
Smooth Progress, Emergent Capability Jumps

A mind-expanding conversation on how models actually work | Emmanuel Ameisen (Anthropic)
Topic 8 of 18
Published Oct 18, 2026 · Clip 28:31–33:02 (4:31)
Emmanuel explains 'emergence': progress is a smooth trend, but tasks have reliability thresholds, so crossing them produces apparent jumps. Easy, medium, and hard words get solved in sequence; a model at 50% accuracy on medium words wastes your time, but at 90% it's a useful email writer. He describes his lived experience: models went from never writing his code to writing snippets to, this year, writing most of his code as reliability crossed thresholds.