{ "video": { "id": "Zd5b40Jbp_k", "title": "Act, Confirm, or Stop? Smarter behavior for AI assistants, wearables & robots — Amit Desai, Roku", "duration": 1225, "upload_date": null, "channel": "AI Engineer", "source": "AI Engineer", "event_format": "In person" }, "analysis": { "video_id": "Zd5b40Jbp_k", "title": "Act, Confirm, or Stop? Smarter behavior for AI assistants, wearables & robots — Amit Desai, Roku", "one_liner": "A voice-AI veteran argues that you can make an assistant meaningfully more satisfying without improving its accuracy at all — by adding 'stop' and 'confirm' behaviors and choosing their confidence thresholds to minimize a user-cost function he calls OUCH (outcome user cost heuristic).", "summary": "Amit Desai (voice SME, ex-Alexa, now Roku) says there are two knobs for voice assistant satisfaction: model accuracy, which everyone works on, and 'system decision' — what the system does under uncertainty — which is orthogonal and under-exploited. Using a toy smart speaker with 1,000 annotated requests at fixed 79% accuracy, he shows that adding a 'stop' behavior and then a 'confirm' behavior, and picking the confidence thresholds by minimizing a weighted cost of bad outcomes (measured in extra seconds the user needs to reach success), drops average user cost from 2.1 to ~1.26 'ouch points' per turn with zero accuracy change. He argues intuition-picked thresholds (65%) are measurably worse than optimized ones (43%), and that this matters more as assistants move from answering to acting — embodied or digital.", "key_points": [ "Two knobs for user satisfaction: technical accuracy, and 'system decision' — smarter conversational behavior under uncertainty — which is orthogonal to accuracy and currently under-used.", "Running example: 1,000 spoken music requests, human-annotated; 790 correct song (79%), 210 wrong song (21%). Accuracy is held constant for the whole talk.", "Errors get more costly as we move from conversational bots to embodied AI taking physical or digital actions (a phone call, an email, a robot throwing your watch out).", "The bad outcomes are not equally bad: a wrong song costs ~10 seconds (hear it, realize it's wrong, shout to stop it, re-request), a 'sorry, I didn't get that' costs ~4 seconds. Heuristic unit = extra seconds for the user to reach the thing they wanted.", "Naming the heuristic: OUCH — Outcome User Cost Heuristic — and the goal is 'minimize the ouch.'", "Act-only baseline costs 2100 across 1,000 turns = 2.1 ouch/turn. A gut-feel 65% stop threshold gives 1904 (~1.9/turn). Solving for the true minimum gives t = 43% and ~1.27/turn — intuition is measurably worse than optimization.", "Adding a third behavior, 'confirm' ('play Kiss by Prince or play Kiss by Chris Brown?'), splits the distribution with two thresholds into stop / confirm / act. Unit costs used: wrong act 10, confirm-yes 2, confirm-no 6, stop 4. A 2-D heat map gives optimal T1 = 41, T2 = 49, total 1464; the recap progression he shows is 2.1 → 1.9 → 1.27 → 1.26.", "Cost weights are inputs, not truths: setting wrong-act to 20 instead of 10 moves the optimal thresholds. Real systems would use a real-time learned decision model, not offline thresholds — 'but the principle is the same.'", "The approach transfers across surfaces with different values: on a TV, confirm can be a visual list of choices selectable by remote (cheaper, so lower confirm cost), while launching a channel that kicks you out of your current state makes a bad act much more expensive.", "He built the interactive threshold/cost explorer and the heat-map optimization with a coding assistant and demoed it live on stage." ], "takeaways": [ "Stop treating accuracy as the only lever: explicitly design and tune 'stop' and 'confirm' behaviors as a second knob you can move without touching the model.", "Assign relative costs to each bad outcome — a defensible heuristic is the extra seconds a user needs to get back to success — and pick confidence thresholds by minimizing that cost function, not by gut feel.", "Log the distribution of confidence scores against human-annotated correct/wrong outcomes; that distribution plus the cost weights is all you need to solve for thresholds.", "Re-derive the cost weights per surface and modality: a multimodal confirm (on-screen choices picked with a remote) is cheap, and an action that destroys the user's current state is expensive.", "As assistants take real actions, budget for higher act-cost weights — the optimizer will naturally shift toward confirming and stopping." ], "topics": [ "voice-ai", "conversational-design", "uncertainty", "confidence-thresholds", "embodied-ai", "cost-functions", "multimodal", "user-experience" ], "tools": [ "Alexa", "Roku", "coding assistants (unnamed, used to build the interactive threshold/heat-map demo)" ], "quotes": [ { "text": "if the robot throws your watch out with the trash, it's a lot worse than playing the wrong song.", "at": "02:15", "url": "https://www.youtube.com/watch?v=Zd5b40Jbp_k&t=135s" }, { "text": "The bad outcomes are not equally bad. They're not the same thing from a user perspective.", "at": "07:58", "url": "https://www.youtube.com/watch?v=Zd5b40Jbp_k&t=478s" }, { "text": "So this whole thing is called an outcome user cost heruristic. So that spells the word ouch and that is some expression of pain.", "at": "09:55", "url": "https://www.youtube.com/watch?v=Zd5b40Jbp_k&t=595s" }, { "text": "So effectively we haven't changed the accuracy at all. The system is not any smarter in that sense. But with some clever system behavior, conversational behavior is what we'd call it and some optimization and a cost function called ouch. Um we have from the user's perspective produced a more satisfactory assistant.", "at": "11:44", "url": "https://www.youtube.com/watch?v=Zd5b40Jbp_k&t=704s" } ], "words": 3585 }, "summary_url": "/#Zd5b40Jbp_k", "transcript": { "html": "/transcripts/Zd5b40Jbp_k.html", "txt": "/transcripts/Zd5b40Jbp_k.txt", "vtt": "/transcripts/Zd5b40Jbp_k.vtt" } }