

✓co-founder+engineer(dust.tt); alumni(openai.com, stripe.com, stanford.edu, polytechnique.edu);



Starting with HOL exploration and then moving on to generating the first LLM proofs on Metamath[0][1] to finally pushing LLM proving capabilities in Lean[2][3] between 2019 and 2022. I have to say.... I can feel the AGI. [0] I thiiiiink these are the first ever authored LLM proofs in a formalised maths library: us.metamath.org/mpeuni/r19.2… us.metamath.org/mpeuni/rntrc… (LLM authored March-July 2020) but I'm sure @ChrSzegedy would argue otherwise :p [1] arxiv.org/pdf/2009.03393 [2] arxiv.org/pdf/2202.01344 [3] openai.com/index/formal-math…

The Lethal trifecta... of getting beautiful shit done. Figma <> Dust <> Snowflake Ship an interactive information-rich Frame with the latest data AND the latest designs.

Now that I'm not in back to back meetings... happy to announce we have added 20 new MCP partners! And many more to come.


With terraform AND agents the cost of managing an infra has plummeted over the past few months. The buy vs maintain tradeoff has likely violently shifted recently. For EU that means EU cloud providers have become massively more competitive overnight vs GCP/AWS/Azure whose stickiness is directly correlated with the cost of maintaining services. Will the market adjust? If so, how long will it take?

desargues got backed by @spolu Stan's gpt-f, pact and minif2f built my foundation in ai4math minif2f showed what a public benchmark can do for a field, and we're building one for canonical lean

The more your company relies on AI, the more control over it matters. Maintaining choice is maintaining control.

Pretty cool work from desargues.ai team working on a proof quality benchmark for Lean desargues.ai/preview/formal-… Proving is not just about correctness it’s also about taste, style, and “interestingness”. Pushing agents to be better there seems like a pretty important thing if we want to maintain our capability to understand and interact with modern synthetic formal mathematics.

Must watch. @spolu unrolls how they do preserve an understanding of the system and do verification with agents using code contracts code-contracts.cc talk: youtu.be/eg8u-0102lw?t=717

Full talk at DotAI 2026 on code contracts and why I believe code will remain the artefact we use to build and understand our future software systems (even if we don't write it). Vibe coding and spec driven dev / light off software factories have extended the line on the criticalitly/formality graph, but if you're maintaining a critical business system, code will remain the only valid artefact possible to understand it: you don't understand a system if you can't maintain it, and no, a set of informal product and architecture specifications is not "maintainable". Call me old-school, but if I'm the one getting paged at 4am, I sure as hell won't let you vibe code your way in my codebase or send a swarm of light-off agents to deface it. Slides: app.dust.tt/share/frame/2cef… Video: youtube.com/watch?v=eg8u-010…

"France built 30 fission reactors, Concorde and TGV all within 20 years." is the best quote I heard today from Rustem who's working on Firefly Fusion in France. That's Commonwealth Fusion, Boom supersonic, and {insert your future high-speed train startup name here}; all seeded, scaled, and S-curved 30-years ago at incredible speed for the time. I/We need to study how we pulled that one off. What were the finance of the country like? What was the mix of private/public implication in these? How regulated? How was it funded? Many questions.

Take notice. It’s coming and it’s gonna get wild.

h/r adbq

Some more thoughts on R>1: