Apprentice AI

Skilled trades cannot train people as fast as they lose them. The US needs about 207,000 new electricians, HVAC technicians, plumbers and industrial mechanics every year, mostly to replace workers who leave (BLS), and an electrician apprenticeship is 8,000 supervised hours. What an expert knows lives in their hands: the order of the work, what "done" looks like, the mistakes they learned to avoid. Shadowing passes it on one beginner at a time. ApprenticeAI records that knowledge once and hands it to anyone. An expert does the job on Ray-Ban Meta glasses and talks through it. Meta's models compile that one recording into a guide: each step's instruction, a photo of the finished step, why it matters in the expert's words, and the mistakes he warned about, kept separate from mistakes the model adds itself. We built it with Chibuoke, a PhD researcher at Purdue's Herrick Labs, on a task new lab members get wrong: wiring a type T thermocouple into channel 16 of a National Instruments module, where swapping the copper and constantan wires silently corrupts every reading. He recorded it once, in under three minutes; Muse Spark compiled six steps from it. Then a beginner put on the glasses. Chibuoke wasn't in the room, but his voice was: every instruction is his (cloned with his written consent), and when the beginner asks "which side does the copper go on?", Muse Spark finds the moment he explained it and plays his actual recording. SAM 3.1 watches for the beginner's hands to leave the work; on steps the camera can judge, Muse Spark compares the frame with his done photo before moving on. Results from 18 recorded live runs: 51 step advances since we switched to the hands-off signal, every one triggered by the beginner's pause, none by a manual override. All 3 runs on the final six-step guide finished, the best in 2:32, at $0.55 to $0.61 per run.