**RT @gdb:**
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.
