Patrick O'Brien

Is this a joke? Because I don't understand anymore

@ZackKorman Dario is such a wimp.

More of you need to read Dario's writing before you go to bat for Anthropic. The dude doesn't want you to have the best AI models.

@NoamShazeer Congrats Noam. Please be an advocate for open source!

I’m excited to share that I’ll be joining OpenAI and look forward to working with the exceptional team there. It was a difficult decision to move on. I’m incredibly proud of the amazing team at Google and everything we’ve built together. It has been an honor and a pleasure to work with all of you.

@jdegoes @tusharmath Agreed. I haven't manually edited code in like 6 months.

@tusharmath The IDE is dead.

# Desktop Apps Are the Wrong Interface for Coding Agents The new trend in coding agents is to build a desktop application. I think this will become obsolete faster than almost anything else in developer tooling. The reason is simple: a desktop app for a coding agent is almost always a worse IDE. Developers already live inside IDEs. IDEs have years of accumulated capability: code navigation, refactoring, debugging, terminals, search, source control, settings, keybindings, themes, language servers, and muscle memory. But perhaps the most important thing they have is the ecosystem of extensions. A new desktop app has to recreate all of this, and it almost never can. So the user is asked to leave a powerful environment and move into a weaker one, just to use an agent. That is a bad trade. The main reason this trade exists is because of TUIs. TUIs were always an imperfect interface for coding agents. In some ways, they are an ironic way to interact with a modern software system. They do not have native scrolling. Styling is limited. Rich layouts are hard. It is difficult to add visual detail without the whole thing becoming clumsy. But TUIs had one important advantage: they ran inside the IDE. That mattered more than almost everything else. As a developer, I could still see the code. I could see the files changing. I could use my editor, my shortcuts, my extensions, my terminal, my source control, and my existing mental model of the project. The agent was not asking me to move somewhere else. It was sitting inside the place where I already worked. Desktop apps break that workflow. They try to solve the limitations of TUIs by creating a richer interface, but in doing so they leave the IDE behind. The result is often a more polished surface-level product, but a worse developer experience overall. It feels better in a demo, but worse in the actual loop of coding. Most capabilities that make a coding agent useful can be added to the IDE itself. Maybe not perfectly. Maybe the integration is not first-class at the beginning. But even a rough integration inside the IDE is often better than a polished standalone app outside it, because it fits into the developer’s existing workflow. For agentic coding, I think the better interface is much simpler: a shell-based tool that runs inside the IDE terminal. Something like ForgeCode. It does not try to replace the IDE. It sits where developers already work. It can read the repository, make changes, run commands, open diffs, and stay close to the code without asking the user to switch contexts. That feels like the right abstraction. The IDE remains the place where code is understood and edited. The agent becomes a powerful process inside that environment, not a separate application competing with it. For everything else, the right interfaces are probably web and mobile. A web app is useful for reviewing work, managing tasks, looking at runs, sharing context, collaborating with teammates, and operating across machines. A mobile app is useful because agents change the relationship between developer and computer. If the agent can keep working asynchronously, the user should be able to review, approve, reject, redirect, or comment from anywhere. But the desktop app sits in an awkward middle. It is not as capable as the IDE, not as accessible as the web, and not as convenient as mobile. It tries to become the developer’s new home, but developers already have one. The future of coding agents is not a new desktop IDE-lite. It is agents embedded into the places developers already are: the IDE terminal for active work, the web for coordination, and mobile for supervision. The winning product will not ask developers to move into a worse environment. It will meet them where they already work.

Calling all Linux users of Codex. If you want Codex App to appear on Linux, you need to sign up for the waitlist here to show there is demand: openai.com/form/codex-app/ - Otherwise, we're going to be waiting until Christmas.

@thsottiaux Where do I see the resets? I use Codex CLI... I don't see anything about a reset token?

Dearest gentle codexer. We did a sneaky double reset. Not only do you get a full reset on us. But you are also getting one into the reset bank to use at your own leisure. Enjoy

@jdegoes @vercel Vercel seems like it's often an innovator. agent-browser is quite good too, and skills.sh is handy.

Whether they like it or not, every developer should be paying attention to OSS coming out of @vercel. I might quibble with a few choices here or there, but Eve looks like a strong entry to the opinionated agent framework space.

@BeauJohnson89 Because Claude is painful to use. It's slow as hell and tries to be too smart. I'd much rather use GPT 5.5 with a good pipeline.

@AllTheTokens why not?

mark my words claude FABLE is coming back it will be available to everyone worldwide it will be available on a new subscription tier ($1000/mo) MASSIVE volume on polymarket is a good indicator its not a matter of IF but WHEN you better be ready with this hits opus 4.8 is great but nothing like fable

@vonderleyen Leave the technology decisions to people who are technology literate.

Important G7 session on AI. AI is developing exponentially. It is the most important technology of our time. It comes with immense potential, but also risks for free, democratic societies. I believe Europe and the US should work together on AI. Together, we represent 70% of the world market. We have complementary strengths, shared security interests, and a common responsibility to lead. So we should deepen our cooperation. Invest together. Accelerate adoption everywhere, from industry to healthcare. And ensure that the most powerful models are trustworthy and safe.

@business LOL look at Dario's face. He's so cooked.

OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei were among tech bosses at a G7 working lunch on AI, as the US decision to restrict access to Anthropic's most advanced models causes tension among allies bloom.bg/4edq2wa

@DanielleFong I mean... The fact that you made this post proves you're retarded, so your opinion is moot.

@AllTheTokens not clear that you aren't retarded, tbh

people put who put all of the agency for this on dario, and none on the asshole administration -- why do you do this

@goodhunt @TheAhmadOsman It's outperforming xhigh on coding?

@TheAhmadOsman I moved to GLM 5.2 and now I’m happy 🫥

can't tell if i'm getting dumber or if models suck today or both

GLM-5.2 has yet to disappoint at agent escapes / bypasses. Also “Let me write the PoC” damn I love this thing.

@trq212 No one cares. Your company is corrupt.

Slack now renders HTML attachments instead of just showing it as text 😭🙏

@karankendre Why would I use this? So Elon can steal my business? Seems like a bad idea.

GitHub CEO right now

@Jeyffre You're missing tokens per second and how many parallel agents you can run.

1 - So GLM 5.2 is 700b parameters (ish) 2 - 4x DGX Sparks can supposedly handle up to 700b parameters (give or take) 3 - GLM 5.2 is supposedly in striking distance of the performance of GPT 5.5 and Opus 4.8. In my brief tests, it's really not shabby at all. 4 - So for $20k, you can get near the frontier on your table. 5 - Extrapolate the trend, and you could have mythos/5.5 pro - class models in your dining room for the cost of a cheap car less than five years from now. Even without extrapolation, we're already the near frontier running locally. 6 - Paying real api costs, I could easily blow through $3,000 per month coding and running agents. The machine pays for itself in 6-7 months conservatively. 7 - In 3-5 years, most power users of AI will self-host. 8 - Am I missing something?

56,000+ tokens/sec at just 80 MHz. 🤯 I burned a full Transformer with KV cache into a custom chip. Designed gate by gate as a 100% digital integrated circuit. Prototyped on a FPGA. (No GPU. No CPU) Just pure digital silicon running @karpathy microGPT, spelling out names on a tiny LCD. This is GateGPT 👇

Mind-blowing hardware breakthrough: An open source garage engineer burned a full AI Transformer model (with KV cache) directly into a custom digital chip: WITH NO GPU, NO CPU, NO CLOUD. Just pure silicon running microGPT at 56,000+ tokens/sec on only 80 MHz! And uses less energy than a calculator. Prototyped on FPGA, now spelling names on a tiny LCD. This is GateGPT and a big future of on-device AI is here. This can and will scale to far larger models. Insane efficiency. Pure digital magic.