Leo (@biosemiote) header
Leo (@biosemiote) profile picture

Founding Member of Technical Staff @axiommathai. Simian. Not post-trained. https://teapawt.org/t.co/Lu9iOQuHRq

Joined March 2022
Photo by Leo (@biosemiote)Photo by Leo (@biosemiote)Photo by Leo (@biosemiote)Photo by Leo (@biosemiote)Photo by Leo (@biosemiote)Photo by Leo (@biosemiote)Photo by Leo (@biosemiote)
Pinned Tweet
Check out our revised lattice deduction transformers paper and the updated github repo at github.com/lcrh/lattice-dedu… Solves sudoku extreme at 99.999% with minutes of training compute. All tables are fully reproducible from the repo. See you at @NeurIPSConf
3
3
9
756
Leo retweeted
In our new NeurIPS paper, we show how to use epiplexity to select and generate data for out-of-distribution generalization! Paper: arxiv.org/abs/2608.11746 Joint work with Andres Potapczynski, @ShikaiQiu @edwardfhughes and @andrewgwils 🧵
19
51
7
339
29,569
Epiplexity helps us understand why certain data leads to better downstream generalization. Now we can use epiplexity directly for data selection, data mixing, curriculum learning, and even purely synthetic data generation! So excited for this research effort.
In our new NeurIPS paper, we show how to use epiplexity to select and generate data for out-of-distribution generalization! Paper: arxiv.org/abs/2608.11746 Joint work with Andres Potapczynski, @ShikaiQiu @edwardfhughes and @andrewgwils 🧵
5
30
336
14,510
Over the past week, I’ve been overwhelmed by the love and support of friends, industry colleagues, and people I've never met. Thank you to everyone who has shown kindness to my family and me. It means the world to me. I came to the United States 53 years ago, and becoming an American citizen is one of the greatest gifts and proudest moments of my life. NVIDIA has been my life’s work. Building it into the company it is today alongside my colleagues has been a privilege and an honor. Nowhere but in America could a company like NVIDIA be possible. I love our country, and I love NVIDIA. I did not live up to the standards expected of everyone at NVIDIA, and I have made serious mistakes in my personal life. I created problems for the Board that I deeply regret. I own my mistakes completely, and I am committed to cooperating with the Board’s independent investigation. Since my departure, I’ve seen an outpouring of calls from NVIDIA’s employees, investors, business partners, and customers for me to return as CEO. Many continue to reach out to me directly as the company navigates complex technology, market, competitive, and geopolitical dynamics. NVIDIA is a once-in-a-generation company. If the Board of Directors decides that my return is in the best interest of our shareholders, our employees, the company, and America, I would be honored to answer these calls and return as CEO. Thank you again to everyone who has shared their good wishes. I am profoundly humbled by your support. To my family: I love you, and I will spend the rest of my life being worthy of yours. Jensen
60,695
121,649
24,466
1,172,900
58,224,720
Proudly wearing the scars of doing formally verified, large scale software development using agents.
1
18
713
I've been playing with this a little bit and its very cool. Social networks spiral towards the equilibria of their signaling games, this one uses AI to change those games in an interesting way. Reminds me a little of some of the ideas in @robinhanson's age of em.
Bond is your personal AI memory. Meta profits off your data. ChatGPT trains on it. You get nothing. Bond works for you and gets you paid. Speak your mind to earn $20+ a month. Free on iPhone. Link below 👇
1
1
122
Recently, we formalized the Classification of Finite Simple Groups, the longest proof in math. Also the largest open-source software repo. Beyond certifying the result, we ask: Is it reusable? What else can we formalize? The answer is YES. 30 new results, built on top of it!🍇
13
26
6
134
13,030
Leo retweeted
This one weird trick improves alignment generalization! Just add random noise during training, telling the model to keep their values stable under perturbation
30
30
2
301
14,320
Thank you to everyone who attended our SGLang talk "Training Systems for Formalizing All of Mathematics"! @tobiges and @ithinkimrishi discussed how we at Axiom post-trained models and ran coordinated AI agent swarms that write proofs in Lean, resulting in projects like Stacks!
3
5
1
46
3,944
Proving binaries one at a time won't scale, even with AI writing the proofs. Prove the language, compiler and runtime once. Then what each program proves at source level carries to the binary. That is where I expect verification investment to go next.
9
12
1
82
3,700
It was a pleasure speaking about Training systems for Math Formalizations at the SGLang conference @sgl_project with @tobiges.
4
25
1,510
NextLat has been my baby that I've worked on for the past year and a half. We just released a new, and possibly final, revision of the paper on arXiv (arxiv.org/abs/2511.05963): 1/ Multi-token belief states and other insights into scaling NextLat for large-scale pretraining. These came out of our efforts to bring NextLat to the next generation of Phi models, though I'm sure y'all can guess how that went from @JohnCLangford's recent blog post. 2/ We show that NextLat also benefits linear-attention models (specifically, Mamba). 3/ More ablations to further validate NextLat's benefits. 4/ Added references to relevant work we previously missed (my sincere apologies!). 5/ NeurIPS template! 🥳 It's been quite a journey, and I'm incredibly proud of how far this work has come. Huge thanks to my fellow collaborators for their support and guidance!
19
39
2
319
15,944
Single card draw: The Tower
1
103
In 1922, Mordell asked a beautiful question: can a curve of genus at least two have infinitely many rational points? Six decades later, Faltings proved the answer is no, winning the Fields medal for his work. Today, Axiom announces the first machine-verified proof of Faltings’ theorem.
9
34
9
144
16,846
Congrats to Colin for leading this effort --- we want it to be an example of how to present formalizations of mathematics in a way that is helpful to students and experts seeking to learn, to teach, and to discover.
Today, Axiom Math is releasing a Lean formalization of Intro to Cluster Algebras Chapter 1-7 by Fomin, Williams, and Zelevinsky. The theories in this text show up from geometry to particle scattering. We showcase AxiomProver’s library learning capabilities across many domains.
4
21
1,297
Today, Axiom Math is releasing a Lean formalization of Intro to Cluster Algebras Chapter 1-7 by Fomin, Williams, and Zelevinsky. The theories in this text show up from geometry to particle scattering. We showcase AxiomProver’s library learning capabilities across many domains.
5
42
7
230
18,688
my vocal synth models the vocal tract as a network of branching acoustic tubes, unlike vocaloid which stitches together and pitch-shifts phoneme recordings. still working on improving naturalness by tuning it against recordings of real singers, but I'm already impressed at how cool it sounds!
I only make serious software
638
2,628
632
23,692
3,261,930
Excited to be speaking next week at SGLang Summit by @lmsysorg on how we've built scalable post-training infrastructure to push the frontier of math
Replying to @lmsysorg
An in-house training stack for mathematical formalization. Tobias Gessler from @axiommathai shares how the team approaches building training infrastructure for this specialized domain.
3
4
1
23
3,134
The circuitry of mathematics
The Stacks Project, autoformalized by Axiom Before intelligence runs on silicon, it runs on structure.
1
2
26
1,874