
Living inside the AI exponential

Why AI is going vertical (again) | Dianne Penn (Anthropic)
Topic 3 of 18
Published Jul 26, 2026 · Clip 13:51–23:34 (9:43)
The conversation turns to the pace of AI progress: Dario's correct prediction that coding would be solved and the sense of now living inside the exponential, where each improvement is a massive jump. Diane emphasizes adaptability — using evals, safety testing, red teaming, and prototypes since you can't predict the exact model or moment — plus first-principles reasoning to pull plans forward when models surprise, citing Cowork, Skills, and Tag. She explains discontinuous emerging capabilities from the original scaling-law papers (models jumping from unable to reliably do 1+1 to able), which is also why safety is hard, and names 'product overhang and user overhang' even on current models. Reacting to Gary Tan's idea that spending $100K/year on tokens means living like 2028, she reframes token spend as input and experimentation as the output, adding that some of the best thinkers spend heavy time with every research model. She highlights Anthropic's 'working in public' Slack culture from the early days, where company-wide experimentation surfaced new use cases within about 10 requests, arguing experimentation is not an individual sport.