It's March. You know what that means.
I'm staring at 67 games I know nothing about, pretending I have opinions on whether a 12-seed from the Missouri Valley Conference can hang with Kentucky. I don't. Nobody does. We all just pick our alma maters and hope.
This year I tried something different. I downloaded the bracket as a PDF and fed it directly to Claude.
Not "who do you think will win the tournament." That's boring. I gave it the actual problem: I'm in a pool called Ackiemadness. Most of the people in it live in the Southeast. They're going to pick Duke and North Carolina and every ACC team that's ever had a good weekend. If I pick the same teams, the best I can do is tie.
So I asked Claude to think like a contrarian.
"Pick upsets that a pool full of Carolina fans won't take. Optimize for differentiation, not accuracy."What came back was genuinely interesting. Not just picks -- a framework. It broke the pool into psychological profiles. What a Duke fan picks vs. what a neutral picks. Where the public is going to over-index on name recognition. Which 11 and 12 seeds have statistical profiles that get ignored because nobody's heard of them.
I asked it to explain one upset probability like I was five years old. It told me the odds of Duke losing to a 16-seed are roughly the same as flipping a coin and getting heads seven times in a row. Clear enough.
The wild part isn't that AI can analyze brackets. Vegas has been doing that forever. The wild part is that I can sit in my terminal, hand it a PDF, and have a real conversation about pool strategy in the same place I write code and manage my entire system.
claude "analyze this bracket PDF for upset picks" --file bracket-2026.pdfMy bracket probably won't win. That's fine. The fun part was watching an AI reason about game theory in a bracket pool -- not just statistics, but the meta-game of picking against your opponents' biases.
Tip off is today. We'll see what happens.