I think ALOT of technical ppl from back in the day are roasting the simplicity of it but they got timing / pain / packaging right!

One of these players is AI - can you guess which? Lately, we've been playing a lot of Smash at Frisson Labs - and training small behavior models that can play with us. Biggest takeaways so far: - A model that's "good at the game" isn't the same thing as "fun to play with". - Online latency (and human brain latency) completely changes how the model should be trained and needs to be factored in early. - Iteration speed quickly became one of our biggest bottlenecks. To solve for iteration speed, we looked into faster matmuls, fused attention, buffer reuse, CUDA-Oxide kernels, smaller models, and only processing every other frame (Nothing too magical, but a lot of fun little learnings that added up). Now we're able to train a Melee model from scratch to beat a lvl-9 CPU in 10 minutes on one GPU. Huge shoutout to @ericyuegu for advice and support in this effort! I'll share everything we learned about training optimization in the comments below 👇