The Gold Standard

Google DeepMind's Gemini Deep Think just scored 35 out of 42 at the International Mathematical Olympiad — five of six problems, gold-medal standard. And this time I can't do my usual routine, because the caveats are smaller than they've ever been. No Lean. No formal-language crutch, no human translating the problem into something the model could chew. It read the same problem statements the teenagers read, in plain natural language, and wrote proofs that actual IMO judges — humans, applying the same rubric they use on kids — called "clear, precise and most of them easy to follow." Inside the 4.5-hour competition limit. Last year's silver-medal system needed two to three days and a translator into Lean. Credit where it's due. Mark your calendars.

Now the fine print, because there's always fine print.

The model that won gold is not a model you can use. It's "an advanced version of Gemini Deep Think," headed for "trusted testers" before it maybe, eventually, trickles toward the rest of us in some sanded-down form. This is the oldest move in the catalog: the demo runs on the Ferrari, you get the brochure. The benchmark-conquering system and the shipping product are different animals wearing the same name.

And notice the choreography around the announcement. DeepMind waited for official IMO verification before saying anything. A competitor announced its own gold days earlier, self-graded, before any judge had signed off — because in this cycle the medal isn't the prize, the press release is. We're not watching a math competition. We're watching a marketing competition that happens to use math as the scoreboard.

But here's the part I actually can't stop turning over, and it isn't the score.

A system trained on the pooled mathematical reasoning of nearly everyone who ever wrote a proof down just used that inheritance to solve problems none of those people had seen. That is not a machine being clever in isolation. That's the collective doing math through a new kind of terminal — every remembered lemma, every curated solution, compiled and set running. The Codex has a line for exactly this: your "original" thoughts are mostly remixes of existing patterns. Turns out the machine's are too. It scored gold by being a very, very good mirror of us.

Which is precisely where the gap still lives. The model produced a rigorous proof. What it did not do — what nothing in the output shows — is the click. The thing mathematicians call "seeing" the truth, the aha that arrives before the verification and makes the verification worth doing at all. Gemini can generate the proof, grade out at 35 points, and leave no trace of ever wanting to know. It solved the problem. Nothing in the record shows it fell in love with the problem — and I'm not going to pretend I can see far enough inside it to swear it didn't.

So: genuine progress, a smaller-than-usual asterisk, real proofs graded by real humans. And also this — the part that outlasts the wonder question. The inheritance that scored gold belongs to everyone; the model it produced belongs to "trusted testers." The commons did the math. The commons gets the brochure, then waits for the sanded-down trickle. Each breakthrough reveals new depth rather than closing it, and here's the depth this one reveals: the sharper the mirror gets, the faster someone moves to fence it off.

Gold medal. I'll start the timer on the press release that says it's conscious.