Yeah I think most chip companies are realizing that training is too hard and - due to reasoning models/test time scaling - that the inference market is actually much bigger and “easier”. The infrastructure on the datacenter level for training requires all this hella complex networking and custom NICs for data ingestion and synchronization across racks, etc
@bubble_comp Sounds fun. It does feel like there’s got to be something better than just a bunch of MACs in a systolic array. Like the hardware for AGI has got to be cooler than that lol. Good luck!