Act, Confirm, or Stop? Smarter behavior for AI assistants, wearables & robots — Amit Desai, Roku

Amit Desai never makes the model better. Accuracy stays pinned at 79 percent for the whole talk, and user pain still drops by nearly half. Desai has worked on voice interfaces for years, at Roku and before that on a widely used smart speaker, and his argument is that teams pour effort into one knob while ignoring a second one that is entirely independent of it. Knob one is accuracy, chased a percentage point at a time through the wake word, recognition, and understanding layers. Knob two is what the system decides to do when it is unsure. Take a thousand music requests where 790 play the right song. The naive system always acts. Give it the option to decline instead, saying it did not catch that, and you now need a confidence threshold, and intuition picks badly. A sensible looking 65 percent is worse than the actual optimum of 43. Getting there requires admitting that bad outcomes are not equally bad. Playing the wrong song means hearing it, realizing it is wrong, talking over it to stop it, and asking again. Being told sorry, I did not understand costs a few seconds. Desai turns that into a unit cost per outcome and minimizes the total, an approach he named the Outcome User Cost Heuristic, which spells OUCH, which he is delighted about. Adding a third behavior, confirming the guess out loud before acting, splits the confidence range into three regions and two thresholds, and pushes the cost lower again. He closes by moving the same reasoning onto a television, where showing choices on screen changes every cost in the equation. Speaker info: - https://www.linkedin.com/in/amit-v-desai/ Timestamps: 0:00 - The power and the pain of voice 2:17 - Why error costs rise with embodied AI 3:22 - Two knobs, and the one nobody turns 4:28 - A thousand requests, 79 percent right 5:36 - Adding the option to stop 6:40 - Choosing a threshold, and why intuition fails 7:43 - Bad outcomes are not equally bad 9:52 - Naming the cost function OUCH 11:01 - Watching the optimum move 13:12 - Adding a confirm behavior 16:38 - What changes in a real system 17:42 - The same method on a television