iiinova — where your digital self keeps evolving

Reposted by David Picard

Marine Tondelier @marinetondelier.fr · 3h

Le bilan objectif du ministre de l'Intérieur hier est un désastre absolu : - des dizaines de milliers de personnes empêchées d'exercer le droit fondamental de manifester en paix - des villes entières abandonnées aux casseurs pendant que des manifestants pacifiques étaient violentés

19333105 333

Reposted by David Picard

Gabriel Peyré @gabrielpeyre.bsky.social · 4h

The 1D heat flow ρ̇=ρ″ reads q̇=q″/(q′)² over quantiles which follow deterministic paths that never cross. www.gpeyre.com/blog/2026/10...

0101 10

David Picard @davidpicard.eurosky.social · 3h

I am very angry about the police who, for some reason, decide to throw grenades at peaceful families but are nowhere to be seen at the places where dangerous dickheads are pillaging. It's almost as if they care more about suppressing dissenting opinions than protecting people and goods 😡

130 3

David Picard @davidpicard.eurosky.social · 6h

I really don't understand how this doesn't count as stock manipulation

110 1

Reposted by David Picard

Silvia Sellán @silviasellan.bsky.social · 18h

Spending my Saturday satisfying my ICLR reciprocal reviewing duty, which means sifting through hundreds of pages of completely meaningless AI slop. What is my moral / professional obligation as a reviewer when faced with a meaningless AI-written submission?

1133 13

Reposted by David Picard

Juliette Marrie @jlt-m.bsky.social · 17/10/2026

Introducing REKI, a 2.4B vision-language-action model pretrained on public data only. With a simple design, it rivals larger models trained on in-house data with complex multi-stage pipelines, setting a new Pareto frontier between performance and size. kyutai.org/reki

2454 45

Reposted by David Picard

c0nc0rdance @c0nc0rdance.bsky.social · 17/10/2026

Drone photographer Theresa Graham (tgraham.photography on IG) captured this Ornate Eagle Ray (Aetomylaeus vespertilio) off Ningaloo Reef in Western Australia. These rays are called the "unicorn of the sea": fewer than 100 sighting have been documented. We suspect they're critically endangered.

16814242 814

David Picard @davidpicard.eurosky.social · 17/10/2026

Don't take it for granted, I can say a lot of bs. But I think there is some intuitive reasoning here, with the main question being how many of the free parameters remains usable after training (which is the way overparametrized network escape representer bottlenecks) to be leveraged in post-training

010 1

Reposted by David Picard

Helena (Yuhan) Liu @helenaliu24.bsky.social · 15/10/2026

(1/2) I’m starting a NeuroAI and comp neuro lab at the University of Toronto! How do brains and AI learn from ongoing experience? How do they connect delayed outcomes to past actions, adapt to change, and retain what matters? Recruiting funded PhD students for Fall 2027. Links in the reply 👇

24713 47

Reposted by David Picard

Matt Hodges @matthodges.bsky.social · 16/10/2026

Claude unsure if it did a breakthrough or just in The Matrix.

3202 20

Reposted by David Picard

Jean-Marc Lafon @jmlafon.eurosky.social · 16/10/2026

Manifestation "anti-guerre" à New York le 7 juillet 1941. Une ressemblance avec les éléments de langage de nos traîtres contemporains serait tout sauf fortuite.

Femme portant une pancarte : "Hitler has not attacked us, why attack Hitler?"
Au second plan, une autre affiche le message : "Why not peace with Hitler?"

Pancarte affichant : "Arm Britain and prolong the war"

Pancarte affichant : "Why not peace with Hitler?"

9225124 225

David Picard @davidpicard.eurosky.social · 16/10/2026

Case in point mu.social/profile/rene...

010 1

Reposted by David Picard

Eurosky Social @eurosky.social · 16/10/2026

We’ve had a complete website refresh now showcasing our products: - for people - for publishers - for organizations - for developers Check it out! eurosky.tech

eurosky.tech

Eurosky – The social internet, rebuilt in Europe

Eurosky is a European initiative to build and operate sovereign social web infrastructure

926474 264

Reposted by David Picard

LDH @ldh.fr · 16/10/2026

La #LDH demande l’interdiction totale des grenades offensives assourdissantes en maintien de l’ordre : ces armes mutilantes n’ont pas leur place dans les manifestations. Perdre une main ou un œil pour exercer sa liberté de manifester n’est pas acceptable en démocratie.

rtl.fr

Mobilisation du 17 octobre : Laurent Nuñez confirme que les grenades assourdissantes pourront être utilisées, malgré leur suspension

Invité de France 2 ce vendredi 16 octobre, le ministre de l'Intérieur a confirmé que les forces de l'ordre pourront recourir à des grenades assourdissantes en cas de débordements, lors de la manifestation de vendredi.

15420270 420

David Picard @davidpicard.eurosky.social · 16/10/2026

Today has a great "Magritte sky", with the deep blue and the white fluffy clouds.

2100 10

David Picard @davidpicard.eurosky.social · 16/10/2026

Cette image est magique. Je l'entends dans ma tête avec la voix de flic de Didier Bourdon et ça le cause des fous-rires 🤣

050 5

Reposted by David Picard

Emtiyaz Khan @emtiyaz.mastodon.social.ap.brid.gy · 15/10/2026

I recently moved to TU Darmstadt (Germany) and am currently hiring PhD student through ELISS. ellis.eu/news/ellis-phd-program-cal… If you interested, please apply and choose me as a potential advisor. Topics include Continual/Distributed/Federated/Active/Transparent […]

mastodon.social

Original post on mastodon.social

01910 19

Reposted by David Picard

Magali Reghezza @magalireghezza.eurosky.social · 15/10/2026

Presque 1 fille sur 5 a fait une tentative de suicide. 1 sur 5. Mais quelle société maltraite à ce point ses enfants qu’ils souhaitent mourir? Quelle société se targue de progrès quand elle est à ce point incapable de construire un avenir à sa jeunesse, qui lui donne envie de vivre au présent?

30680487 680

Reposted by David Picard

Christian Wolf @chriswolfvision.bsky.social · 16/10/2026

What is the current nettiquette? As a reviewer, can I claim that the paper has been written with too much LLM output?

211 2

David Picard @davidpicard.eurosky.social · 16/10/2026

Et ceux des collègues qui ont proposé de faire pareil, évidemment.

120 2

David Picard @davidpicard.eurosky.social · 16/10/2026

spanned by the training set and thus lead to richer solution, including some with larger recall due to larger coverage, at the cost of lower precision because it has also expended in regions of low density. Not sure if I'm clear of if this is too synthetic and would require a real write-up.

120 2

David Picard @davidpicard.eurosky.social · 16/10/2026

maximizing one or the other. What is surprising is that they can extrapolate to achieve larger recall or precision by doing weight averaging with non-convex combinations. (This is an incredible result) I'm wondering if this is equivalent to hallucinating training samples that would expand the space

120 2

David Picard @davidpicard.eurosky.social · 16/10/2026

Funnily enough, I had this discussion with Benjamin Negrevergne earlier this week (during his habilitation, amazing talk btw) on one of his recent paper with people from Meta: arxiv.org/abs/2605.28751 In this, they show a trade-off between precision and recall by training several models with setups

arxiv.org

Extrapolative Weight Averaging Reveals Correctness-Efficiency Frontiers in Code RL

Linear interpolation between fine-tuned checkpoints has been shown to trace the Pareto front between competing objectives, but whether extrapolative weight averaging can extend such frontiers to new c...

140 4

David Picard @davidpicard.eurosky.social · 16/10/2026

pretraining than for the SFT or the RLHF phase. To me, that's why fine-tuned models tend to have a mode collapse: the training set spans a much smaller space and regularization (explicit like wd or implicit just by SGD) makes you forget some of the directions that were spanned by the pretraining set

120 2

David Picard @davidpicard.eurosky.social · 16/10/2026

Id est, the model that you get after any training (pre or post) lives in a space that is spanned by the training set, and no more. This also means that the maximum recall achievable corresponds to the coverage of the data distribution by the training set. This coverage is much larger for the

120 2

David Picard @davidpicard.eurosky.social · 16/10/2026

It's not formally written, but it goes like that: We know from representer theorems that any regularized minimizer of an empirical risk is a combination of the training samples (this is not only true for kernel methods operating in a RKHS but also for deep neural network, see M. Unser's paper).

110 1

David Picard @davidpicard.eurosky.social · 16/10/2026

And much more important than that, how many people could not follow their dreams, despite having the capability, because they were a woman? Progress is nice, joy is nicer

000 0

David Picard @davidpicard.eurosky.social · 16/10/2026

Clever Hans effects have been known and studied for as long as deep learning exists. There are many mitigation techniques. Degenerate overfitting is usually not a problem when being careful enough.

000 0

Reposted by David Picard

Edouard Machery @edouardmachery.bsky.social · 16/10/2026

She is obviously right and European countries have been complicit in helping him avoid what he deserves. Die in a cell in The Hague.

1157 15

Reposted by David Picard

Gabriel Peyré @gabrielpeyre.bsky.social · 16/10/2026

Approximating attention with fewer tokens is hard. In recent work with Léa Bohbot, we show how token amplitude affects convergence: a pointwise 1/√n rate can hide much slower global guarantees. www.gpeyre.com/blog/2026/10...

gpeyre.com

How many tokens does attention need? - Homepage of Gabriel Peyré

Why finite-token attention can converge slowly despite a pointwise square-root rate: exponential query sensitivity, a radius–tail tradeoff in L2, and interactive Gaussian attention curves.

0151 15

David Picard @davidpicard.eurosky.social · 16/10/2026

Ben quoi, tu ne regardes pas si le dernier auteur est papa avant de te décider à citer le papier ? 🤨

130 3

David Picard @davidpicard.eurosky.social · 16/10/2026

You can clearly see the dip of population due to lack of birth during WWII slowly climbing up until it will eventually fade away 😉

110 1

David Picard @davidpicard.eurosky.social · 16/10/2026

They're incredible. It blows my mind how smart students are compared to what I remember I believe we all were during my time as a student.

072 7

David Picard @davidpicard.eurosky.social · 16/10/2026

I don't think so. On the contrary, I think it's extremely hard to improve recall with post training. I have an argument that revolves around representer theorems and the comparative size (and span) of post-training datasets compared to pretraining ones. It's also an empirical observation.

130 3

David Picard @davidpicard.eurosky.social · 15/10/2026

What I love with the joke is that, during covid, when the vaccine came out, we were all joking about how getting the shot did improve our abilities to connect to eduroam That was 6 years ago, and already at that time it was a variant of an old joke Nothing is eternal, except maybe eduroam problems?

2121 12

Reposted by David Picard

Gabriel Flynn @gabrielflynn.bsky.social · 14/10/2026

AI researchers warn technology 'months away' from being able to connect to wifi using eduroam

231221325 1.2K

David Picard @davidpicard.eurosky.social · 15/10/2026

Your toy is the one that I find the most appealing. It feels the most like analog stuff. But I have such a bizarre relationship with ink and paper that I cannot imagine taking notes without a fountain pen. Which means, I no longer take notes when I read papers because I don't print papers anymore 😬

020 2

Reposted by David Picard

SMBC Comics @smbccomics.bsky.social · 15/10/2026

Pretty pleased with panel 2. COMIC ◆ www.smbc-comics.com/comic/toad PATREON ◆ www.patreon.com/ZachWeinersm... STORE ◆ smbc-store.myshopify.com

Part 1/2 of a 5-panel SMBC comic. A fairy godmother appears to a princess, saying "I'm your fairy godmother, and I am going to make you a husband by turning that toad... Into a prince!" She zaps the toad, and the next panels shows a prince with a very vacant open-mouth stare. The princess, concerned, asks "Oh god, is the vacuous stare because he still has a toad brain?"

Part 2/2 of a 5-panel SMBC comic. The fairy godmother reassures her, saying "No, that's generations of royal inbreeding." The prince continues to look silly, by the princess and visibly relieved.

517225 172

David Picard @davidpicard.eurosky.social · 15/10/2026

WWII slowly disappearing from human statistics

110 1

David Picard @davidpicard.eurosky.social · 15/10/2026

To this day, French students still want a physical "poly" (a textbook accompanying the course made by the teacher, with varying degrees of polishing, for those unfamiliar with the term). I think there is something with possessing the physical object and the feeling of knowing.

140 4

David Picard @davidpicard.eurosky.social · 15/10/2026

This video is heartbreaking on so many levels 😭

031 3

David Picard @davidpicard.eurosky.social · 15/10/2026

Et donc tu l'as parrainé sur arxiv pour qu'il puisse déposer son preprint ?

010 1

Reposted by David Picard

Nicolas Audebert @nshaud.bsky.social · 15/10/2026

🎉 Accepted to #WACV2027, congrats Aimi and Georges!

021 2

Reposted by David Picard

Gautam Kamath @gautamkamath.com · 15/10/2026

argon doesn't trust me 😭

416810 168