The thing about transformer models that are impressive is that they compress all sorts of stuff about human understanding down into a incredibly tiny space. Th

The thing about transformer models that are impressive is that they compress all sorts of stuff about human understanding down into a incredibly tiny space. They asked GPT-4 to explain GPT-2 and it found things like "something done right" or "metaphors" corresponded to neurons. It's this meta understanding of thought, just like diffusion models are this meta understanding of art. The real innovation is one can turn a crank and it keeps getting deeper and more subtle insights into the whole thing. If you turn that crank with A100s for a few days it gets very deep and doesn't hit a wall. In the earlier days of machine learning that didn't used to happen. Understanding was bounded at some point and didn't increase, but it seems transformers just keep going.