David Picard

#KostasThoughts: Stay organized and keep intermediate, raw project results. You never know when you, or your supervisor 😉, will need them later for a talk.

J'espérais en trouver une qui s'en approche, illustrée dans une vieille bd, mais rien ne correspond vraiment. À la place, voici le corbeau et le renard, pour le plaisir des yeux.

Whereas this is more like some 90s Blair Witch Project stuff.

Yeah, it's bizarre. It's the same model that continues training. I don't get why it's grainy, and where the low res 20th century aesthetics comes from. Those ones have some 60s-70s sci-fi vibe.

Exploring more weirdness.

OH "the S in MCP stands for security"

Let's say you give it the rights to execute pdftotext and read the output, so that it can access new types of documents. Seems harmless right? What if there's a hole in pdftotext? Can someone craft a prompt and a pdf that exploits the hole?

I was going to post exactly that! I'm way too dumb to learn on my own, but with this job, I'm constantly learning from the best. That's awesome.

Tsss, goal posts can always be moved. I find your lack of faith disturbing.

But then you're missing some functionalities. For example you'd like your agent to convert files into a readable format, which requires executing some commands (pdftotext or similar).

Or whole-body layouts, to get the hands and interaction.

✨ New 3D pose estimation method from my lab! #FMPose3D allows for monocular (i.e. single camera) 2D➡️3D 🔥

Led by Ti Wang & w/ Xiaohang Yu #FMPose3D is SOTA on human & animal 3D benchmarks, & will be integrated into @deeplabcut.bsky.social ⬇️

📝 arxiv.org/abs/2602.05755
➡️ xiu-cs.github.io/FMPose3D/

Et encore, si ce truc était beau, on pourrait parler de compromis. Mais là, premier prix du parpaing d'or, au moins.

Put the "warning graphic content" flag on, please. Every cell of my (French) body recoiled and screamed in horror 😱🤮

Of course you're not going to "prompt" entire games. But I mean that asset creation, physics calculation, and rendering will be done using generative models. It's much easier than trying to handcraft all of those. What makes a game is the gameplay, not technical details like textures and collisions.

YouTube video by gameranx Is The Gaming Industry COOKED?

Interesting video showing the gamers' viewpoint on recent video generative models: www.youtube.com/watch?v=Xsae...

I think this is not going to age well. Here is a prediction I feel confident about: in 10 years, most big title video game will rely on generative world models.

Ce qui est complètement ubuesque, c'est que sans l'affaire Epstein, il y serait encore à 86 ans. C'est fou de se dire qu'il n'y avait pas un seul jeune de 60 ans pour le remplacer 🙄

You're not going to convince me that "you can do it" by merely saying you do, I can see the outcome you know.

🤐 about mathematics

Most importantly, it does not scale. In some domains, you have 20+ applications per reviewer. Good luck reading even only 3 papers per candidate (in addition to the rest of the application) in one month for the pre-interview selection jury. That's already a hefty full-time job.

I get the spirit of LLM-2) about sharing - papers under review are under NDA, communicating them to a 3rd party is disclosure - but now that there are many solutions for running a local model (eg, github.com/janhq/jan), maybe this should be more specific to avoid sounding arbitrary.

Check out what’s new at #ECCV2026 Malmo 🇸🇪

Full details: eccv.ecva.net/Conferences/...

Est-ce que ça aurait du sens de calculer ces cartes avec des durées utilisant les transports en commun, le budget des étudiants étant ce qu'il est ?

Trop bien!
Je ne suis pas sûr de bien comprendre comment les temps ont été calculés. Je regarde la carte d'Île de France, et j'ai l'impression que plus de la moitié des cellules <30min n'est pas à 30min. Par exemple, n'importe quel bled autour de Melun ou Meaux, c'est l'enfer.

It's the most illogical of them all, right?

Secrets of Intelligence Services I asked Claude code to build Claude code. You won’t believe what happened next.

I asked Claude Code to build Claude Code. You won’t believe what happened next.

We are recruiting four positions connected to Machine Learning, Statistical Learning, and AI for Science in the Applied Mathematics department at École polytechnique. Join our vibrant community at IP Paris and Hi! Paris IA center. List below🧵 tinyurl.com/3jpw9t26

Actually, writing a cool program in 6502 assembly could be a great team building fun event...

seems way more scalable because you see many problems in one go. Reading is way faster than writing.
So no, I don't think learning how to code is obsolete because of coding agents.
But it's true it enables some people to do things they couldn't do before.

you don't have the skill to do 2. So it's just very inefficient. You're going to manually test every corner case and then tell your agent that it's not correct and that the code must change. It just doesn't scale.
Doing 2. on the other hand, reading the output and giving the feedback to the agent,