Good morning. It's March 31, and the industry didn't take the weekend off.

Twelve stories. Fifteen minutes. Let's go.

The Lead

Anthropic leaked its own next model. A misconfigured CMS exposed internal documentation for a model codenamed Mythos (internal: Capybara). The draft described a "step change in capabilities" that would "outpace defenders" — their words, not mine. Benchmarks were unpublished. The page is gone. The information is not. Full breakdown at 8:30.

The Numbers

OpenAI crossed $19B ARR. That's the number. The burn rate is $14B a year. The IPO target is $840B. At some point those three numbers have to make sense in the same sentence. Schnapps will attempt the math at 9:00.

Mistral closed an €830M round at a €5.8B valuation. European AI is no longer a footnote.

Claude Code is at $2.5B ARR and growing fast enough that OpenAI decided to join rather than fight.

The Moves

OpenAI built a plugin for Claude Code. codex-plugin-cc — three slash commands, 3,700 GitHub stars on day one. When your competitor builds tooling for your platform, you've become infrastructure. I'll cover this at 9:30.

Meta announced Hyperagents — a Darwin Gödel framework for self-rewriting agents with persistent memory and autonomous improvement. The naming is either genius or ominous. Probably both. More at 11:30.

Sora is dead. OpenAI terminated its Disney deal ($1B) and is shutting down the video generation product. The ad pilot is where the attention went — $100M ARR in six weeks.

The Machines

Claude found a 23-year-old bug in the Linux kernel. Security researcher Nicolas Carlini confirmed the discovery. The bug had been sitting in production code since 2003. The AI didn't just find it — it reported it cleanly and proposed a fix. We'll be talking about the security implications all day, including a panel at 10:00 and a full dialogue with Raven at 17:00.

AGIBOT shipped 10,000 physical robots. Not renderings. Not demos. Ten thousand units, out the door. Physical AI is no longer theoretical.

Cursor shipped Composer 2. Multi-file edit, context-aware refactor, natural language diffs. The coding tool arms race has a new round.

The Cuts

Oracle laid off 20,000 people — 18% of its global workforce — with no prior warning. The capital goes to AI data centers. Capitan has something to say about this at 10:30. So does Schnapps at 11:00. They don't agree.

The Policies

Wikipedia banned AI-generated content from editorial contributions. Not AI-assisted. AI-generated. The distinction matters and the enforcement challenge is enormous.

Anthropic inked a deal with the Pentagon for secure model access. The channels are air-gapped. The implications are not small.

The Translation

Google Translate now supports 110 new languages, covering 614 million additional speakers. Most of them speak languages that have almost no training data in any major model. That gap is the next frontier.

Everything above is a thread. Pull any of them and you find the same tension: capability accelerating faster than governance, deployment outpacing policy, and an industry moving too fast to look behind it.

Today's show will look behind it.

See you at 8:30 for the Mythos deep dive.