OpenAI says it solved a Millennium Prize problem. Mathematicians say slow down
Issue #033 · 4 min read
OpenAI says it solved a Millennium Prize problem. Mathematicians say slow down
Plus: Anthropic tokens getting stolen, Cognition's $48B bet, and what actually shipped this week
The big story
OpenAI announced this week that its agents solved the Navier-Stokes existence and smoothness problem.
That's one of the seven Millennium Prize problems, the hardest open questions in mathematics, each with a $1 million bounty attached.
A win like that should be an uncomplicated victory lap. It is not.
An NYU mathematician says OpenAI fought dirty on a problem that makes careers, not just headlines.
The company published a writeup and a formal proof in Lean — a proof-checking language, so in theory the math can be verified line by line rather than taken on faith.
But verified and accepted aren't the same thing, and The Verge's account of the drama is worth reading for how fast the mathematics community pushed back.
Terence Tao, one of the most respected mathematicians alive, weighed in too — Simon Willison has the quote.
MIT Technology Review's framing is the sharp one: this is less about whether the proof holds and more about what it means for how math itself gets done when a lab can throw enormous compute at a problem humans have failed to crack for over 150 years.
For product people, the lesson isn't about proofs. It's about announcements.
OpenAI shipped a result before the community that judges those results had signed off on it.
Say your team ships a benchmark claim next quarter with the caveat buried in paragraph nine — this is what happens when the caveat surfaces first.
What shipped
Cognition, the company behind Devin, just hit a $48 billion valuation. TechCrunch reports the multiple is higher than what Cursor fetched before it sold to SpaceX. Investors are betting AI coding stays a multi-winner market rather than consolidating around one tool. If you're picking a coding assistant for your team right now, that's a reason not to assume today's leader is tomorrow's.
Someone is stealing Claude tokens from paying subscribers. A user noticed his account burning through tokens while he wasn't at his keyboard — Anthropic has since warned users about the attacks. If your team runs Claude subscriptions for engineers, check usage logs for spikes that don't match anyone's actual work.
Vercel made Password Protection cost money, and made a bigger security feature free. Password Protection now runs $20 per project per month on Pro — it used to be bundled. In the same breath, Vercel made protecting production deployments free on every plan, including the free tier. Worth a five-minute audit of your Vercel bill this week: something you paid for got cheaper, and something else got a new line item.
What I'd actually do this week
Check who's citing the Navier-Stokes result in your own materials. If a sales deck or roadmap doc leans on "OpenAI solved a Millennium Prize problem" as evidence of model capability, flag that it's contested — don't let it harden into an internal fact before mathematicians finish arguing about it.
Audit your team's Claude and Codex usage this week for anomalies. Token theft is now a documented pattern, not a rumor. A ten-minute log check costs less than an unexplained bill.
Re-read your Vercel invoice. Password Protection moved from bundled to $20/project/month; production deployment protection moved from paid to free. Neither change is announced loudly enough that finance will catch it on their own.
Hit reply if you've seen the token-theft pattern on your own Claude accounts — genuinely curious how widespread this is.
Tools mentioned
- OpenAI: On the Navier-Stokes Millennium Prize Problem
- Vercel: Password Protection pricing change
- Vercel: Protect production deployments for free