Ex-Apple/Berkeley Researcher Launches Perceptual Labs for Weight-Space Programming

jzlegion · x · 2026-11-18

KJHMiao announces Perceptual Labs, betting that weight-space programming will shift AI from training models to engineering them — and eventually to self-improving models. Pitch: precise edits (change one behavior without forgetting), efficiency (edits written in language, no new training runs), and self-improvement (auditable weight edits). Use cases include model debugging via neuron tracing, behavior removal directly in weights, and capability transplant without new datasets. Waitlist is open.

Original post →

More from Companies & People

Companies & People channel →