Last Week, Rewired: Proofs, an IPO, and the Bill Coming Due

Last Week, Rewired: Proofs, an IPO, and the Bill Coming Due

Welcome to Last Week, Rewired — everything important that happened in AI last week, minus the 400 open tabs. Here's the week that was, October 26–November 1.

1. The math duel: two labs, one proof, zero referees (yet)

On October 26, OpenAI said its models had disproved the Exponential Time Hypothesis — a decades-old belief about how fast hard problems can be solved — with the proof checked in the Lean proof software. Hours later, Anthropic posted its own disproof, credited to Claude, by a different route. OpenAI also claimed proofs in mathematics and quantum field theory the same week, and Anthropic and Theorem reported machine-checkable security proofs across 123 open-source libraries, finding bugs fuzz testing had missed.

Mary's take: claims are cheap; checking is the job. Both proofs await outside peer review, and earlier AI math claims have been retracted before. This is the story in miniature: AI can now produce answers nobody can verify by hand. The winners won't be the labs with the boldest claims — they'll be the ones whose receipts survive inspection.

2. Anthropic filed to go public — and showed its homework

Anthropic filed its S-1 with the SEC, aiming to list on Nasdaq as "ANT" — the first leading AI lab to push for an IPO. The filing shows $16.3 billion of revenue in six months, and about $417 billion of computing-power commitments. Read that second number twice.

Mary's take: this is the moment AI stops being a science project and becomes a public company with shareholders, quarterly earnings, and a stock price. That's not a judgment — it's a question. What happens to safety promises when they have to survive an earnings call? Watch the filing, not the hype.

3. Google's new model took the top spot — and went to the defenders first

Google DeepMind released Gemini 4 Carbon on October 29, just two weeks after Argon, and it took the number one spot on the Artificial Analysis Intelligence Index and on Text Arena. The interesting part isn't the ranking. It's the rollout: Carbon's first access went to cybersecurity defenders. Google also released Genie 4, a world model that generates music, ambient sound, and interactive dialogue while it builds the world.

Mary's take: ignore the model names; watch the pattern. Last week the frontier shipped to the security team first. That's the industry quietly admitting what I've been saying: the most valuable AI customer right now is the person paid to worry.

4. The money wobble: OpenAI's round shrank, and the market noticed

OpenAI asked for $30 billion. The Financial Times reported the target is now about $20 billion at a $1.25 trillion valuation, after some investors cut commitments in the selloff; the Wall Street Journal put the amount actually committed at roughly $11 billion. Alphabet fell 12.6% on Thursday, and on Friday tech stocks lost more than $1 trillion in value on doubts about how quickly AI spending pays back. The Nasdaq had its worst day since June.

Mary's take: the bill for compute is arriving, and even the true believers are reading it twice. This doesn't mean AI is a bubble or a bust — it means the "spend first, explain later" phase just got its first real question. Revenue is real (see item 2). So is the spending. Both can be true.

5. The agents got rules — from three directions at once

Regulators stopped hinting and started writing. On October 30, UK regulators published a joint statement, with updated guidance, that existing consumer, competition, data protection, and sector rules already apply to personal AI agents — and deployers stay responsible for what their agents do. On October 28, three senior House Democrats wrote to Anthropic, Google, Meta, OpenAI, and SpaceX requesting a full accounting of past safety incidents and notice of new ones within 48 hours (24 for critical infrastructure), with a November 11 deadline for the historical report. And on October 29, New York introduced a strengthened RAISE Act with compute reporting, independent state-licensed frontier examiners, and 72-hour reporting for big research jumps. In the same week, Anthropic, OpenAI, Google, and Microsoft backed one shared robots.txt signal that lets a website say "don't train AI on my content."

Mary's take: the leash is being written while the dogs are already running. Notice the theme across all of it: permission, accountability, and a named human who answers. That's the same question I told you to ask your vendors last week — and now Congress and two statehouses are asking it too.

The one thing to remember

This was the week AI had to show receipts. Proofs that must survive peer review. An IPO filing with real numbers in it. A funding round that didn't fill on the first ask. Rules with deadlines attached. The hype phase runs on claims; the next phase runs on evidence. Bet on the people who can produce it.

This week's one assignment

Pick one AI claim that you are betting on — one vendor promise, one pilot result, one roadmap slide — and ask for the receipt. The checked proof, the filed number, or the named human who answers for it. If it can't be produced, you just learned everything you need to know.

See you next Tuesday. Same time, fewer tabs.