Smart glasses, on paper

01 / What it was

Two phone notes from 6 May. The first is a product sketch for rugged AI glasses: no AR display and no cloud, just a camera, microphones and open-ear speakers, with a vision-language model running on the glasses themselves.

The second is a short report on whether a very small open vision-language model could run on wearables, based on published papers. Nothing was built or measured by us.

02 / How it worked

• Always-on camera, microphone array and open-ear speakers; no screen in front of the eyes.
• All inference local, so it works without a phone or internet connection.
• The report looked at a 256M-parameter open model: the papers it cites put it under 1 GB of GPU memory, under 500 MB when quantized to 8 bits, with time to first token as low as about 23 ms after optimisation.
• Privacy was part of the case: keep sensitive camera data on the device.

04 / Where it led

It stayed on paper; the question of where intelligence should run (“what should remain local?”) later became one of the questions on omo.sh.

Sources

1. iPhone note, product sketch. iPhone note, 6 May 2026 No AR. No cloud dependency. No distractions.
2. iPhone note, product sketch. iPhone note, 6 May 2026 You talk. It understands. You move. It adapts. You explore. It delivers.
3. iPhone note, technical report. iPhone note, 6 May 2026 Technical Performance Report: SmolVLM-256M for Wearables
4. omo.sh research page (Jul 2026). omo.sh, July 2026 version Agency What should remain local?