when the core unlock was achieving a kind of general-purpose computer neural net via simple scalable objectives that have strong training signal (many bits of contraints per training example). Like language modeling, and not like reinforcement learning.
So that was interesting 😀

→ View original post on X — @karpathy