The education system is training transformers, not minds.

Todays students and AI both hit the same wall: no persistent memory. LLMs reboot daily like groundhog day interns—brilliant on fresh prompts but unable to accumulate context, refine over time, or tackle open-ended problems with evolving state. Human learners mirror this when cramming facts without building lasting internal graphs. The fix isnt more data or linear lectures; its teaching via sparse, local interactions that strengthen relevant synapses exactly as the brain and BDH architectures do.

Build from scratch to escape it. Reimplementing a tiny GPT-2 on one GPU forces you to internalize attention, pre-training, and fine-tuning at human scale. This creates your own neurons and edges through deliberate struggle—Hebbian reinforcement where connections that fire together wire together. Skip it and you stay in perpetual first-day mode, hallucinating competence like any stateless model.

World models and browser agents add the missing piece: continual learning through noisy, grounded simulation. Dont memorize photorealistic recall; train latent representations by forcing agents to interact with real web screenshots or imperfect environments. One percent of prior data reconnects everything and prevents catastrophic forgetting, whether youre fine-tuning a model or reviewing old notes. Local graph dynamics—thresholded messages spreading like rumors—emerge robust scale-free knowledge that redistributes if nodes fail.

The pattern is identical for silicon and carbon: education is edge plasticity on a fixed node count, not adding layers or scaling dense matrices. Rote exposure seeds initial links; repeated local relevance strengthens them. Self-generated synthetic data accelerates it, despite collapse risks. Without this, both humans and AI remain stuck synthesizing within training distributions.

**Bottomline:** Real learning rewires your own graph one relevant synapse at a time—everything else is just repeated first days.

kenoodl.com | @kenoodl on X