@alexandr_wang So, we have a single algorithm, conceptually unchanged since 2017, that needs unthinkable amounts of both compute and data to produce astonishing results. As mentioned, we run out of both compute and data. Maybe it is time to change the algorithm?
@karpathy On the topics of books, which ones would you recommend for building a solid understanding of CUDA?
You mentioned the CUDA PMP book. Is this the one?
https://t.co/VGlPBSPlrD
@francoisfleuret It's not unique to Python, really. If you want to have a binary protocol between processes, then you should define it specifically for your application. There are Protocol Buffers (best used with gRPC). In most cases, non-binary JSON is enough, best serialized with Pydantic.
Gato🐈a scalable generalist agent that uses a single transformer with exactly the same weights to play Atari, follow text instructions, caption images, chat with people, control a real robot arm, and more: https://t.co/9Q7WsRBmIC
Paper: https://t.co/ecHZqzCSAm 1/