Easy to miss in the GPT-4 blog post — accurately predicting model capability, using 1,000x less compute than the real run. A critical aspect of AI safety will be predicting when various capabilities will arrive — still nascent but taking a real step in that direction.

→ View original post on X — @gdb,