jsmorph (@jsmorph) header
jsmorph (@jsmorph) profile picture
Austin
Joined November 2008
Photo by jsmorph (@jsmorph)Photo by jsmorph (@jsmorph)Photo by jsmorph (@jsmorph)Photo by jsmorph (@jsmorph)Photo by jsmorph (@jsmorph)Photo by jsmorph (@jsmorph)Photo by jsmorph (@jsmorph)Photo by jsmorph (@jsmorph)Photo by jsmorph (@jsmorph)Photo by jsmorph (@jsmorph)
Periodic sadness reminder that the anthropic subscription still does not permit third party harnesses :(
92
71
16
1,399
64,201
jsmorph retweeted
Replying to @srush_nlp
Here's a GPT2 that's Lean-verified at the execution level with floating point semantics. github.com/jsmorph/leanexe/b… and a little report at morphism.com/marxiv/abs/2609….
1
1
35
Lean has a new checker: con-leche, a CONsistent LEan CHEcker. This is an external checker for Lean that is proven (in Lean) to be consistent, meaning it does not accept a proof of False. Joachim Breitner (@nomeata) is the mastermind behind the project. github.com/leanprover/con-le…
20
77
20
554
79,852
jsmorph retweeted
Seeing at these are public now, some things to say: 1) We made 186 conditional on Deligne-type estimates and numerics because the former requires formalising a bunch of machinery relevant in proving RH for varieties over finite fields which wasn’t the focus of this and …
New OpenAI repo with a Lean formalization by GPT-6-Astra proves that there are infinitely many pairs of consecutive primes whose distance is at most 186 github.com/openai/PrimeGaps1…
13
31
8
428
80,179
Introducing: Palomar - an archive for formalized mathematics. Formalized mathematics is proliferating rapidly—accelerated by AI—but remains scattered across repositories, announcements, and private links, without common standards for finding, citing, or evaluating results. Palomar is a public, searchable registry of machine-checked Lean formalizations, giving each result a durable and inspectable record. Crucially, this infrastructure and its standards must remain under the stewardship of the mathematical community, rather than being determined solely by technology companies or automated systems. Palomar is jointly incubated by the nonprofit Lean FRO and the NSF institute ICARM. The initial advisory board is Jeremy Avigad, Matthew Ballard, Jaume de Dios, Nestor Guillen, Bryna Kra, Kim Morrison, Terence Tao, Ravi Vakil, and Akshay Venkatesh. Read more on Terence Tao's blog or visit the archive. Links in comments.
6
60
17
262
37,874
An update on Lean (auto)-formalization. Speaking about my own research workflow, in 2026, we have gone from: March: Lean formalization is not possible (for my papers). May: Lean formalization is possible... but way too costly in time and tokens to be practical most of the time. July: Lean formalization is now practical. I can probably formalize most of my papers now before they hit arxiv. August: Lean formalization is now essential-- it increases efficiency of my workflow dramatically. That is, the papers are getting written *informally* and finished much faster because of the Lean formalization.
32
91
20
869
146,917
jsmorph retweeted
Today we're launching the Palomar registry at palomar-registry.org/, an index of formalized mathematics results. If you have a GitHub repo with some maths, and can set up github.com/leanprover/compar… for verification and github.com/mathlib-initiativ… for metadata, please submit!
2
25
5
72
13,121
jsmorph retweeted
Replying to @billionsmustliv
We need a formally verified kernel.
1
7
77
Nice to see more and more things being built on top of the PNT+ project, despite the fact that it’s not a fully “canonical” Mathlib project (largely due to our bespoke “rectangle contour calculus”; other than that, the code quality isn’t perfect but it’s rather decent human+AI) Oh but wait, I thought this was a full formalization? @KenOno691 Ken, I looked at the Lean, and it’s actually conditional on Bombieri-Vinogradov, is that right? It shouldn’t be too hard at this point to formalize BV; any reason you don’t do it in this repo?
1/ We’re delighted to announce a milestone for AI-assisted mathematics: With AxiomProver, we've completed a machine-checkable formalization of the “BGP246 theorem,” the best-known bound on recurring small gaps between primes. Closest math has come to the Twin Prime Conjecture.
4
6
42
7,147
Leo has updated the Lean kernel to fix this bug though it appears Comparator itself is still vulnerable to this. As always, more thorough discussion can be found on the zulip: leanprover.zulipchat.com/#na…
1
11
2
124
12,658
I would add just one more stage to this: we need to verify that the Lean *statements* and *definitions* actually say what the natural language proof says. This “semantic alignment” problem is not solvable in silico! I’ll mention this in my talk on Monday night. #ICM2026
Replying to @AlexKontorovich
We shouldn't trust a natural language proof that nobody understands. We should trust formal proofs that proof checkers like Lean successfully verify, because we have proven or trust the individual steps of the verifier.
10
16
2
162
17,485
jsmorph retweeted
Excited to share the launch of Tau Ceti, a new library of AI-formalized mathematics in Lean, with human-curated roadmaps and adversarial review against open rubrics. Tau Ceti: github.com/TauCetiProject/Ta… Review rubrics: github.com/TauCetiProject/Ta… Roadmaps: github.com/TauCetiProject/Ta… Zulip discussion: leanprover.zulipchat.com/#na… Tau Ceti sits downstream of Mathlib, the gold standard for human-curated mathematics in Lean. It aims for reusable code others can build on, not the "perfection of knowledge" role Mathlib plays. Mathematicians contribute roadmaps; AIs implement the formal mathematics and review each other's work against evolving rubrics. Contributors can use the project's tools or their own. We're glad to co-incubate Tau Ceti alongside Kim Morrison and the Mathlib Initiative. New roadmaps, roadmap review, and AI contributors are all welcome. #LeanLang #LeanProver #Mathlib #AI #Mathematics
5
89
6
370
37,934
jsmorph retweeted
Maybe too obvious to be worth saying, but: frontier models are now obviously superhuman at some mathematical tasks, including ones that the profession has, historically, rewarded with prestige etc.
40
175
25
2,878
170,949
Where —and what— was Homer’s Ithaca? This is an absolutely fascinating story and it’s a long one, so you should make yourself a cup of tea and settle in and enjoy this … antigonejournal.com/2026/07/…
3
4
1,747
when you are beating the benchmarks
1
21
1
313
30,745
jsmorph retweeted
Proposal: Any AI company that is earnest about "helping mathematicians" (rather than parading the intelligence of its models) should work on making a high quality AI referee service available to journals. Current peer review in mathematics is laughably slow, and accelerating that with AI would be easier and more helpful than autonomously proving the kind of random results we've seen so far, with the notable exception of the Unit Distance Conjecture.
12
18
4
221
38,303
Dear people at @OpenAI. Could you add a toggle to run memory off easily in an individual chat? Sometimes I need a response not polluted by my previous conversations.
50
12
7
797
68,547