Roku

Watch “Act, Confirm, or Stop? Smarter behavior for AI assistants, wearables & robots — Amit Desai, Roku”. The talk separates model accuracy from the policy governing what a system does when confidence is low. Desai uses an illustrative smart-speaker dataset and relative user-cost assignments to compare acting, stopping, and asking for confirmation. He calls the optimization framework the Outcome User Cost Heuristic (OUCH). The numerical thresholds and costs—including the 79% accuracy example and optimized act/stop/confirm regions—are illustrative recorded claims, not current product specifications.

Use the same recording to follow a practical sequence: begin with the fixed-accuracy comparison, examine why errors and rejections can impose different recovery costs, then study how confidence thresholds are optimized against those costs. Continue to the three-region stop/confirm/act policy and the claim that production systems would use learned real-time decision models rather than fixed offline thresholds. This is useful as a framework for evaluating interaction policy, but the catalog does not establish whether Roku currently deploys OUCH or any particular thresholding method.

Focus on the later interface comparisons. The recording argues that outcome costs should be recalculated for each interface: a television may present multiple confirmation choices, while an incorrect action may disrupt the user’s current state. The talk extends this reasoning toward wearables and robots, where mistaken actions may have greater consequences. These are arguments made in the presentation, not evidence of current Roku product capabilities or plans.