178K Parameter Neural Net That Wins Poke(rogue)like
178K parameter model beats a Pokemon roguelike after the author rage-quit hundreds of times.

Runs PPO training entirely in-browser via TinyJit WebGPU kernels.
ML engineers and WebGPU enthusiasts
TensorFlow.js · WebLLM
Requires WebGPU.
178K parameter model beats a Pokemon roguelike after the author rage-quit hundreds of times.
Vectorized multi-agent RL combat sim with deterministic checkpointing and telemetry logging.
178K neural net beats Pokémon roguelike with clever 1386-dim state encoding.
Two RL clones duel in-browser with epiplexity selection — no human tuning the explore-exploit knob.
PPO beats classical elevator dispatch 6x on wait times; niche but rigorous.
You play a minimalist web game while a neural net trains in real time, so the opponent starts weak and adapts to your playstyle. The landing page is extremely spare — the demo is quick to load and oddly charming, but there’s little signposting about the model, implementation details, or source code. Fun as a curiosity or teaching toy, but not novel enough to wow ML folks who’ve seen in-browser learning demos before.