Ugh... missed the 2 hour window. Initially I was focused on the training and memory requirements, but as I thought about it while doing other things, it occurr

Ugh... missed the 2 hour window. Initially I was focused on the training and memory requirements, but as I thought about it while doing other things, it occurred to me that the same things that work for LLMs should work with your idea. Use an autoencoder to try to reduce the dimensionality of the data, while preserving as much information as possible. This gains you orders of magnitude data compression while remaining useful, and reducing compute requirements for the next steps by that amount squared. Once the autoencoder is sufficiently effective, then you can try to predict the next state at some point in the future. If you have any tagging data, then you can do the whole gradient descent, repeat for a trillion iterations thing. The thing is, trillions of cycles aren't really a barrier these days. Start with deliberately small systems, and work up. Good luck!