Jason Rute

AI Researcher @ Mistral AI | Formally IBM Research | Former Mathematician/Logician/Data scientist | Building AI for math and reasoning

[Jason Rute @JasonRute]

So excited to introduce Leanstral 1.5, an open source @leanprover code agent model which is SoTA on FATE-H/X, 587 on PutnamBench, and solves all of MiniF2F with only 6B active parameters! It is free to use, so give it a try!

[Albert Jiang @AlbertQJiang]

https://t.co/mWNOMLuWtN
Leanstral 1.5 is here. SoTA on FATE-H/X, 587 on PutnamBench, saturating miniF2F, all with an Apache-2 6B active params model. We are having fun verifying code properties and catching bugs in Rust repos!

Tech report covering training environment and evaluations: https://t.co/4zVUZfArn3
We also open-source LeanstralSafeVerify and FLTEval.

[Jason Rute @JasonRute]

@octonion Here is a list of at least some of them, but I’d love to find a more comprehensive list of these super broad conjectures: https://t.co/A5jYHkZ5sd

[Jason Rute @JasonRute]

@ereliuer_eteer Going to do a set of Wang tiles next? Or a set of group generators?

[Jason Rute @JasonRute]

@MoritzFirsching @Almost_Sure Actually I think MLC might imply both hyperbolicity and zero boundary area as well, so that alone would resolve it.

[Jason Rute @JasonRute]

@MoritzFirsching @Almost_Sure If you do, also add the hyperbolicity conjecture for the Mandelbrot set and that the boundary has zero area. Those two combined would resolve my conjecture positively.

[Jason Rute @JasonRute]

@ElliotGlazer $20k vs 1 dollar. I’d take those odds. You can even hold onto my dollar! 😈

[Jason Rute @JasonRute]

@LawrPaulson Interesting, because a lot has been made of the errors in the paper and how they don’t match the Lean.

[Jason Rute @JasonRute]

@ChrSzegedy @ColmezPierre Yes, LANA didn’t get very far with the formalization (and was overly PC). A full completion of the LANA formalization with just a small remaining gap left for Mochizuki to fill in would be interesting and hopefully valuable to the anabelian geometry community.

[Jason Rute @JasonRute]

@tonylfeng @aaswaminathan01 How challenging is it to *state* the GLC in current Lean. Maybe GLC should be added as a challenge to LeanEval. Then @IlinVasily29521 will be compelled to fill in the proof.

[Jason Rute @JasonRute]

@deancureton @schweingehtbree @geoffreyirving It would even better if this proof was in Tau Ceti (alongside the CFSG statement). It would act as a good unit test on the statement which is robust to library refactors.

[Jason Rute @JasonRute]

@ElliotGlazer When we formalize 90% of interesting published math, it will be interesting to see what gets left behind and for what reasons.

[Jason Rute @JasonRute]

@ElliotGlazer Lean has 4CT now. Computation might not be a problem.

[Jason Rute @JasonRute]

@t0tientqu0tient @CarinaLHong To be fair, @ElliotGlazer’s prediction was for constructive proofs of those theorems. We are not there yet. 😀

[Jason Rute @JasonRute]

@AndrewCritchPhD I think @ElliotGlazer would be interested in wagering you on some of those using your odds?

[Jason Rute @JasonRute]

@ereliuer_eteer @nihilunbounded The conjecture as stated is for ZFC (follow links).

[Jason Rute @JasonRute]

@ShriramKMurthi @PommeTomateRiz @rs_datascience I don’t mean “check” in the sense of 100% confidence, but I mean greatly improve the confidence that the spec is right and useful toward the goal. Humans can review other’s specs so why can’t machines? You can even use proofs to improve trust in the spec. I’ll DM you tomorrow.

[Jason Rute @JasonRute]

@PommeTomateRiz @rs_datascience @ShriramKMurthi Ok, I was confused because you responded to my post responding to the claim that Lean can prove false. (And because when people talk about “verifying the verifiers”, they often are taking about proving false as well.) Anyway, thanks for clearing that up.