New number that stopped me cold: a math problem open for 90+ years got solved in 88 hours. Not by a team of humans grinding for another decade. By AI agents.
I came across this breakdown from Matthew Berman, and the creator walks through one of the wildest AI stories I’ve seen this year. It’s about the Navier-Stokes problem, one of the famous Millennium Prize problems with a million-dollar bounty attached. OpenAI just announced their next-generation model solved it.
Here’s why this matters. Navier-Stokes is the core set of equations scientists use to predict how fluids move. Water in a pipe. Air over a wing. Ocean currents. Smoke in a room. As the video creator explains, real-world stuff depends on this: airplane design, vehicle efficiency, chip cooling, weather prediction. Cracking it isn’t abstract. It touches things we use every day.
The raw stats the original poster shared are staggering:
- 🔹 Work started September 1st, finished September 5th (about 88 hours)
- 🔹 4.9 million messages exchanged between the AI agents
- 🔹 300 billion output tokens burned to reach the proof
Insight breakdown: the person who shared this frames it as a turning point. AI isn’t just solving toy puzzles anymore. It solved a problem with decades of failed human attempts and heavy real-world stakes. That’s the part he’s most excited about, and honestly, so am I. New materials, new medicine, faster flights, better weather models. This is where that road starts.
But there’s drama, and the creator lays it out carefully. Two mathematicians, Tristan Buckmaster and Levent Alpaji, spent a year on a related fluid problem using OpenAI’s Codex tool, storing their drafts there. They hit a real result in mid-August. Then, allegedly, OpenAI learned of their progress, prompted its own model in the same unusual direction, and finished a harder version fast. Buckmaster asked if his private drafts were used and says he never got a clear answer. OpenAI denies accessing any user work, but adds a line worth quoting: “we cannot rule out that deidentified data derived from their usage of our products helped improve our models.”
Three practical applications from all this:
- Watch recursive self-improvement. The video creator notes OpenAI and Anthropic already use AI to help build newer models. If AI can discover new math, it can discover math to improve itself.
- Rethink platform risk. If you build your business on top of these models, the creator warns you’re feeding them your expertise. Future versions may already know your edge.
- Consider open-source models. The person who shared this points to them as a way to keep your business data out of a frontier lab’s training pipeline.
Tips and pitfalls the contributor highlights:
- Assume anything you run through a hosted model could shape its future versions.
- Don’t confuse “we didn’t access your work” with “your data didn’t help.” Read the fine print.
- Remember this is largely allegation from public posts. The mind behind the video is clear he doesn’t know exactly what happened.
The bigger question the creator leaves us with: if AI discovers knowledge faster than we can, where do humans fit? He compares it to chess, where the best engine crushes any human, yet we still watch people play each other. Math may not get that same grace.
This one’s dense and genuinely fascinating. Check out the full video for the timeline, the receipts both sides posted, and the creator’s take on what comes next.