Pinned Tweet
Replying to @rasbt
Hard to judge without seeing it in full training, as itâs not the case here. Iâm a bit skeptical about how this could scale without a norm layer, and also without skip connections on networks much larger than 100 million parameters. We always come back to the basics transformer
Matteo Robino đ retweeted
Ca signale une forte ambition de déploiement massive des robotaxis de Tesla également en Europe. Il y a 10 ans les taxis bloquaient des routes contre les VTC, d'ici peu on verra les VTC et les taxis bloquer ensemble des routes contre les robotaxis.
Un million de voitures sans volant ni pédales : Tesla va investir 3 milliards d'euros dans son usine prÚs de Berlin pour produire ses taxis autonomes Cybercabs
l.bfmtv.com/qEd4
Matteo Robino đ retweeted
the main problem with current full duplex models is that almost everyone is iterating on the same moshi architecture which imo has some real bottlenecks.
in moshi every text token is locked 1:1 to an audio frame that gets spoken right away (apart from padding and special tokens). because of this rigid lockstep the native architecture cant generate a burst of text without advancing audio time. if the model generates 20 text tokens time has to move forward 20 frames so exactly 1.6 seconds of audio gets synthesized alongside. you cant pause the audio clock to let the LLM write out a block of text.
within base moshi you could in theory train it to output reasoning text or JSON tool calls while forcing the audio stream to output silence tokens. but since the clock cant be decoupled that "silent thinking" still runs at 12.5 tokens per second. a 50 token thought chain or tool call means the user sits through 4 solid seconds of awkward dead air.
the other option is synthetic data where it keeps talking while producing the JSON or the thinking. but there is no data in the wild for that so you generate a lot of it and now your data is the bottleneck. the voice will only ever be as natural as the most natural TTS you can use it will always carry the quirks of the harness you built to create it and if whoever is making that data cant think through the scenarios of dyadic conversations well, you are cooked.
so people came up with delegation system which is a pain in itself. you train the frontend model to be smart enough to hold a decent conversation and to hand off tool calling or thinking to a much larger, more capable model. this is engineering and its the only thing that has made the 2 or 3 full duplex models we have like openai gpt live possible at all.
thinking cant be done in speech. the signal to noise ratio is low and there isnt enough speech data to give the model a good enough world model in its weights. how much data is there of people doing maths out loud? even we humans dont think in speech. yes there is "thinking out loud" but speech is not the medium humans think in. thinking or reasoning needs a dense representation like text or maybe some latent but pure speech is a fuck all dream.
i saw dessa a few days ago. it seems good and sounds good but god knows how much data they trained it on and even it needs text, or some good engineering like ASR over its thinking speech, to actually interact with the real world or be even remotely useful. train on all the call centre data u want then ask it a basic question.
this also reminds me of a question from @juliarturc something like: if you converted every text available into speech and trained on that, what would happen? idk it might be smart or it might not. it depends on a lot of things like whether the speech encoder you use preserves everything needed for speech understanding. codecs are good enough these days but i believe they are still a bottleneck.
even if you solve that,context window is a problem and so is context degradation in speech compared to text LLMs current speech models have the memory of a goldfish.
there are just too many bottlenecks but at the rate we are going we might get there next year only if a lot of people come out in the open and open source the data they are sitting on and stop giving fake demos by training moshi on their privately acquired data.
Matteo Robino đ retweeted
Attention, quelles que soient vos opinions politiques par ailleurs, cette mesure de LFI est extrĂȘmement nocive pour la France en terme Ă©cologique, gĂ©opolitique, et Ă©conomique.
LâAllemagne a fait pareil et câest un fiasco, qui lui a coĂ»tĂ© trĂšs cher.
Elle sâest retrouvĂ© dĂ©pendante du gaz russe et du nuclĂ©aire français pour continuer son activitĂ©, et a mĂȘme dĂ» rouvrir des centrales Ă charbon ultra-polluantes.
Et ce malgrĂ© un investissement dans les renouvelables de 500 milliards dâeuros (que la France nâa clairement pas les moyens de sâoffrir !)
Et en terme gĂ©opolitique, lâAllemagne se retrouve dĂ©pendante du gaz russe et amĂ©ricain.
La France se retrouverait complĂštement soumise aussi Ă des pays hostiles comme la Russie, lâAlgĂ©rie, les USA, etc. sur son Ă©nergie.
La question de lâĂ©nergie est trop vitale et importante dans un espace mondialisĂ©, lĂ câest pas juste une question dâopinion politique, câest la survie du pays qui est en jeu : ne votez surtout pas LFI.
Jean-Luc Mélenchon balaie le programme EPR2 : la France insoumise veut sortir des énergies fossiles avant 2050 mais aussi du nucléaire pour passer au 100% renouvelables
l.bfmtv.com/s8Le
Matteo Robino đ retweeted
Yann LeCun and Randall Balestriero just Proved the Platonic Representation Hypothesis, and found a better way to achieve it.
The Platonic Representation Hypothesis suggests that AI models trained on different modalities, like images and text, eventually learn similar representations of the world. But this paper shows that how models are trained matters more than simply scaling them up.
So they mathematically show that, under certain conditions, two JEPAs trained completely separately can learn essentially the same underlying representation of the world. One learns from different images of the same scene, while the other learns from different descriptions of that scene, without ever being trained on image-text pairs.
Their approach achieves the same image-text alignment with 42x fewer training samples than next-token prediction. A 109M-parameter text JEPA even outperforms LLMs up to 72B parameters on cross-modal alignment benchmarks.
alphaxiv.org/abs/2610.39302
Matteo Robino đ retweeted
Today, we release the weights of a new and improved Le Chonk, advancing the open frontier for agentic coding and cybersecurity.
đ» Built to build: Le Chonk ranks #1 on @cognitionâs FrontierCode among open models, and sets a new benchmark for Pareto efficiency on @artificialanalysis.
đĄ Built to defend: State-of-the-art in cyberdefense, put through the paces by the world's leading cyber vendors, including @CrowdStrike, @cloudflare, @rubrik, @zscaler, and others.
Built for sovereignty: Make it yours. Download the weights on @huggingface , customize it with Mistral Forge, and deploy it wherever you need it.
Read more from our team and customers đ:
Lors de la journĂ©e de mobilisation nationale du 17 octobre, Pierre Collin, gĂ©rant du Micromania de Nice a Ă©tĂ© agressĂ© alors qu'il tentait de protĂ©ger sa boutique. La sĂ©quence est devenue virale. Nous l'avons rencontrĂ© Ă la sortie du commissariat de Nice oĂč il a portĂ© plainte.
Le gérant du Micromania agressé à Nice témoigne
Matteo Robino đ retweeted
Replying to @KenneyNL
This is bullshit, Kenney. Iâm the same tech tree following optimist who moved from C to C++, from 2D to 3D, and from closed source to open source. I run a company. I write code. I contribute to published papers. Iâve upped my expectations of AI. Up yours!
Matteo Robino đ retweeted
Some reflections on AI in math. I found it helpful to simply list out the various parts of what a mathematician does, why math is valuable, and where the funding comes from. From there, what happens if AI gets more capable?
youtu.be/LKrwhjOI8Nc
Je ne veux plus fréquenter les gens qui disent « à quoi ça sert ? ».
Jâai 50 piges sur mon passeport, 12 ans dans ma tĂȘte (et 18 ans dans mon slip, dirait ma femme).
Et je compte bien que ça dure.
Je ne supporte plus les pisse-vinaigre professionnels du scepticisme.
à quoi ça servait de vouloir voler ? De traverser les océans ? De marcher sur la Lune ?
Ă rien. Jusquâau jour oĂč quelquâun lâa fait.
Jâai toujours eu une tendresse infinie pour les originaux. Les inadaptĂ©s. Les rebelles. Les emmerdeurs. Ceux qui ne rentrent pas dans les cases.
Ceux qui ont des idées absurdes, des ambitions démesurées, des projets improbables.
Parce que ce sont eux qui font avancer le monde.
Pas les gestionnaires du déclin.
Les gens suffisamment fous pour croire quâils peuvent changer le monde sont prĂ©cisĂ©ment ceux qui finissent par le changer.
Et ce qui me terrifie aujourdâhui câest lâĂ©tat dâesprit que nous transmettons Ă nos enfants.
On les élÚve dans une atmosphÚre de fin du monde permanente. Entre malthusianisme, injonctions à la décroissance, culpabilité écologique et anxiété organisée.
Quelle terrible inversion des valeurs !
LâĂ©cologie devrait ĂȘtre un formidable dĂ©fi lancĂ© Ă lâintelligence humaine, pas une pĂ©dagogie de la rĂ©signation.
Je ne veux pas dâune gĂ©nĂ©ration qui sâexcuse dâexister au point de ne plus faire dâenfants
Je veux une gĂ©nĂ©ration qui brĂ»le dâenvie de dĂ©couvrir, de crĂ©er, de baiser, de comprendre !
lâoptimisme nâest pas de la naĂŻvetĂ© mais le courage de croire en notre capacitĂ© Ă tout affronter.
Le progrĂšs nâest pas une fatalitĂ© mais une conquĂȘte.
Câest mon ode aux rĂȘveurs, aux inventeurs, aux enthousiastes, aux insoumis, aux doux dingues et aux emmerdeurs magnifiques.
Ă tous ceux qui disent « tâes cap ? » plutĂŽt que « Ă quoi ça sert ? ».
Jâai 50 ans, et je refuse de devenir vieux dans ma tĂȘte.
Aux fous qui veulent changer le monde !
Matteo Robino đ retweeted
There has never been a better time to go into math: youtube.com/watch?v=LKrwhjOIâŠ
Matteo Robino đ retweeted
Should Europe leave frontier AI to the Americans and the Chinese?
The idea has gained ground in Europe. It is a fundamental mistake.
Build or surrender: for now, the choice is still ours.
Full article below.
Matteo Robino đ retweeted
"Ils ont annihilĂ© mon domaine de recherche": nouveau coup de massue dâOpenAI sur le monde des mathĂ©matiques
En rĂ©solvant des problĂšmes extrĂȘmement complexes par centaines, l'IA provoque les larmes d'Hugo Duminil-Copin, mĂ©daillĂ© Fields
lemonde.fr/sciences/article/âŠ
Matteo Robino đ retweeted
Nah, it's a nothingburger. The final statement is checked for semantic alignment (part of Deep Mind's Formal Statements), and sorry-free with standard axioms. They just missed step 3:
Step 1 (hard!): have your AI solve the problem and write a natural language paper.
Step 2 (super hard!!): have your AI formalize the paper, which obviously requires correcting any typos in the argument, including possibly errors in any references on which it relies.
Step 3 (easy): have your AI go back over your NL writeup, compare it to the formal, and fix the issues surfaced, making sure to also point out any corrections from references that were fixed.
The fact that an intermediate lemma uses m+4 regularity in NL but m+5 in the formal proof? Nothingburger.
Replying to @ElliotGlazer
IDK⊠I donât think this is a nothing burger for mathematicians though. I think if one verifies a faithful translation of the problem into lean, and provides a lean proof (with no sorries) and provides a prose proof, the implication is that the prose proof reflects the arguments in the lean proof (and vice-versa). What other purpose could it have?
I think the community hasnât landed on what a âproofâ is with the introduction of lean certifications, but I think itâs a meaningful problem if the prose proof and lean proofs donât closely parallel each other.
Matteo Robino đ retweeted
Replying to @KrzakalaF
Au contraire.
C'est une nouvelle Úre qui s'ouvre pour les mathématiques.
Une Ăšre oĂč la dĂ©monstration formelle est largement automatisĂ©e et oĂč l'accent sera reportĂ© sur le dĂ©veloppement de nouveaux concepts, nouvelles abstractions, nouvelles dĂ©finitions, et nouvelles conjectures.
L'invention du bateau a réduit l'importance de la nage, mais a permis la découverte de nouvelles terres.
Matteo Robino đ retweeted
We serve those as well! Sorry if using nuclear energy is a real blocker
Berlin-based search engine Ecosia, used by government services, is dropping its French AI partner, Mistral, its CEO told POLITICO.
The company will now bet on open-source models, including those from China.
vist.ly/5mh5y
Matteo Robino đ retweeted
Meet Mistral Large 4, aka Le Chonk.
âą 1T parameters, natively multimodal. 49B active.
It is the best open weights model from US or Europe on aggregated benchmarks.
âą State-of-the-art on critical workloads, including cyber defense, manufacturing and finance and it surpasses closed frontier models on visual grounding.
âą Forged in Europe end-to-end and is deployable from Europe via our own Mistral Cloud infrastructure.
âą Available to all via API today. Working with cybersecurity partners privately.
Open weights release end of October.
đ€ Made with AI
Matteo Robino đ retweeted
Fast LeWorldModel
arxiv.org/abs/2606.26217

























