Best of your X follows: May 31

Today's digest covers: OpenAI Robotics is hiring for the first time, Greg Brockman's hands-on Codex moment, Ethan Mollick on open-weights fragility, Greg Brockman on AI expanding scientific ambition, Paul Graham on CEOs who build, the OpenAI IPO filing, and CSU's bumpy ChatGPT rollout.

Sam Altman posted the clearest signal yet that OpenAI's robotics push is live, not just a slide deck. The division — built from the World Simulation research program led by Aditya Ramesh — is now publicly hiring across the hardware and ML stack: engineers who want to build "robots for skilled workers to build our future infrastructure." The long-term vision is personal robots. Applications go directly to robotics-recruiting@openai.com.

What changed: this is the first time OpenAI has named a robotics leader, named a program, and posted a public job channel in one announcement. It's a step beyond the occasional robot-demo tweet.

Greg Brockman posted from personal use rather than from a press release, which is worth something. "Viscerally compelling" is the kind of phrase people use when a product does something that surprises them even though they built it. Earlier this week he also noted Codex is useful for "any kind of work done with a computer," which is a broader claim than its original code-focused pitch.

Mollick pushed back on Epoch AI's benchmark data placing open-weights models 3–4 months behind frontier. His read: the benchmark gap understates real-world fragility. Open-weights models perform well on in-distribution tasks but break down faster when the query is out-of-distribution, and the existing evals don't capture this reliably. Whether or not you agree with his inference, it's a useful corrective to the "open-weights models are almost caught up" narrative that circulates every quarter.

Yesterday morning's tweet was quieter but pointed: "AI for accelerating research, by expanding what mathematicians and scientists dare attempt." The emphasis on dare attempt matters. The argument isn't just that AI makes existing research faster — it's that it changes the set of problems people consider tractable, which is a different and larger claim.

Last night's take: "The only thing worse than having the CEO knee-deep in building stuff with AI is not having the CEO knee-deep in building stuff with AI." It's a pithy inversion of the usual concern about executive meddling — and probably the most-shared Graham line this week.

OpenAI filed confidentially for an IPO targeting roughly $1 trillion. Per Reuters and multiple follow-on reports, Goldman Sachs and Morgan Stanley are on the deal, with a target public debut in September 2026. The filing comes as Anthropic hit $965B in valuation last week — so within roughly a month, both major frontier labs have either crossed or approached $1T. For anyone watching AI's capital structure: the era of private AI unicorns is closing.

California State University renewed its OpenAI contract — and students aren't happy. CSU agreed to $13M/year for three years (down from $17M) despite facing $144M in budget cuts. The survey numbers are striking: 65% of students are skeptical AI benefits education overall, 80% are uncomfortable submitting AI-generated work, and 40% of faculty either discourage or ban AI in class. The university described the original deal as a "huge branding opportunity" in planning documents. The contract renewal went through anyway.

One more for context: Simon Willison flagged a 3 — last 28 days of consumption-based sales × 13, plus monthly subscriptions × 12. Worth knowing before citing that number.