Post by Théo Martin

🤔 AI doesn't democratize expertise. Not yet. Everyone says it levels the playing field. Juniors ship faster. PMs build products. The skill gap closes. But in production, right now, expertise still determines the outcome. The interesting part is that this is changing. When Donald Knuth's collaborator used Claude to crack an open combinatorics problem this year, it took 30+ guided iterations. Claude did the search. The human knew which directions were worth exploring. Without that domain knowledge steering the loop, Claude degraded and got stuck. The human was the reason it worked. I call this the summoning problem. The model is the search engine. You are the query. Knowing what to ask, in what order, with what constraints, is the entire skill. But that summoning problem is shrinking. 2 years ago, you needed deep domain knowledge to get useful agent output. Today, the unknown unknowns the model can't cover are fewer. Each quarter the floor rises. For complex production work, expertise still matters. A PM vibe-coded a system that hammered an endpoint 10,000 times per hour because they didn't understand the system they were building on top of. That still happens. But for simpler tasks, the model covers more ground on its own now. The window where deep domain knowledge is a primary multiplier is real. It's also closing. The expert with AI still covers 10x the ground of a novice. But that ratio was larger a year ago and will be smaller next year. The summoning problem is a snapshot of a narrowing gap. How much of your edge is durable, and how much is just a head start?