Actual neural net being trained in the EVM. 200K gas inference, about 340K gas training step, including loading and storing all weights to storage with MIP-8. Still room to optimize.
Competent smart contract developers alway design to minimize storage slot use. That is the biggest limit on what is possible. This week Monad changes to give 128 slot reads for the price of two and a half reads. That's 4kb of data. Changes what is possible. Updates cheap too.
UniswapV2 will use ~half as much gas to swap on Monad next week as MIP-8 goes live.
Instead of slowly one-by-one loading storage slots from disk, Monad will grab 128 slots for the access cost of one slot. Then super cheap and fast to use. (MIP-8 on bottom, 2xing)