AI's slowest phase is the one that matters

Arvind Narayanan’s ICML keynote, “What will be left for us to work on?”, makes an argument we keep coming back to: AI is transformative on the scale of electricity, but its impact arrives through the slow phases — diffusion and adaptation — not through some milestone reached in a lab. His history is the tell. When electricity replaced steam, factory owners first bolted electric motors where the boiler used to sit, and it barely helped. The real gains took forty years and a reorganization of the whole factory around portable power: new layouts, new training, new labor law.

We find the reliability part especially clarifying. Narayanan separates capability from consistency — a “70 percent accurate” agent that works dependably on a knowable 70 percent of tasks is something you can deploy; one that fails unpredictably three times in ten is close to useless. And, he notes, the benchmarks don’t distinguish the two. That gap between a headline number and a tool you’d actually trust in production is the whole game for anyone who builds software with care.

His speculative future — “extreme personalization,” software shaped to a team rather than one app for a billion people — is the part that lands for a small studio. We’d rather build the thing that fits, slowly and well, than the drop-in replacement that doesn’t. It’s a good frame for the years ahead: bet on amplification, and build the taste and judgment no model can hand you.