Studio SF

@KaiXCreator Why not? The spark, the strategy, and the orchestration still come from you. If you have the vision to direct an AI to ship something secure, scalable, and harmonious, you aren't just a founder, you’re an effective modern founder

Can you call yourself a founder if your entire product was built by Claude?

@_mohansolo Man, yesterday I said hello he said no thanks come back later 🤦‍♂️

We heard concerns that Antigravity consumes many tokens for simple tasks now. So, we're adding Gemini 3.5 Flash (Low) as a way to optimize token usage for these tasks. In our internal testing, it generates around 45% fewer tokens than Gemini 3.5 Flash (Medium) and generally outperforms Gemini 3 Flash (High) on SWE tasks. We've also gone ahead and reset Gemini quota across all paid plans to make sure you have all the tokens needed to build for the next week 🙂

excellente analyse, lucidité rare. Tu mets le doigt sur le vrai problème : on confond souvent un problème de talent avec un problème de logiciel mental. ​La France excelle à fabriquer des "moteurs" ultra-puissants (merci la rigueur des prépas et l'esprit analytique), mais elle leur injecte un carburant fait de peur du ridicule, de perfectionnisme paralysant et d'un rapport presque névrotique à l'argent et à la vente. ​L'expatriation aux US agit finalement comme un jailbreak : elle ne rend pas les Français plus intelligents, elle les autorise enfin à exécuter. Le tacle de fin sur le "syndrome des six mois de WeWork" est d'ailleurs ultra-juste... C’est la différence entre adopter une posture et vivre un vrai pivot culturel

bon bon bon comment dire (j’ai TROP de choses à dire que je ne sais même pas par où commencer mdr) déjà je crois que bcp de gens n'ont pas saisi la vraie lecture de ce graphique, on va revenir à des fondamentaux mathématiques c'est pas le nombre brut de founders français aux us c'est leur densité rapportée à la population, pour un million de français expatriés on a 45x plus de founders de licornes que pour un million d'américains de naissance, c'est une statistique énorme qui dit quelque chose de très précis sur la qualité brute du talent français quand on le sort de son cadre culturel d'origine ensuite sur le fond je vais donner mon opinion avec ma propre expérience des États-Unis car ce que ce chiffre révèle vraiment c'est que le système français produit des esprits exceptionnellement bien formés techniquement mais leur installe en parallèle un système d'exploitation cognitif qui les empêche d'exprimer ce potentiel sur leur propre sol, c’est un sujet dont j’ai longuement parlé ici à travers plusieurs threads pour présenter les différences culturelles entre les US et la France mais par ex il y a la peur viscerale de l'échec public, le besoin permanent de validation par les institutions,, le réflexe de cchercher une caution avant d'oser, le mépris culturel envers la vente et l'argent assumé,le réflexe de discuter une idée jusqu'à sa mort plutôt que de la tester sur le marché pour itérer… bref autant de bugs (la liste est encore + longue) profondément inscrits par + de 10 ans de prépa et de grandes écoles françaises qui ne se désactivent que par une vraie immersion prolongée dans un autre système culturel avec des gens venus de régions du monde pour qui la capacité à exécuter une idée prime sur tout le reste (clairement les asiatiques de la baie ) je pense les français sont éclatants à l'étranger exactement quand ils acceptent de désinstaller ce logiciel et de réinscrire dessus des réflexes radicalement différents, ship fast et itérer sur les retours du client, parler de revenue dès le jour un sans complexe bourgeois, recruter des gens meilleurs que soi sans craindre l'éclipse, vendre soi-même avant d'embaucher un commercial etc étc, ceux qui réussissent vraiment ce réapprentissage montent des empires comme hugging face, ceux qui refusent par orgueil de désapprendre rentrent au bout de deux ou trois ans rejoindre bpifrance ou un fonds parisien en se présentant comme ex-entrepreneurs de la baie sur leur linkedin alors qu'ils ont surtout fait six mois de wework à san francisco je m’arrête là même si j’ai encore tant de choses à dire sur le sujet mais le talent français brut est parmi les + puissants au monde mais c’est le logiciel COLLECTIF français qui le stérilise patiemment depuis quarante ans (et pas uniquement le politique, la fiscalité ou encore l’administration française)

Engineering is magic

@spolu The cost structure is real. What's also real: the founders who built critical infrastructure in France over the last decade didn't do it because the system was easy. They did it despite it. The question is whether the next generation has the same tolerance for friction.

Fuck off

@elonmusk Impressive ! The fault tolerance architecture is doing exactly what it was designed to do

Breathtakingly Beautiful

@_mohansolo The hardest part of building a useful coding environment isn't the intelligence layer. It's whether it works inside existing authentication, tooling, and review workflows without asking the team to change how they operate. That's the adoption constraint

Thanks for all of the Antigravity feedback over the last couple of days, especially around the IDE. Our intention was never to remove the IDE support for developers, and we should have been clearer with that in the product from the beginning. We’ve made it clearer in 2.0 on how to connect to the IDE, fixed issues with opening the IDE on Windows machines, provided instructions to restore IDE settings & extensions, and more. New releases for the Antigravity IDE and Antigravity 2.0 have rolled out with these changes. We should have done better so we’re going to reset everyone’s Gemini quota for the week again.

@sama The coordination problem inside large organizations. Not the intelligence gap, not the data gap. The structural inability to move decisions from the people who have the context to the people who have the authority. That's where the friction lives in every company

what problem do you most hope AI will solve in the future? maybe we can help!

@GoogleDeepMind Safely deploy AI at scale' means something very specific at the government level: auditability, data sovereignty, incident accountability. The interesting part of this partnership isn't the capability. It's which of those three constraints they solve first.

We’re expanding our partnership with Singapore to help safely deploy AI at scale. 🇸🇬 Together with country experts, our new programs will focus on accelerating scientific discovery, advancing pandemic preparedness, and improving healthcare. Find out more → goo.gle/49jGwjv

@GoogleDeepMind Create anything from anything' works in a demo environment. The question I'd ask: does it work in a context where the inputs are partial, the data is siloed, and the integration team had no say in the architecture? That's the real deployment test.

We’re dropping Gemini Omni: our first step towards a model that can create anything from anything - starting with video. It combines Gemini’s intelligence with our generative media systems - representing a leap forward in world understanding, multimodality, and editing 🧵

@GoogleDeepMind @GeminiApp @Google Material traceability was a legal requirement in my supply chain work. If a component failed, you needed to trace it back to the batch, the supplier, the process date. AI content provenance is the exact same problem, just in a different domain

SynthID, our imperceptible watermark for AI-generated content, is expanding to more partners. We’re also adding new ways to find out if content was generated using AI - just ask in the @GeminiApp or in @Google Search.

@AnthropicAI In manufacturing QC, you don't test only the finished product. You test every weld, every joint. AI-generated code has structural joints everywhere, and most of them are invisible to the developer who deployed it. Glasswing is the right framing

Last month we launched Project Glasswing, our collaborative AI cybersecurity initiative. Since then, we and our partners have found more than ten thousand high- or critical-severity vulnerabilities in essential software.

@karpathy HTML is a better output format for pipelines because it was designed for machine parsing, not human reading. Markdown was built for humans writing for humans. When you're building structured workflows, that's not a minor distinction

This works really well btw, at the end of your query ask your LLM to "structure your response as HTML", then view the generated file in your browser. I've also had some success asking the LLM to present its output as slideshows, etc. More generally, imo audio is the human-preferred input to AIs but vision (images/animations/video) is the preferred output from them. Around a ~third of our brains are a massively parallel processor dedicated to vision, it is the 10-lane superhighway of information into brain. As AI improves, I think we'll see a progression that takes advantage: 1) raw text (hard/effortful to read) 2) markdown (bold, italic, headings, tables, a bit easier on the eyes) <-- current default 3) HTML (still procedural with underlying code, but a lot more flexibility on the graphics, layout, even interactivity) <-- early but forming new good default ...4,5,6,... n) interactive neural videos/simulations Imo the extrapolation (though the technology doesn't exist just yet) ends in some kind of interactive videos generated directly by a diffusion neural net. Many open questions as to how exact/procedural "Software 1.0" artifacts (e.g. interactive simulations) may be woven together with neural artifacts (diffusion grids), but generally something in the direction of the recently viral x.com/zan2434/status… There are also improvements necessary and pending at the input. Audio nor text nor video alone are not enough, e.g. I feel a need to point/gesture to things on the screen, similar to all the things you would do with a person physically next to you and your computer screen. TLDR The input/output mind meld between humans and AIs is ongoing and there is a lot of work to do and significant progress to be made, way before jumping all the way into neuralink-esque BCIs and all that. For what's worth exploring at the current stage, hot tip try ask for HTML.

@karpathy The choice to go into pre-training rather than products or agents is the signal here. The people who understand what the base model learns, and why, will have the deepest leverage on what comes next. Good decision

Personal update: I've joined Anthropic. I think the next few years at the frontier of LLMs will be especially formative. I am very excited to join the team here and get back to R&D. I remain deeply passionate about education and plan to resume my work on it in time.

@fchollet Indeed. In manufacturing (my field), the first time an operator understands why a process drifts instead of just that it drifts, everything else becomes learnable. The first genuinely novel insight seems to unlock a meta-capacity. I've seen this consistently

If you can learn one thing that's genuinely novel to you, you can learn anything.

The personal AI expert for health/finance/education is compelling. The harder version of that problem is the same AI operating inside an organization's decision chain, with incomplete data, accountability constraints, and a compliance team. That's where 'beneficial' gets structural

three of the things we are most excited about: 1. AGI accelerating research 2. AGI accelerating companies 3. personal AGI accelerating everyone in achieving their goals today it was great to announce the unit distance result. yesterday it was great to announce that we are offering to invest $2M in openai credits into every YC company. now we need to increase our efforts on the third!

@emollick Early dismissal of industrial tools follows the same pattern: initial capability looks like a trick, real utility only shows at the second or third deployment cycle. The 'strawberry' arc is textbook.

Its funny how much the whole "strawberry" thing, which turned out to be o1-preview, was dismissed as overhyped at launch when it is clear in retrospect that it was way underhyped. A direct line from models unable to do basic math to solving unresolved math problems in 18 months.

@emollick The 'different' that matters in enterprise isn't the capability demo. It's whether it works inside an existing workflow with credentials it doesn't own, on data it wasn't trained on. That gap is where most deployments fail

I think people don't realize why Gemini Omni is different than other video AIs. It is fully multimodal, so it can edit video natively, too I took the famous "train " movie from 1896 & made it a bullet train, LEGO, added a time traveler, a centipede, muppets... (see reflections?)

@emollick Fact-checking is a production problem, not a benchmark problem. The gap shows up when you have live data dependencies, partial context, and time pressure. Does GPT-5.5 Pro hold up in those conditions, or only on structured test sets?

GPT-5.5 Pro is a very solid fact checker. I can throw entire chapters at it and it will hunt down every key reference accurately. The only real annoyance is that it loves nuance, so returns a lot of “the general idea is right, but you are not taking into account tiny detail X”

Google vient d’annoncer une architecture IA massive : 9 600 TPU par Superpod, 134 000 puces connectées, 47 pétabits/seconde de bande passante. Mais le vrai sujet n’est pas la puissance brute. C’est le scaling quasi-linéaire. En clair : ajouter plus de machines permet presque d’obtenir autant de performance en plus. Dans le monde réel, c’est extrêmement rare. Dans une usine, si vous doublez les machines, vous ne doublez pas forcément la production. Très vite, les bouchons arrivent : attente, coordination, qualité, logistique. Google essaie de faire l’inverse : une architecture où les puces travaillent ensemble comme un flux continu, sans embouteillage. C’est pour ça que cette annonce est importante. Ce n’est pas seulement une course au meilleur modèle. C’est le début d’une vraie industrialisation de l’IA. Demain, l’avantage ne viendra pas seulement de l’algorithme. Il viendra de ceux qui maîtrisent toute l’usine derrière.