Biological Anchors: A Trick That Might Or Might Not Work
Astral Codex Ten Podcast
00:00
The Contribution of New Algorithms to Computer Performance
My model is something like for any given algorithm, e g, s vms, alphago, alphabeta pruning cove nets. E g, there is an effective compute regime where dumping more compute makes them better. If you go above this regime, steep, diminishing marginal returns occur. In the relatively small regimes of old agarithms, new argarithms and old Agarithms perform similarly. With small amounts of compute using alphago instead of alphabetta pruning doesn't get you that much better performance than, like, an order of magnitude of compute. I think the studies that karl cites above are decent evidence that multiplitive factor of compute
Transcript
Play full episode
