Next week I’m giving a talk and this article has some ideas I might use. Here’s the series — it’s open and online, come take a look!
https://cs.santarosa.edu/ai-talks-2026
A new kid joined my high school group last week. I think he was just escaping the larger group. When I asked him if he was here to stay, he asked what we were doing.
I said “AI.”
He said “I hate AI.”
I asked him why.
He mumbled about it drinking all the water. We started off with a quiz (Kahoot!) that covered how much of your phone’s magic is AI-powered (like unlocking it, your TikTok/Insta feed, predictive text, etc.) and how generative AI like ChatGPT is but one facet of AI.
We went on to explore cool programs other kids have made in Scratch (with and without AI) and he started having fun. He found one that played the piano as notes fell down like rain onto the keys. In Scratch you can remix any program to make it your own and that’s what he did.
Programming is creative. Creating is fun. Computers are fun. AI is fun.
It’s a polarizing time, so even when you don’t fully understand something, you’re expected to pick a side. Something else is happening and I think this is the biggie — a feeling that your future is outside your control. A new generation sees us handing over a mess to them and now AI is coming for all the jobs, ruining all the art and making us all stupid.
Push through the noise to find your agency.
Last year I introduced my class to The Millennium Prize Problems. They’re seven of the hardest and most important math problems. So far one has been solved. Each carries a million-dollar prize. Some, like the Traveling Salesman Problem, seem eminently solvable. But no, our biggest brains haven’t been able to solve the rest.
Last month, OpenAI claimed to have solved one.
It’s called Navier-Stokes1 and it describes the flow of fluids. As you might imagine, this set off a firestorm across the mathematical community.
Actually, there were two separate AI proofs. OpenAI announced the first one here. A rumor oozing out of Twitter suggested that OpenAI’s rival Anthropic was close to solving Navier-Stokes. Angst motivated OpenAI to unleash a 10,000-agent swarm on the problem. 88 hours later they had a solution. While the proof seems to be correct, it was done by exhaustively checking different approaches rather than deriving a reusable method.
Shocker — the Twitter rumor was incorrect.
It wasn’t Anthropic the company but rather one of their researchers who’d teamed up with an NYU math professor2 and they worked on Navier-Stokes for over a year. They took a tightly scoped and atypical approach to the problem, tweaking their methods many times using multiple tools and AI systems.3
Once they had a viable proof, it took them another month to work through it and write it up. They published their results here. These results are recognized by mathematicians as novel work and a bridge to new understandings in fluid mechanics.
Here’s how they characterized the contribution of AI:
We happily used both Claude and Codex to iterate on our proof …
The first writeup that we produced iterating with Claude was, in our opinion, the worst writeup we had ever seen in the history of mathematics … It was then our task to make a presentable and understandable writeup and we iterated over the weeks on the writeup with Claude and Codex ...
If you think of Navier-Stokes as a complicated combination lock, OpenAI’s agent swarm tried combination after combination until one opened it, while two mathematicians working with AI uncovered how the lock worked.
AI completed versus AI assisted. Both teams had AI but chose to use it differently.
This is not the first collision of new tools and old trades.
In 1852, university student Francis Guthrie made an interesting claim: any map can be colored with four colors so that no two bordering regions share a color.
This is a similar problem to Traveling Salesman as it seems easy enough to verify, yet the proof is elusive. The four color theorem went unsolved for over 100 years until 1976 when it was proved by a couple of math professors at the University of Illinois.4
They figured out 1,800 unique map configurations and had a computer check every single one, taking 1,200 hours of computer time. At the time, many mathematicians wouldn’t accept the proof because no human could check it by hand.
Ultimately it was accepted. AI déjà vu.
My current high school stint has taken a turn. Rather than taking a few kids on a coding sidequest, I’m now working with the teachers on a school-wide mini-golf challenge where the kids design an 18-hole course.
The microbit computers I talked about in this article a few weeks back will automate each hole.
Yesterday I sat with the same kid and we imagined a hole where you tee off towards an octave of piano keys and the ball plays the right note. I took a look at some of the other ideas these kids are coming up with — super creative and ambitious.
It’ll take all our collective brains and every tool in the shop (including AI) to pull them off.
It’s a problem in fluid mechanics: specifically if stable fluid equations can break down.
Tristan Buckmaster at NYU and Levent Alpöge at Anthropic.
Including Codex from OpenAI as well as internal Anthropic models.
Kenneth Appel and Wolfgang Haken.