Zero-training: LLM directly writes neural network weights implementing Pong in 8 minutes

Sl33py_4est · reddit · 2026-11-06

A Reddit user demonstrated "weight space programming": an LLM (Astra, labeled GPT-6) directly fabricated neural network weights implementing Pong's rules with zero training — about 8 minutes and 200k tokens (mostly output), using multiple task-specific small networks. Movement, collisions and scoring all emerge from the weights, with no rule engine running alongside.

The model also built a live editor to tweak running weights mid-game: changing one neuron stopped the ball from moving left, so it eventually overlapped the right paddle. The model distinguishes this from editing a pretrained model — it's constructing a network from explicit rules, not rewriting learned representations.

Original post →

Zero-training: LLM directly writes neural network weights implementing Pong in 8 minutes

More from Fun