I figured my math followers would especially appreciate this! Just got these, they are originals from 1740 and 1751. Include the original publication of Euler's solution to the Basel problem, among other very cool things. So crazy to have these pieces of history on my bookshelf.
imo this is confused thinking; you can automate safety research while being able to verify the results and therefore not having handed off human oversight. in fact, our powers of formal verification will only grow the better we understand these systems
@GiulioRebuffo@Citrullin Yes, but the subtlety in "super close" is very practically useful! In Lean/Rocq, you can prove consistency of n universes with n + 1, metatheory properties, or relative consistency with extensions of ZFC. I'm convinced that these will become table stakes for new proof assistants.
@Citrullin@GiulioRebuffo It's usually not worth it, but there is work that includes modeling physical failures (power loss, etc) into their verification. As with all software verification, it's a matter of how much detail you want to model, and how accurate this model is.
@Jack______M@scottnarmstrong I think I agree. One characterization I've liked from OAI is calling these AI-generated papers "raw research data". They should be immediately available in a release, with the recognition they don't fill all roles of traditional publication (trust, communication, etc.)
Very sad to see comments like this. Work on the human-curated formalization is certainly not "done". Besides the fact that the proof they're formalizing is a modern version, there is certainly value to investing time in a higher quality of formalization than LLMs can provide.
ICYMI, this part of the Fable 5.2 release has its own report: https://t.co/mp5upicgiX
This a a second full formalization of FLT, taking fewer tokens, a much more compact formalization, and 1000 agents figuring out on their own what to do without a tool like Prove2me.
A team of 1,000 Fable 5.2 agents fully formalized Fermat’s Last Theorem in 30 hours, translating the proof into a language a computer can check step by step.
With the same setup, Opus 5.5 completed only 26% in 18 hours, and barely made any more progress after 30 hours.
Formal methods didn't start with us.
@ClarkBarrett7 on the people and tools &Truth builds on, from Turing, Hoare, and Dijkstra to Lean, Separation Logic, and SMT solvers:
https://t.co/5GFDoYj6IA
@Enrohk3j2@GoogleDeepMind@axiommathai There is never an absolute guarantee, with or without AI. Inherently it's a nontrivial human judgement of a meaningful alignment with informal mathematics. The value is that with enough careful review and theorems proven that match the literature, we gain greater certainty.
Autoformalization is now mature and widespread enough that we need to think closer about the different reasons we formalize and the different degrees of informal/formal proofs being aligned. (Short thread)
Reminder. Formalization is void if you cannot prove that the lean certificate matches exactly the proof.
I suspect some people with bugs in their papers are going to produce lean certificates of the main result that avoids the holes that have been filled bu the clankers.
Deciding how to help non-expert humans engage with formal verification specifications seems like one of the more important non-verifiable UX problems facing the rise of formal verification. 🧵
@rel_zeta_tech There are many projects that revolve around human-curated statements, with people thinking hard about how they align with the literature. Those doing high quality autoformalization are at minimum carefully reviewing definitions. (Anything else is extremely irresponsible IMO)
@GoogleDeepMind@axiommathai To me, this is the argument for new formalizations, if they aim for anything past verifying correctness, to have a human-curated core isolated from agent development. Beyond correct specifications, you should retain control of the engineering choices of your formalization.
@GoogleDeepMind@axiommathai In my view, current capabilities are simply not here yet. I think it's admirable that for instance the Stacks formalization (also by @axiommathai) put thought into both aligning semantically and with good engineering, but ultimately this remains removed from Mathlib quality.