AI4Math & Autoformalization PhD Student @EPFL Student Researcher @GoogleDeepMind Previously @mathematics_inc @OpenAI
Joined June 2023
- Tweets51
- Following174
- Followers214
- Likes260
Auguste Poiroux retweeted
To say that we are living through a transformational time for mathematics is not an overstatement. AI's growing ability to find proofs coupled with ability to check them quickly due to the pioneering work of the formal mathematics community is leading to remarkable answers. But mathematics will never be "solved": every answer opens up new questions.
In many ways, it is just the beginning - a time to work on asking new and even more important questions than before, and understanding the answers. We look forward to working with the mathematical community in this new era.
Auguste Poiroux retweeted
16 months ago, the largest LLM-generated Lean formalization project was ~3.5k lines of code. OpenAI’s new release has 26 million - a ~7,400x jump in scale!
the chart below: across 5 selected frontier releases, codebase size has doubled every ~38 days.
(thanks to @augpoi for the original chart! this one plots total LOC, not LOC/day, since i don't have reliable development times, so may look different)
When will most existing mathematics be formalized?
Possibly 2027, based on the current trend.
Frontier Lean formalization throughput (LOC/day) is doubling every ~35 days. Anthropic's FLT result lands almost exactly on the frontier previously paced by Math Inc.
ALT Log-scale scatter plot titled "Formalization Throughput", showing Lean formalization throughput in lines of code per day from June 2025 to September 2026. Eleven projects rise from about 10² LOC/day to over 10⁶ LOC/day. Four highlighted frontier milestones, de Bruijn ABC, Strong PNT, Sphere Packing, and Anthropic’s Fermat’s Last Theorem formalization, lie close to an exponential trend corresponding to a doubling every ~35 days. Other projects include M2F, Meta algebraic combinatorics, Unit distance, OpenAI ten proofs, S⁶ formalization, an Erdős backlog sweep, and Axiom Prover LeanEval. A broader trend fitted to all 11 projects doubles about every 39 days. Anthropic FLT is the highest-throughput point, at roughly 1.2 million Lean LOC/day.
Auguste Poiroux retweeted
Introducing Gemini 4 Argon – our new frontier model.
It’s built for complex workflows across coding, enterprise knowledge work, and cybersecurity defense – rolling out today to a set of trusted testers through our Fairwind Program.
Auguste Poiroux retweeted
There’s a lot of excitement about AI + formal verification right now.
I’ve worked on this since 2018. I share the excitement—but the progress has also made me rethink what formalization can solve.
What Happens When Formalization Becomes Cheap?
yangky11.github.io/blog/ai-a…
Auguste Poiroux retweeted
Whoops, chose the wrong week to share our paper…
Something about someone named Navier and Stokes????
But hey the code is out now! Go play with latent identification probes 🤾
💻 Code: github.com/bayazitdeniz/late…
1/🚨 New paper #emnlp2026
Do multilingual #LLMs work in English? It depends on how you ask. In "Lingua Franca or Probing Artifact?", we probe the same hidden states with three latent language identification methods and get different answers. 🧵👇
#interpretability #NLProc
Auguste Poiroux retweeted
One on math: super-human mathematician AI by June 26. The other: 95% at the end of 2025 on the current test set of of ARC.
We can work out the details on the first. I feel much less confident about the second, esp. given the current challanges with evaluation on proprietary models.
Auguste Poiroux retweeted
If I did not believe in the intrinsic and underrated usefulness of mathematics, I would be somewhat sad about what AI is doing to it.
I think Jevons paradox will strike again: much cheaper and deeper mathematics will become even more of the fundamental infrastructure of the future than it is today.
Math will no longer be centered on conjectures posed by humans. It will be continually postulated, solved, processed, and utilized by AI systems as they work on the hard scientific, engineering, and computing problems of the future.
Auguste Poiroux retweeted
1/🚨 New paper #emnlp2026
Do multilingual #LLMs work in English? It depends on how you ask. In "Lingua Franca or Probing Artifact?", we probe the same hidden states with three latent language identification methods and get different answers. 🧵👇
#interpretability #NLProc
Auguste Poiroux retweeted
Yes, this result cost millions of dollars.
But remember that when @OpenAI announced o3 it cost ~$500,000 to score 87.5% on ARC-AGI 1. Today, Astra scores higher for ~$20.
In 2025 it took us and GDM an enormous amount of compute to achieve IMO gold. For the 2026 IMO, anyone with a $20/month ChatGPT subscription could do it.
Massively scaling test-time compute gives us a glimpse of the future. I believe that a year from now everyone will have an AI at their fingertips capable of solving problems of this caliber.
We’re sharing a solution to the Navier-Stokes Millennium Prize Problem, one of the deepest problems at the frontier of mathematics.
The proof was produced by a group of agents, using an OpenAI next-generation model significantly more capable than GPT-6 Astra.
The problem concerns whether the description of smooth three-dimensional fluid motion modeled by the Navier-Stokes equations can break down. It has remained unresolved for roughly 90 years.
When will most existing mathematics be formalized?
Possibly 2027, based on the current trend.
Frontier Lean formalization throughput (LOC/day) is doubling every ~35 days. Anthropic's FLT result lands almost exactly on the frontier previously paced by Math Inc.
ALT Log-scale scatter plot titled "Formalization Throughput", showing Lean formalization throughput in lines of code per day from June 2025 to September 2026. Eleven projects rise from about 10² LOC/day to over 10⁶ LOC/day. Four highlighted frontier milestones, de Bruijn ABC, Strong PNT, Sphere Packing, and Anthropic’s Fermat’s Last Theorem formalization, lie close to an exponential trend corresponding to a doubling every ~35 days. Other projects include M2F, Meta algebraic combinatorics, Unit distance, OpenAI ten proofs, S⁶ formalization, an Erdős backlog sweep, and Axiom Prover LeanEval. A broader trend fitted to all 11 projects doubles about every 39 days. Anthropic FLT is the highest-throughput point, at roughly 1.2 million Lean LOC/day.
At a ~35-day doubling time, the difference between a 1B and 100B corpus is only ~8 months.
At the current trend, ~100M Lean LOC/day is reached around April 2027, enough for a 1B-LOC corpus in about 10 days.
Auguste Poiroux retweeted
Did not expect a shoutout from the White House:
Formalizing the Prime Number Theorem
& follow-on conversation with Terry Tao on the future of mathematics
In response to @POTUS's call for a revival of America's scientific enterprise, today I’m releasing Science: A New Golden Age.
This report is a blueprint for renewing American scientific leadership for the 21st century.
Auguste Poiroux retweeted
mistral.ai/news/leanstral-1-…
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: github.com/mistralai/Leanstr…
We also open-source LeanstralSafeVerify and FLTEval.
Auguste Poiroux retweeted
Just caught up with @augpoi who interned at OpenAI with me. It is so cool to see his trajectory in the field of Maths and AI.
@augpoi @KunhaoZ Mantas and Marco all interned at OpenAI with me on LLM x Formal Maths. 4 out of 4 stuck with Maths and AI. If there is one achievement I’m proud of of my time at OpenAI, it’s certainly this one.
@augpoi worked at Maths Inc is now at DeepMind. @KunhaoZ is now at MetaAI. Mantas and Marco both at Cambridge / Imperial with strong ties to labs as well as @ProjectNumina
Strong reminder that at end of the day, the accomplishments are not about the what or the money, it’s all about the people. Yes cheesy. But also true.
Auguste Poiroux retweeted
FormalQualBench is a new formal math benchmark simulating the rigor and level of expertise expected of a math PhD qualifying exam. We introduce a new benchmarking standard called specification-based evaluation, which uses the Lean FRO's Comparator tool to enforce a strong standard of correctness.
Unlike many benchmarks which are private, wrap model APIs with an undisclosed harness, and are marketed to data teams at commercial AI labs, FormalQualBench is open-source, evaluates generally available tools run with their default settings, and is aimed at practitioners who want to choose the best tools to reach the frontier of autoformalization capabilities as effectively as possible.
math.inc/formalqualbench
Auguste Poiroux retweeted
Today, at the @DARPA expMath kickoff, we launched 𝗢𝗽𝗲𝗻𝗚𝗮𝘂𝘀𝘀, an open source and state of the art autoformalization agent harness for developers and practitioners to accelerate progress at the frontier.
It is stronger, faster, and more cost-efficient than off-the-shelf alternatives. On FormalQualBench, running with a 4-hour timeout, it beats @HarmonicMath's Aristotle agent with no time limit.
Users of OpenGauss can interact with it as much or as little as they want, can easily manage many subagents working in parallel, and can extend / modify / introspect OpenGauss because it is permissively open-source. OpenGauss was developed in close collaboration with maintainers of leading open-source AI tooling for Lean.
Read the report and try it out:
Auguste Poiroux retweeted
🚨 Math Inc is introducing FormalQualBench: an open-source benchmark for end-to-end auto-formalization capabilities with math PhD qualifying exam level problems.
We build this benchmark for the Lean community, allowing anyone to compare different auto-formalization agents!
