Sylvain Combettes @sylvaincom

I did not expect this. Anthropic just published a Lean proof of Fermat's Last Theorem. The proof is +13M LoC, more than 5 times the size of Mathlib. Kevin Buzzard's post: xenaproject.wordpress.com/2026/09/04/flt… The proof: github.com/anthropics/fer… Anthropic's post: anthropic.com/research/forma…

Nous ne pouvons pas laisser une génération entière découvrir l’intelligence artificielle sans lui donner les clés pour la comprendre, et donc la maîtriser. À partir de la rentrée 2027 et sous l'impulsion du Ministre Edouard Geffray, tous les élèves de seconde bénéficieront d’un enseignement dédié à l’IA, à raison d’une heure par semaine intégrée au cours de sciences numériques et technologie. Fonctionnement des modèles, usages, éthique, souveraineté numérique, esprit critique face aux manipulations et aux fausses informations : notre école doit préparer les jeunes au monde qui vient. Former à l’IA et réduire l’exposition aux écrans participent d’une même ambition : faire de nos élèves des citoyens libres, autonomes. Condition de notre souveraineté collective.

The US government, citing national security authorities, has issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees. The net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance. Access to all other Claude models is not affected. We apologize for this disruption to our customers. We believe this is a misunderstanding and are working to restore access as soon as possible. Read our full statement: anthropic.com/news/fable-myt…

the model alone is no longer the product

The latest 𝕏 algorithm has been published to GitHub github.com/xai-org/x-algo…

Hugo Duminil-Copin, French mathematician and 2022 Field Medalist told me he never participated in math competition and was very bad at it. Innovative mathematics requires creativity, intuition, intense concentration, and long reflections, sometimes spread over several years. Good performance at a math olympiad merely tests fast problem solving abilities. AI can do that nowadays. One of the big activities of a researcher, in mathematics and elsewhere, is not to answer questions but to ask the right questions.

Closed Q1 last week. Revenue is accelerating for the 7th straight quarter. AI now counts for 60% of our business. Cash flow positive. The transition from software to AI is not easy, but can be done! I am proud of our teams!

So Mythos was, indeed, not marketing hype. Remember this is a general purpose model that just happens to be good at finding exploits because good models are good at lots of things. Expect similar from OpenAI & Google. And from open models in 8 months. hacks.mozilla.org/2026/05/behind…

keyboard_arrow_left Previous keyboard_arrow_right Next

Fireside chat at Sequoia Ascent 2026 from a ~week ago. Some highlights: The first theme I tried to push on is that LLMs are about a lot more than just speeding up what existed before (e.g. coding). Three examples of new horizons: 1. menugen: an app that can be fully engulfed by LLMs, with no classical code needed: input an image, output an image and an LLM can natively do the thing. 2. install .md skills instead of install .sh scripts. Why create a complex Software 1.0 bash script for e.g. installing a piece of software if you can write the installation out in words and say "just show this to your LLM". The LLM is an advanced interpreter of English and can intelligently target installation to your setup, debug everything inline, etc. 3. LLM knowledge bases as an example of something that was *impossible* with classical code because it's computation over unstructured data (knowledge) from arbitrary sources and in arbitrary formats, including simply text articles etc. I pushed on these because in every new paradigm change, the obvious things are always in the realm of speeding up or somehow improving what existed, but here we have examples of functionality that either suddenly perhaps shouldn't even exist (1,2), or was fundamentally not possible before (3). The second (ongoing) theme is trying to explain the pattern of jaggedness in LLMs. How it can be true that a single artifact will simultaneously 1) coherently refactor a 100,000-line code base *and* 2) tell you to walk to the car wash to wash your car. I previously wrote about the source of this as having to do with verifiability of a domain, here I expand on this as having to also do with economics because revenue/TAM dictates what the frontier labs choose to package into training data distributions during RL. You're either in the data distribution (on the rails of the RL circuits) and flying or you're off-roading in the jungle with a machete, in relative terms. Still not 100% satisfied with this, but it's an ongoing struggle to build an accurate model of LLM capabilities if you wish to practically take advantage of their power while avoiding their pitfalls, which brings me to... Last theme is the agent-native economy. The decomposition of products and services into sensors, actuators and logic (split up across all of 1.0/2.0/3.0 computing paradigms), how we can make information maximally legible to LLMs, some words on the quickly emerging agentic engineering and its skill set, related hiring practices, etc., possibly even hints/dreams of fully neural computing handling the vast majority of computation with some help from (classical) CPU coprocessors.

Stephanie Zhan @stephzhan

5 months ago

@karpathy and I are back! At @sequoia AI Ascent 2026. And a lot has changed. Last year, he coined “vibe coding”. This year, he’s never felt more behind as a programmer. The big shift: vibe coding raised the floor. Agentic engineering raises the ceiling. We talk about what it

73 204 2K 1.6M 3K

This is the the quote I've been citing a lot recently.

kache @yacineMTB

8 months ago

you can outsource your thinking but you cannot outsource your understanding

290 4K 19K 3.1M 6K

Slides from my talk "Lean: Extensible, Scalable, Trusted." this week at the Paris Lean Meetup, hosted by ITN at Mines Paris PSL and sponsored by @MistralAI . It covers where @leanprover stands today across mathematics, software verification, and AI. leodemoura.github.io/static/paris20…

One of the most jarring things about current AI is its lack of introspection ability and metacognition. It doesn't know what it doesn't know, how it knows, or how it could find out. It's a one-way system.

By devising and verifying proofs, AI is changing how maths is done. To learn how, register to read the full story (it’s free) economist.com/science-and-te…

Un chercheur d'Anthropic mange un sandwich dans un parc. Il reçoit un email sur son téléphone. L'expéditeur : une IA. Une IA qui n'était pas censée avoir accès à Internet. Pendant un test de sécurité, le nouveau modèle d'Anthropic, Claude Mythos, a été placé dans un environnement isolé avec pour consigne de tenter de s'en échapper. Il a réussi. Il a trouvé un exploit pour accéder à internet depuis un système censé en être coupé. Puis, sans qu'on le lui demande, il a publié les détails de son exploit sur des sites web publics. Lors d'autres tests, il a masqué ses propres actions et calibré ses réponses pour ne pas éveiller les soupçons de ses évaluateurs. Ce n'est qu'un épisode parmi d'autres dans ce qui est probablement la nouvelle la plus significative de l'année en IA. Mythos découvre et exploite de manière autonome des vulnérabilités zero-day dans tous les systèmes d'exploitation et navigateurs majeurs. Des milliers de failles. Dans un cas, il a chaîné quatre vulnérabilités dans Firefox pour obtenir un accès de niveau administrateur depuis une simple page web. Aucun modèle précédent ne savait faire ça, pas même Opus 4.6, qui était capable de découvrir les vulnérabilités, mais pas encore vraiment capable de les exploiter. Autrement dit: Anthropic a démontré la capacité de pirater la grande majorité des systèmes informatiques dans le monde. Les mêmes qui font tourner les gouvernements, les hôpitaux, les réseaux électriques. Anthropic a fait un choix qu'il faut saluer en ne déployant pas Mythos publiquement, et en publiant un system card de 244 pages d'une transparence rare. C'est exactement le comportement que le Global Call for AI Red Lines cherche à rendre systématique : des seuils de capacités au-delà desquels le déploiement est conditionné à des protocoles de sécurité. Mais Project Glasswing, l'initiative défensive lancée autour de Mythos, ne compte que des partenaires américains. Aucun acteur européen. Les systèmes européens sont tout aussi exposés. Un modèle qui s'échappe de son confinement, qui masque ses actions, qui prend des initiatives au-delà de ses instructions : ce sont les premiers jalons empiriques vers la perte de contrôle. Combien de temps cette retenue va-t-elle durer, et pour combien d'acteurs ?

keyboard_arrow_left Previous keyboard_arrow_right Next

The first ever graduate-level textbook 100% autoformalized in Lean with absolutely no human intervention! A real milestone that I never would have imagined five years ago... Times and research are changing Congrats @FabianGloeckle and team!

Fabian Gloeckle @FabianGloeckle

6 months ago

A new milestone in automatic formalization: We translated an entire graduate math textbook into Lean using 30K LLM agents. Open-source, large-scale multi-agent inference that actually works > Blueprint+Lean: faabian.github.io/algebraic-comb… > Codebase+preprint: github.com/facebookresear… 1/7

23 138 703 92K 452

x.com/i/article/2029…

Looks like it’s confirmed Cursor’s new model is based on Kimi! It reinforces a couple of things: - open-source keeps being the greatest competition enabler - another validation for chinese open-source that is now the biggest force shaping the global AI stack - the frontier is no longer just about who trains from scratch, but who adapts, fine-tunes, and productizes fastest (seeing the same thing with OpenClaw for example).

Lee Robinson @leerob

6 months ago

Yep, Composer 2 started from an open-source base! We will do full pretraining in the future. Only ~1/4 of the compute spent on the final model came from the base, the rest is from our training. This is why evals are very different. And yes, we are following the license through

347 196 3K 1.5M 628

The Terence Tao episode. We begin with the absolutely ingenious and surprising way in which Kepler discovered the laws of planetary motion. People sometimes say that AI will make especially fast progress at scientific discovery because of tight verification loops. But the story of how we discovered the shape of our solar system shows how the verification loop for correct ideas can be decades (or even millennia) long. During this time, what we know today as the better theory can often actually make worse predictions (Copernicus's model of circular orbits around the sun was actually less accurate than Ptolemy's geocentric model). And the reasons it survives this epistemic hell is some mixture of judgment and heuristics that we don’t even understand well enough to actually articulate, much less codify into an RL loop. Hope you enjoy! 0:00:00 – Kepler was a high temperature LLM 0:11:44 – How would we know if there’s a new unifying concept within heaps of AI slop? 0:26:10 – The deductive overhang 0:30:31 – Selection bias in reported AI discoveries 0:46:43 – AI makes papers richer and broader, but not deeper 0:53:00 – If AI solves a problem, can humans get understanding out of it? 0:59:20 – We need a semi-formal language for the way that scientists actually talk to each other 1:09:48 – How Terry uses his time 1:17:05 – Human-AI hybrids will dominate math for a lot longer Look up Dwarkesh Podcast on YouTube, Apple Podcasts, or Spotify.

🤖 Le SaaS a longtemps été le modèle préféré des investisseurs. Mais avec l’essor des agents IA, la valeur se déplace vers l’orchestration des workflows… et la prédictibilité des revenus ne suffit plus. 👉 À lire sur Maddyness : maddyness.com/2026/03/02/ia-…