A playground and experiment exploring "how small can a neural network be and still meaningfully play Pong?" You can play against the models directly in your browser, also on mobile, no signup needed. Recommend hitting the fullscreen button for better gameplay.
The discussion https://news.ycombinator.com/item?id=50341174 sparked my interest and I started experimenting a bit with generating neural network weights. I used Codex to program the weights, developing formulas that generate them and refining them using feedback from simulated matches. I also tried one-shot prompting Astra for the weights directly.
The surprise, I think, is how hard it’s been in my training attempts to outperform the programmed model using the same 163-parameter architecture.
The playground also has some explainers to help interpret the weights, starting from a one-hidden-neuron example with just 13 parameters...
It’s all open source, and there is also a league where community weights can be submitted. It would be interesting to see how others would approach this. Can anyone train, prompt or program weights that outperform the initial ones?
Code: [https://github.com/jonaslindblom/pongarena](https://github.com/jonaslindblom/pongarena)
Comments URL: [https://news.ycombinator.com/item?id=50509161](https://news.ycombinator.com/item?id=50509161)
Points: 1