Performance Hints

Jeff Dean argues that AI progress increasingly depends on the systems around models: specialized inference hardware, tools, context, and agent orchestration. He recommends finding bottlenecks with first-principles reasoning and napkin math, and choosing problems where general models still fail while checking whether those gaps are likely to last. Clear specifications, useful skills, evaluation, and search across multiple approaches can make agents more reliable on long tasks. He also encourages founders to build for problems they care about, question entrenched assumptions, and work with complementary, low-ego teams.