Wordless
Gemini 4 NYC HackathonWordless is a one-shot procedure learner and real-time verifier. From one unlabelled demonstration video, Gemini 4 Carbon induces the steps and a forced-alignment pass over per-frame Decisions API posteriors localizes each one; an image model renders them as a wordless picture card. The card then runs as an 8 Hz streaming classifier on the next person: posteriors are Viterbi-decoded, aligned to the card by edit distance, and a missed step is flagged mid-task. On 16 public recordings of other people, 97% of performed steps were never falsely flagged, compared with 91% for a single model call.





