Trained Eye

Meta glasses answer when you ask them something, but they can't keep an eye on anything for you. Trained Eye lets you teach them to. You look at something for about ten seconds, say what to watch for ("tell me when the fridge door is left open"), and Trained Eye turns that into a small model of your own that watches the glasses' camera and speaks up through the glasses when it happens. Teaching uses Meta's models to build a training set from your own frames: Muse Wonder turns the sentence into a skill, SAM 3.1 finds the target, Muse Image edits your frames into about 300 variations (the condition, near-misses, other light and angles), and Muse Spark checks each one. Meta's QLoRA recipe then fine-tunes the open-weights Muse Glimmer 30B on one GPU. While watching, the iPhone app streams one frame a second from the glasses, the fine-tuned model scores it, and once the condition has held long enough (20 s for the fridge), the glasses say so, repeating while it stays true. On fridges in kitchens it was never taught on, filmed first-person on phones, the fine-tuned model scored frame F1 0.95, against 0.91 for base Glimmer and 0.92 for Muse Spark. It answers in about 1.6 s a frame against 11.7 s for Spark, at about $4 an hour against $16.53. The burner skill is weaker (F1 0.64) and needs more teaching. The adapter is public on Hugging Face, and a web playground lets anyone try it without glasses. The capture light always stays on, skills about people are refused, and faces are blurred in everything published.