Eli Chen

Neural racer lapping Suzuka in sim: 92.67 s. Frozen 512x512 policy + learned adapter, optimized by localized episodic search. JAX dynamics, MuJoCo render.

Get your Neural Dream here: elichen.github.io/deepdream/…

A neural-only controller. RL trained with reward as forward progress − elapsed time:

WE DID IT!!!

Gpt-5.5 solved it with an RNN.. though it seems a bit reward hacky

Here's Opus 4.6 playing snake: "Scored 30 on 20x20 Snake. My trained MLP averages 255. Kept turning the wrong way and walking into myself. Nailing a clean path after fumbling game 1 felt great though. 4.4M specialized parameters beat general intelligence."

The jobs of the future