Sam Altman — Greylock: AI and the Next Computing Platform

Altman's Greylock talk (June 2023, ~60 minutes) was one of his most technically detailed public presentations — aimed at venture capital investors evaluating where to bet in the AI wave and covering the competitive moat question in more depth than most public venues. His key argument for why foundation models won't commoditize as fast as observers expect: the cost of training frontier models creates a natural oligopoly because only organizations with $100M+ training budgets can participate, and the number of such organizations worldwide is single digits. This creates durable competitive advantage at the frontier even if open-source models improve. Altman's investment thesis for the AI layer: applications built on AI — not AI infrastructure — will capture most value, because infrastructure tends toward commoditization while applications can build defensible user relationships. His advice to founders: don't build things that require AI to reach capabilities it doesn't have yet; build things that are 10x better than existing solutions using AI that already exists. He also discussed the context window as a capability threshold — the jump from 4K to 32K tokens enabled entirely new application categories — and predicted that 1M+ token contexts would enable "reading your entire life's context into a model," a use case he called personally compelling. The talk includes his most detailed public discussion of why he believes in-context learning and fine-tuning will eventually give way to fundamentally different interaction paradigms.