Act, Confirm, or Stop? Smarter behavior for AI assistants, wearables & robots — Amit Desai, Roku
Amit Desai from Roku shows that you can make AI assistants far more useful without improving the underlying model accuracy.
His point is that developers focus on pushing recognition accuracy higher while ignoring an entirely independent control: what the system does when it is uncertain. With the same 79 percent accuracy, user pain can drop by half by letting the system choose between three behaviors: act, confirm with the user before acting, or stop and say it did not understand. The key is that different failure outcomes cost the user different amounts—playing the wrong song is worse than saying "I did not understand"—and you can calculate these costs and set thresholds that minimize total user cost, a method Desai calls OUCH (Outcome User Cost Heuristic).
Vibekollen prepared this summary with AI from the original publication. The content belongs to AI Engineer.
