Not an expert yet. Just someone genuinely obsessed with LLMs.

Paris, France
Joined March 2026
Enzo Comas retweeted
DeepSeek just published the price of a token, and it should make you angry. From their own DeepInfer docs: 1 cluster serves 155 trillion tokens a day, at $38 of revenue per GPU-hour, and thats at off-peak prices. Their API is $0.30 per million input tokens and $1.20 per million output at peak. Yuchen Jin did the math: even if they rented B300s at $7 an hour, thats about 82% gross margin. Other estimates go higher. So cheap AI is not charity and it is not "dumping". At these prices serving a frontier-class model is a very profitable business, if your inference stack is good enough. Now remember every time a US lab told you that intelligence is expensive. That $200 a month is a fair price. That limits exist because compute is scarce. Part of that is true. Part of it is a business model built on you not knowing what a token actually costs. DeepSeek just showed everyone the numbers and gave away the engine. Every inference company can now compete on price with the same software. Thats how AI gets cheap for everyone.
“A single cluster currently serves 155 trillion tokens/day, generating $38 in revenue per GPU-hour (at off-peak pricing).” Huge flex from DeepSeek. Not only did they open-source their inference engine, but they also open-sourced their revenue numbers. They didn't disclose which GPUs they're using here, but DeepInfer supports both NVIDIA Blackwell and Huawei Ascend. Even assuming they're renting B300s at $7 per GPU-hour, that's ~82% gross margin.
159
470
59
5,600
288,152
Enzo Comas retweeted
I'm an American bank guy who goes to rural France 4-5x/year. My academic background is European History. The conservative forces of rural France have crushed urban liberalism many times since the Revolution (1848, 1871, 1959 and one could argue Vichy in 1940 as well). That's why they're on the 5th Republic. It's setting up for that again next year. Background: France is run by highly educated mandarins for the benefit of the retiree class. There is little economic growth or job opportunity for the young because taxes and regulation are so high. President Macron's mild attempts to reverse this have been very unpopular with the retiree class. The school funding protest is really about lack of opportunity post-graduation, which is valid. The Stalinist LFI and the essentially unemployed immigrant youth are exploiting the government's unwillingness to unleash its experienced and violent but not numerous riot police on the students. Near-term, if the students stay home tonight, the police can start breaking heads with less fear of repercussions. If they don't, the govt is waiting for a spectacular outrage to justify it. Long-term, the next Presidential election is in the spring. One should assume that National Front will win by a lot with dire consequences for the immigrant community. Financially, France's social spending and still very capable (and conservative) military can not be maintained, yet the National Front wants to increase them. A financial crisis and EU exit can't be ruled out. Nor can a transition to a new form of government. TLDR - we haven't yet seen the guys that Gerard Depardieu used to play in cop and detective movies. They about to come on to the stage.
188
232
40
4,423
422,281
Pretty funny to see everyone blame this (rightly condemned, to be clear) on whatever their political bogeyman is: I've just seen people say that this is because of, take your pick, "communism," "Islamism," "third worldism," "Russia," and "China". As if France needed any outside help to riot... Maybe some people need to brush up on our history, or read Victor Hugo's "Les Misérables"... This isn't to romanticize the rioters in the least. A reminder that in Les Misérables there's the figure of Thénardier, a horrendous scoundrel who exploits the chaos around him, robbing corpses and fleecing anyone he can. Every chaotic moment in French history has its share of Thénardiers, and this one is no exception. It's just too easy, and actually useful to those in power, to blame this on outside forces. At the end of the day, this anger was made in France, is happening in France, and can only be solved in France. And, as painful as it is to hear, if one studies history, this is pretty much exactly what's expected at the stage France is at. People have studied this. Whether you take Ray Dalio’s six-stage cycle theory, Peter Turchin’s "elite overproduction" thesis or the Chinese dynastic cycle framework, they all come to the same conclusion: when debt piles up, public services crumble and those in power become incapable, the streets eventually explode. In fact it's almost eerie how France ticks almost every box of most cycle theories: - Fiscal distress (debt is 119% of GDP and we now pay more to borrow than Italy or Greece), - Crumbling public services: a reminder that these protests (now turning into riots) started with high school students protesting over the state of education - Political paralysis and a legitimacy crisis (Macron has an insane 82% disapproval rate, he's definitely lost the Mandate of Heaven...) - Popular immiseration with a record poverty rate at 15.4% and a median wage at only €2,190 net per month for a full-time private sector employee (you can't even afford to rent a decent flat in any major French city with this) "Third worldism" or "Islamism" isn't the cause of any of this: French governments didn't borrow like there's no tomorrow (the last balanced budget dates back to 1974!) because the "Muslim Brotherhood" told them to, and it wasn’t Putin or Xi who left French classrooms without teachers... Sadly the reactions - scapegoating - are ALSO typical of this stage of the cycle: Dalio lists among the classic markers of Stage 5 the emergence of "scapegoat classes who are commonly believed to be the source of the problems" (linkedin.com/pulse/delving-s…), alongside a "loss of truth in the public domain," with people more interested in hunting for culprits than looking for solutions. So what is the solution? Well, the way the Chinese dealt with it was basically a reset - regime-changing themselves with a new dynasty claiming the Mandate of Heaven, usually after decades of rebellion and war. Not exactly a hopeful prospect... There is also the exception where a dynasty managed what the Chinese call a 中兴 (zhōngxīng, "restoration"): profound reform from within before the whole house came down. It succeeded a few times (e.g. the Tongzhi restoration during the Qing dynasty) but usually could only buy a few more decades, rarely more. In 1945, the writer Huang Yanpei famously asked Mao how to break free of the cycle (cn.vlex.com/vid/constitution…), framing it in those terms: "People are careful and hard-working at first, for bold measures are the only possible ways to break through difficult conditions. But when things gradually get better, people begin slacking off: laziness may spread through a long history and is impossible to be corrected when it becomes a common practice. Gradually expanded territory requires corresponding increases in talent, but when the latter is insufficient, the central control will become weak. In a word, no one can escape the 'vicious cycle' of dynastic rise and collapse." Mao replied that democracy was the answer, because "as long as the people can supervise the government, the government dare not slacken." And Xi later added a "second answer" (chinadailyhk.com/hk/article/…): a system capable of "resolute self-reform". The jury is still out on whether China has broken the cycle, but if France has any hope of escaping it, it will need both answers at once: a government that truly answers to its people, and very resolute self-reform. One thing the answer definitely isn’t: turning on each other. In all these cycle theories, division and scapegoating appear on the way down, never on the way back up.
⚠️🔴 À REGARDER JUSQU'À LA FIN JE NE PEUX PAS DÉCRIRE CE QUE JE VIENS DE VIVRE À MARSEILLE. #17octobre
244
1,116
221
4,391
337,057
This is very true. Here is Qwen 3.8 27B compared to all the sota models 1 year ago today on artificial analysis. If 27B released a year ago, it would be the best model in the world by a wide margin.
The biggest takeaway from this is a 27b model that runs locally is 35% more intelligent than the leading SOTA model 1 year ago. Imagine what will run locally next year.
54
127
52
2,223
238,996
Enzo Comas retweeted
to me what's beautiful is that rea has gotten people asking: 'what can i do with ai?' instead of 'what can ai do for me' it's only a subtle difference in framing, but the future i see is beautiful
10
9
1
121
5,769
Enzo Comas retweeted
BREAKING NEWS🚨Mistral has caught up to frontier AI models from April 2025 It's time for European AI excellence
34
20
4
739
25,692
Enzo Comas retweeted
Met a kid today who taught himself AI with LLMs at 14 while barely speaking English, worked a 1yr long internship in a top AI company, flew across the world to attend a top 10 US undergrad, dropped out in 2 weeks, and has 5 ~$1M+ job offers from the top neolabs in SF at age 17. Traditional education is so dead. The young kids today are just built different.
249
338
55
5,678
367,625
Enzo Comas retweeted
Vibecoding with no understanding and learning just in time is the default for all major codebases School brainwashed you into thinking you need to learn something first completely before doing it. No, you just do it and learn along the way. If you get stuck learn more. Everything is figureoutable. I like the term code archaeology but I prefer exploring the code labyrinth. You never know what you will find
it is very hard for humans to fully understand all the code even on human written codebases. I’ll share a real example from a codebase i helped maintain: React no one, even the OG team members can fit the whole thing in their head. we have always relied on code archaeology and imperfect/JIT understanding to deal with complex software
40
86
15
1,358
41,065
You don’t solve poverty by teaching everyone to become millionaires. You solve poverty by building an economy where ordinary work can produce a decent life.
287
23,448
364
108,148
1,159,015
Ok les gars, écoutez-moi bien. Après deux semaines à San Francisco, je suis arrivé à une autre conclusion : c’est qu’on a un niveau tech de fou en France, en vrai. J’en discutais hier avec un autre ingénieur de très haut niveau installé ici, et il m’a expliqué que 80 % des gens qui font de la tech ici ont un niveau éclaté au sol et utilise l’IA de la pire des facons. Les 20 % restants, ce sont les gens des labs, des grandes boîtes, les indies qui buildent, revendent et recommencent en boucle. Et l’IA a accentué de fou cette différence de niveau. Et pour la première fois, je me dis que la bulle IA existe peut-être, mais elle n’est pas économique. C’est une bulle de possibilités et de connaissances. C’est surtout du Dunning-Kruger, quoi. Soyons clair, je veux pas être médisant et dire que les gens qui utilise de l’IA pour coder ont tort; c’est ok d’utiliser l’IA à foison, je le fais moi le premier mais quand on prend de la hauteur et qu’on regarde l’écosystème , le bruit de bas niveau ajouté est ultra méga violent. En fait dit plus clairement c’est que les gens pas tech qui code des produits avec l’IA ont direct un « effet wow » (et tant mieux) mais ils pensent que vu que pour eux c’est extraordinaire , il vont pas aller chercher plus loin, iterer , construire, répondre à un vrai problème qui va emmener de vrai user service.. je sais pas si je suis clair 😂😂😂. Je m’exprime trop mal je pense. Ici les événements publics sont remplis à 99 % de gens qui ont découvert ce qu’ils pouvaient faire avec l’IA et qui pondent tous les mêmes wrappers d’agents pour X, pour Y, pour Z. Et c’est ok, mais il y a pas de truc vraiment innovant ou wow. Je suis assez déçu sur ça en vrai. Là, je discutais avec un VC de Sunnyvale. Il m’expliquait que les boîtes adorent récupérer les idées de toutes ces « néo-startups jetables » et les refaire, car ça leur fait des tests de backlog produit gratuits. Et surtout qu’aujourd’hui, le truc qui compte, c’est l’expertise propre des gens dans une niche, et leur capacité à l’appliquer vite et bien pour résoudre un problème spécifique. Et que oui, avec l’IA, ça va plus vite, mais ça reste une loterie. Bref, tout ça pour dire qu’en France, les seuls trucs qui nous manquent, c’est : - Des VC qui arrêtent de financer les gens qui leur ressemblent (i.e. 90 % de fils à papa de HEC) et qui attendent juste le tour suivant pour sortir comme des piranhas. - Des événements de communauté tech, bien tech. Mais ça, ça demande de changer la mentalité des tech bros en France, qui aiment trop écraser les autres ou les gens qui apprennent, avec un sentiment de supériorité (je le sais, j’étais comme ça aussi). Et juste accepter de créer un environnement d’émulation. Et surtout, si chaque tech bro passait autant de temps à proposer et construire des choses qu’à critiquer, on serait top tier. Parce que c’est vraiment ça le truc de fou à SF. Bref, un changement par le haut et par le bas, et je vous jure, on éclate la tech US easy, vu les profils. Désolé, c’est décousu, j’écris ça devant mon café dans un diner bien Americaaaannn donuts donuts ! 🍩
64
59
12
800
98,748
Enzo Comas retweeted
Hugo Duminil-Copin (médaille Fields 2022) : la résolution de 350 problèmes majeurs par OpenAI, c'est comme si des camions m'avaient écrasé ; tous les problèmes (et donc pistes de recherche) que je mentionnais dans mes exposés, articles et demandes de financement, ont été résolus.
Astra au sujet de ces résultats : "Voici les annonces qui me paraissent les plus susceptibles de compter dans l’histoire des mathématiques" ; "Une seule de plusieurs de ces annonces, confirmée intégralement et réellement nouvelle, pourrait déjà constituer un événement majeur".
60
352
80
1,897
427,495
Enzo Comas retweeted
After six years in stealth, Atomic Machines is out. Our mission: on-demand, universal command of matter. First beachhead: the Matter Compiler, an AI-native manufacturing system that builds working micro-machines from code alone. No per-product tooling. No process development. Different code, different machine.
786
2,264
918
18,252
3,028,500
Enzo Comas retweeted
trying to imagine explaining to myself 10 years ago that we made a Markov chain so long that the pope is mad about it
17
147
6
2,570
58,578
Enzo Comas retweeted
When you “hit a wall” in something you are trying to learn, it’s typically just a massive debt of unlearned prerequisites that are finally being called due.
17
418
29
3,094
208,912
Enzo Comas retweeted
something to think about
96
7,308
349
52,632
1,031,246
Enzo Comas retweeted
Steve Jobs favorite quote
The way Steve taught us to care about every detail, every experience, every person, still guides our work today. Forever grateful.
30
946
23
11,077
268,561
Ok, donc c’est moins bien que les modèles chinois sortis il y a plusieurs mois, et Mistral qui s’arrange encore, comme à son habitude, avec les benchmarks et la manière de les présenter. La distillation n’aura pas suffi. Ils vont se faire exploser dans 2 semaines. C’est bien, hein, de release des modèles, mais bon. Faut juste regarder si ils ont réussi à faire sauter les safeguard des modèles chinois qu’ils ont utilisés, ça serait éventuellement le seul avantage. Mais sinon circulez, il n’y a rien à voir. Mais tous les maxeurs vont y aller de leur « wow ». Ok, ok. GL, HF.
62
17
9
340
68,964
Enzo Comas retweeted
If money wasn’t a problem, I’d probably spend my whole life studying. A master’s in philosophy, then astrophysics, physics, ancient history, psychology… Whatever makes me understand this weird existence a little better. Not for a career, not for a degree. Just because there’s so much to know and one lifetime already feels too short.
1,165
2,381
313
15,151
290,708
Enzo Comas retweeted
My mentor taught me that geo arbitrage will always be the thing that makes money You just copy whatever is making money in the west, make it a little bit better and more local, and sell it in other countries Me and a friend grew to $350k mrr making a malaysian openclaw alternative and we pretty much just stole the whole playbook There are literally endless opportunities like this
81
249
11
5,104
230,887
Enzo Comas retweeted
🚨 CODEX Users!! We will get free codex resets everyday For the next 28 days If OpenAI don't ship one meaningful improvement every single day for Codex + ChatGPT Work So basically: 📈 Improvement drops = W 🔄 No improvement = FREE RESET 🔥 28 days of Codex updates OpenAI just turned the next month into a Codex upgrade marathon What would you like openAI To ship?
Over the next 28 days, each day we’ll either ship one thing that is a clear improvement and relevant for most codex/work users or ship a full reset. Let the improvements begin.
223
227
84
4,170
782,986