US-China AI Model Gap To Widen With Compute Scaling And RSI
Reactions from ranked influencers
2 posts
in the full version, most takes are undeniably correct except the political
the rest is mostly slop but this seems true though again, "RSI" as a discrete event. LLMs have been doing autoresearch for a while. Even Kimi can do autoresearch. This is an acceleration already
Uh huh.. https://x.com/kente_clarke/status/2078443306426519644?s=46 (2) The model gap between China and US, while it seems to have been closed, hasn’t. Gap has actually been pretty similar for the past two years, and I believe the gap will actually widen from here. •The perception is that gap has tightened but main reason is (1) delays on model releases from frontier labs due to US government / security risks and (2) underappreciated reality that internal models at labs significantly far ahead of public releases. Very few know this second point, (3) china benchmark maxxing, they seem better on benchmarks than they are in actual use. •Why gap will widen? Over last 1.5-2 years labs have been compute constrained, they pushed their compute to scale RL vs. pretraining because returns per unit of compute were higher there, now labs in place where compute is scaling faster which now allows them to focus on both pre-train scaling and bigger RL runs simultaneously – model progress accelerates from here. •As RSI reached in next 6-12 months gap widens dramatically from here. (3) While these Chinese models today are open-weight, this will not continue in definitely. In a couple generations of China model improvement, their best models will closed and cost calculations will look different at this point. •If China models get too good, US will either ban them, or China will keep labs from selling them •Opensource industry in US in a few years, will be US opensource labs