distillation is overblown but chinese models still benefit from being able to distill and train on synthetic data produced by US frontier models. i feel like if it were easy for them to completely catch up they would have already given how many companies are attempting to do so there
the more relevant question is "when do chinese models get good enough that the difference doesn't practically matter for 99% of use", which i think we're 1-2 years away from. sooner for coding specifically.
Date: July 1st, 2026 6:10 PM Author: wonderful khaki depressive
Distillation is overblown if being treated as a full magic catch-up. But underblown as an explanation for rapid practical convergence. And it is definitely cheating big time, although probably not legally exact to call it "counterfeiting" it is more in that direction than anything else.
Also I think the part you are missing is that a model has not "caught up" just because it meets benchmarks if it can only generate better capability through privileged access to a stronger teacher. That just makes it a successful parasite.