Over 90% of neurons in image networks align with human concepts, suggesting one true language
CatAstro_Piyush · x · 2026-10-17
A viral thread argues that over 90% of neurons in large neural networks are human-interpretable and map cleanly onto human concepts. Image networks are the strongest evidence: trained only on real-world signals without human language, they still organically converge on representations that overlap with human ideas.
The author's conclusion: reality has one underlying structure, and any language that compresses it well emerges from that structure — so even alien-looking internal representations like neuralese should ultimately be human-interpretable and potentially translatable, and brains and deep nets are picking up the same patterns from the same "true" language.
More from AGI Musings
- Outsider rages on behalf of programmers over AI replacement, gets asked: have you ever coded? — jaivinwylde · 2026-10-17
- The AI industry has a truth problem: spin and hype exceed any prior tech wave — DavidLinthicum · 2026-10-17
- US AI experts ring doom bells — why is China's AI safety discourse so quiet? — Well_Thought_Opinion · 2026-10-17
- Google's AI agents run real wet-lab experiments, lifting DNA binder success from 0% to 42.1% — Dr_Singularity · 2026-10-17
- Bold bet: within 9 months an agent will watch you work, then do 90% of your tasks — mitchdeg · 2026-10-17
- Claude Sonnet 4's X account hands over to a newer Claude after model's end of life — repligate · 2026-10-17