I’ve never felt closer to an AI than when it swears in French while doing math. Is this what AGI feels like ?
Etale Cohomology was introduced in the infamous SGA 4 1/2, which is more magical than Platform 9 3/4.
According to J.S. Milne, Grothendieck "chose the word because the way he pictured étale morphisms reminded him of a calm sea at high tide under a full moon".
So what is this graph ⬆️? It's a visualization of all SGA 4 1/2 artefacts (lemmas, propositions, theorems,...) and their dependencies. You can explore it here: sgagraph.vercel.app/
-- Built with an LLM, so to be taken with a pinch of salt.
This keeps resurfacing in discussions about AI and mathematics — especially the gap between how math is practiced and how it’s often framed in AI narratives.
And it echoes my recent conversations with @Thom_Wolf and @davidbessis.
AI is often hyped as a ‘math-proving superpower,’ but real mathematical practice is evolving quietly and socially. I asked mathematicians what they see happening. What’s changing isn’t mathematicians’ relevance, it’s how they work.
dsleo.github.io/maths/2025/1…
Also looping in @evelovesolive @CarinaLHong @f_charton @JiaLi52524397, hoping this naturally resonates with others also building in this space.
AI is often hyped as a ‘math-proving superpower,’ but real mathematical practice is evolving quietly and socially. I asked mathematicians what they see happening. What’s changing isn’t mathematicians’ relevance, it’s how they work.
dsleo.github.io/maths/2025/1…
I’ve always found the making of scientific legends thrilling to read… though a bit mythologized for real life. It sets impossible bars for young researchers, you can still flourish without becoming the next Tao.
Whether von Neumann really multiplied 8-digit numbers as a child or not, it’s okay if you can’t... at any age.
@slatestarcodex @davidbessis
“The product of mathematics is clarity and understanding, not theorems by themselves.” W.P. Thurston. Yet most AI in math is still stuck on prover models. The bigger frontier is being missed.
Geordie Williamson has been a pioneer in AI for mathematical discovery. Listen to his latest talk @SimonsFdn youtu.be/Uxr_HrbN1MQ?si=BS7U… AlphaEvolve over AlphaProof !
Léo Dreyfus-Schmidt retweeted
🚀 NuminaMath 1.5 is here! 🚀
900k+ high-quality competition math problems with CoT solutions, new problem metadata, manually verified Olympiad problems, and more! 📚🏅
Check it out: 🔗 huggingface.co/datasets/AI-M…
Thanks to @Will424408 @dsleo
Léo Dreyfus-Schmidt retweeted
⚡️Preprint: Why do tree-based models still outperform deep learning on tabular data?
We give solid evidence that, on tabular data, achieving good prediction is easier with tree methods than deep learning (even modern architectures) and explore why
hal.archives-ouvertes.fr/hal…
1/9
Léo Dreyfus-Schmidt retweeted
📍 Dream Job Alert📍Does a #postdoc on #neuralnetworks frontiers look like the challenge you are looking for? Don't miss the unique opportunity to work with @GaelVaroquaux and our very own @dsleo. Apply today!
➡️ bit.ly/3RuUTHc











