David Picard

J'allais faire la même blague

Let me ask the question that is going to have me burn on the pyre: why do we need models to learn good representations already? Why does it matter?

This, but the proof has to contain an error that is not discovered until the first million coins are produced 🙃

I'm not shy about it ;) It's fairly well cited, so I suppose some people found it useful

Torch.manual_seed(3407) is all you need: On the influence of random seeds in deep learning architectures for computer vision In this paper I investigate the effect of random seed selection on the accuracy when using popular deep learning architectures for computer vision. I scan a large amount of seeds (up to $10^4$) on CIF...

Something I wrote a while back when I was frustrated by the lack of robustness in computer vision publications. Not that it has changed since.
arxiv.org/abs/2109.08203

This is such a good illustration! Too bad I didn't have the idea when I wrote that silly little paper a few years ago.

Anyway, remember folks: torch.manual_seed(3407) is all you need
No other seed had been subject to as much scrutiny as this one
And it's all yours for free!

I'm so mad that nobody remembers it's 3407. It's like I wrote that paper for nothing...

Hey, il a pas tort ! Je suis à peu près sûr que le directeur de Verimag a soit l'info directement, soit peut coder un script pour l'extraire ! 🤣

Nothing fancy. I just really like Gustave Moreau. I like this image. I find it soothing, so I (wrongly) thought it would prevent me from being too aggressive on the Linux forums I was visiting at the time. It stayed ever since.

I've used this other Gustave Moreau painting (Arion) on other places.

You cannot make this thing up

Ich dachte es sollte etwas mit Sehnsucht sein

Well, the data is already out there and the conclusion is pretty clear...

Me too, but to be honest, those human excerpts were not at all to my liking. They would not have won against other human written texts that I can think of.

Wow! Never thought I would see it on the internet ever again. I've been using this picture as an avatar for at least 20 years.
Glad it's available in higher resolution than it used to.

How many pages of appendix proofs did you have to go through so far?

Also, we all know it's not even a good seed compared to 3407 🙃

À Verdun, ces villages morts pour la France qui ont encore un maire Six communes voisines de Verdun détruites pendant la Première Guerre mondiale n’ont jamais été rebâties. Elles sont encore administrées par des maires, gardiens

La guerre produit toutes sortes de choses, mais celle-ci est inattendue : depuis plus d’un siècle, nous avons des maires sans électeurs

Merci @pierrejova.bsky.social pour cette histoire de France vécue dans la chair

www.lavie.fr/actualite/so...

Cleopatra

Cleopatra https://www.wikiart.org/en/gustave-moreau/cleopatra

EU Council President Antonio Costa said Russia has emerged as “the only winner” from the Iran War.

Learning Functions of Halfspaces We give an algorithm that learns arbitrary Boolean functions of $k$ arbitrary halfspaces over $\mathbb{R}^n$, in the challenging distribution-free Probably Approximately Correct (PAC) learning model, ...

Well, this seems like a big deal. arxiv.org/abs/2603.087...

"This is the first algorithm that can PAC learn even intersections of two halfspaces in time 2^o(n)."

He should ask all députés to get an account. That's how sovereignty works, get your data at home!

Marianne Clausel News : Thesis Offer Apply before May, 5th With E. Devijever (LIG & ELLIS Member), we are offering a PhD thesis funded by INESIA, the new French AI Safey Institute Towards a framework based on causali...

🏹 Job alert: PhD in Causal Inference at INESIA, the new French AI Safety Institute

📍 Nancy 🇫🇷
⏰ Apply by 5 May
🔗 https://bit.ly/46LVCwR

With a pretty good cats per figure ratio!

ok, this one is going to be real tough 😅

they should add reaction emojis to openreview

That the EU is not announcing a 100+B€ plan to get rid of our dependency on (foreign) fossil fuel as fast as possible is disappointing.

Let's build Si ingot factories, let's build Na-ion battery factories, even if it's not state-of-the-art! old-but-subsidized tech ≫ newest tech

Formalizing a proof in Lean using Claude Code YouTube video by Terence Tao

youtu.be/JHEO7cplfk8 🍿

🧵 New preprint! Towards Understanding Steering Strength w/ M. Taimeskhanov and D. Garreau

Activation steering is a popular way to control LLM behavior at inference. But how much should you steer? We provide the first theoretical analysis of the steering strength α.

📄 arxiv.org/abs/2602.02712

1/7

It will happen. News very soon 😉

Still infinite orders of magnitude better than automating the decision to bomb 12-yo girls in their school 🙄