Show HN: Teaching a 4M-param model to read like a child, with RL and no labels (rjb7731.github.io)

I have been really interested in reinforcement learning and had a go at trying to teach a model through simulation to read and speak through experience.

It’s two small networks (~4M parameters total) trained on a MacBook. It passed generated versions of England’s Year 1 Phonics Screening Check (36.6/40, pass mark 32), and the bit I like most is that it taught itself irregular words like “glove” through reading practice. It was given 42 sound slots and the voice learned from transcribed recordings, so it’s not scaffolding-free. Details on that are in the post.