Is the AI Capability Overhang Real?

Alex Zhang agrees, in some sense, that the capability overhang is real. Even if progress paused, the current state of models still leaves a lot of impact available. The case he is presenting is the easiest form of the idea. Models have jagged intelligence on a lot of things. They are disproportionately good at coding and math, and the abilities can be translated to many other things.

Someone who was an IMO gold can be applied to a lot of different problem-solving domains and figure them out. He does not know whether the pattern applies to models. In GPU code optimization, one question is whether a model as good as Astra is now, with all of the GPU programming data taken out, would still optimize GPU kernels. Another question is whether it would learn in context roughly what it needs to learn, then have some pipeline or come up with some solution for optimization tasks.

If you took a human who was as smart as Astra, or who knew as much as Astra, there is a mismatch between what the human can do and what Astra can do, and he only floats a harness as where the mismatch sits. The human can actually be approximated a lot more. Shawn Wang hears the comparison as very approximate to the continual learning problem, and Alex Zhang calls that the best example.