Zvi Mowshowitz on the AI Progress Dial

Watching this really puts the current "plateau" debate into perspective. We often see AI progress as a steady climb, but Zvi’s framing of a "dial" suggests that we might be seeing temporary stalls that are actually precursors to massive leaps once a new efficiency threshold is hit.

What's most interesting is the implication that we aren't just waiting for more GPUs, but for the specific mathematical insights that allow those GPUs to be used more effectively. It suggests that the "scaling laws" aren't a fixed law of nature, but a variable that can be turned up. If we find a way to get 10x the intelligence out of the same amount of compute, the timeline for AGI shifts dramatically. It's a sobering reminder that the gap between "stagnation" and "exponential growth" is often just a few key research papers away.