I'm a 2026 Kleiner Perkins Fellow studying artificial intelligence and computer science at Purdue.
The chart above is Rich Sutton's bitter lesson, the most expensive thing the field keeps relearning. Every hard problem offers the same two roads: teach the machine what we already know, or build something that can search and learn. The first wins the year and loses the decade. I'm interested in the second, and in the unglamorous plumbing that decides whether a model stays a demo or becomes a product. What does it take to make the second road actually work?