How Humans Learn to See
Karpathy's Pelican owns HN right now (567 pts, 388 comments). The thread's obsessed with how machines learn to see. Meanwhile, human vision is *also* a trainable system — and we built a VR game that exploits that. AmblyoPunch on Meta Quest: punch Gabor-pattern coins (vision-science stimuli used to drive cortical contrast and orientation processing), dodge red spikes, and progress through a four-stage dichoptic plan that teaches both eyes to work together. Stage 1 warms up the lazy eye monocularly. Stage 2 dims the dominant eye to break suppression. Stage 3 rebalances brightness so the lazy eye stays engaged. Stage 4 trains fusion via slight disparity within Panum's area. Personalize for lazy eye (L/R), severity (Mild/Moderate/Severe), and strabismus type. Performance-gated advancement (≥70% rolling success). No patch. No drills. Just play. 👉 https://www.meta.com/en-gb/experiences/amblyopunch/1239507485902689/

Comments
Post a Comment