did in fp-ts a few months ago.

This happened to me last week metaprogramming in lean.

but LLMs give us the opportunity to study this with actual hard data!.

envisioning a large scale study or eval that gets really good data on verification burdens from the LLM's perspective. There's been a TON of folklore about the verification burden in humans, where we have massive confidence intervals about what verification does to labor cost.

In my Iliad/Odyssey submission, my metric _verification burden_ measures the ratio of how many tokens it costs to prove a completion correct against how many tokens it cost to produce the completion. I submitted a couple weeks ago, but I'm just now.

doing it's part to solve fertility collapse

RT @qd_forall: @NathanpmYoung @DKokotajlo So idk when people say "thats all vibes" or "so n so was wrong about the events", in my brain im….

RT @qd_forall: @NathanpmYoung @DKokotajlo I literally spent intimate grind moments with the quantification of surprise for several hours, a….

@NathanpmYoung @DKokotajlo Im obviously not doing the calculation all the time, but I can kinda see and feel shapes of certain sizes in my head when I picture accuracy, calibration, information contribution, etc.

@NathanpmYoung @DKokotajlo So idk when people say "thats all vibes" or "so n so was wrong about the events", in my brain im thinking about a distribution's accurate information contribution once you plug in the ground truth, reminiscing about the formalism and numerical properties of KL divergence.

@NathanpmYoung @DKokotajlo I literally spent intimate grind moments with the quantification of surprise for several hours, and I rely on that inside view when I reason about how many bayes points to award various hot takes as the world changes. I really dont think id be able to do that if I hadnt!!.

@NathanpmYoung @DKokotajlo And im kinda worried that it's a niche, sophisticated skill? I only have it cuz i wrote out the integrals for the KLDivergence numerical tests in the squiggle codebase back when I worked there

@NathanpmYoung @DKokotajlo What im interested in is what mental tools are required for members of the audience to say "the point estimate was off by a bounded amount, but it was still reasonably somewhere in the distribution, therefore was valuable info/estimate in expectation".

@NathanpmYoung @DKokotajlo I noticed watching the video that we expect the essay to be wrong in specifics like the success or failure of an exfiltration attack, etc. Daniel himself has said 2028 makes more sense, since the thing was published.

Some people say AI2027 is all vibes and not research. There's some sense in which this is true as @NathanpmYoung says "forecasting is just vibes plus track record", but it's still a misleading statement, because @DKokotajlo has that track record and cuz rigor in the estimates.

Exciting web content! I have a random AI2027/forecasting thought, thread:.

What manner of beast