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

Popular posts from this blog

London Schools Are Using VR to Treat the Brain. Could Lazy Eye Be Next?

Attention under pressure

London schools are trialling VR for student mental health — the NHS is backing it. Could this same approach treat lazy eye?**