Pat Simmons demonstrates applications using what he describes as OpenAI's Decisions API, contrasting fixed-choice classificationClassification assigns an input to one or more defined categories, such as identifying whether a message is spam or a document is relevant. with generating a free-form answer. The examples include an autonomous game demonstration, a screen-aware focus timer and voice controlA voice interface lets a person interact with software through speech, potentially using AI to interpret requests and produce spoken responses. mapped to a predefined list of Mac actions.
The walkthrough exposes limitations as well as successes. The voice-control prototype cannot perform commands outside its configured action set, while the social-feed cleaner misclassifies ordinary company posts as advertisements. A photo-based calorie tracker also struggles with an unusually large meal because portion choices are coarse.
Further examples cover blurring sensitive information, choosing links in a Wikipedia race and identifying false starts or pauses in a video-editing workflow. Simmons proposes routing uncertain editing decisions to a reasoning modelAI model routing sends each request to a model chosen for that request's complexity, cost, speed, privacy, or other requirements. for a second check. His latency, cost and comparison results are presented as observations from these demonstrations, not independent benchmarks or general guarantees.
Watch on YouTube
