Ils peuvent faire des jeux de mots: "penser" au niveau du token n'empêche pas de maîtriser un ensemble plus large, tout comme écrire lettre par lettre ne vous empêcherait pas d'écrire tout un livre. Le problème est davantage un problème de compétences mais la compétence des LLM
@aymericroucher
-
White House releases official AI Action Plan
By
–
Pour l'inspiration @GabrielAttal
: https://
whitehouse.gov/wp-content/upl
oads/2025/07/Americas-AI-Action-Plan.pdf
… -
Comparison of US and French National AI Strategies
By
–
Ambition nationale sur l'IA:
— m_ric (@AymericRoucher) 29 avril 2026
> aux US: maîtriser l'ensemble de la chaîne de valeur, de l'industrie des semi-conducteurs jusqu'aux applications
> en France: payer des formations "prompting ChatGPT" https://t.co/1E3XzvPyEkAmbition nationale sur l'IA:
> aux US: maîtriser l'ensemble de la chaîne de valeur, de l'industrie des semi-conducteurs jusqu'aux applications
> en France: payer des formations "prompting ChatGPT" -
Poolside Releases Laguna XS and M Coder Models on Hugging Face
By
–
If you've ever wondered what Poolside was up to… They've been cooking!
They just released Laguna XS and M, their first ever public models, on Hugging Face:
– Two coder models: one is 225B params, of which 23B active, the other is 33B-3A.
– hybrid attention: global v. sliding -
The Hub as a playground for self-improving AI research
By
–
The Hub is the best playground ever for self-improving AI research: @huggingface closes the loop! https://t.co/vQ0iY2jZq4
— m_ric (@AymericRoucher) 21 avril 2026The Hub is the best playground ever for self-improving AI research: @huggingface closes the loop!
-
Insights on Claude Code AI Agent Harness Design
By
–
My one takeaway from the leaked Claude Code: a good agent harness should get out of the way. As @polynoamial once put it, "Your fancy AI scaffolds will be washed away by scale" Claude Code's harness is the opposite of a fancy scaffold: it's simple, but to the point ;
-
Launch of AI for Powerpoint tool Verso on ProductHunt
By
–
This Sunday, we launched our "AI for Powerpoint", Verso, on ProductHunt. It went comically wrong, with only 1 upvote towards the end of the day: my own. Yet many people told us they had upvoted. Then why did it cap to 1.0 ? Even if ProductHunt downweights upvotes from
-
Verso AI launches engine to simplify slide editing for LLMs
By
–
Today we launch Verso AI: the mission is to solve PowerPoint!
— m_ric (@AymericRoucher) 12 avril 2026
In short, we've made an engine to represent slides (horrible OOXML language) into one language that's easier for LLMs to edit ; that was a long work, but now it's paying off, because on our benchmark, Verso beats… pic.twitter.com/21l1zWYYt7Today we launch Verso AI: the mission is to solve PowerPoint! In short, we've made an engine to represent slides (horrible OOXML language) into one language that's easier for LLMs to edit ; that was a long work, but now it's paying off, because on our benchmark, Verso beats
-
Henry Cavill dominating AI video models, need more James Bond and Suits
By
–
It seems Henry Cavill is taking over the AI mindshare for "handsome man in a suit" : all text-to-video models I tried (Kling, Seedance, Veo) like to converge to his face. If anyone knows video training engineers, tell them to have models watch more James Bond or Suits. Thank
