Simo Ryu, Stable Diffusion LoRA creator, claims US frontier AI models lead Chinese models by two years

Yes I understand that benchmark wise chinese models are only 6 month behind. But I am saying it will take them 2 years to cross the gap, because US models do not distill from other better model.

In other words, I think chinese models would be 2 years behind had they not used output of US models.

@cloneofsimo This tells me that you haven't meaningfully used Chinese frontier models

@cloneofsimo is this related to recent reports about Korean LLM companies defrauding the govt

@cloneofsimo Uh did you see glm 5.2? It's def ahead of opus 4.5 and probably gpt5.3, maybe others as well, but def ahead of grok 4.3 and gemini 3.1.

you're talking past each other: peak capabilities are much closer than 2 years, but they feel worse because they generalize differently/worse.

China is starting to catch up on pre-training so they have more of that "big model smell", and starting to move away from distillation as a driver for reasoning gains.

(distilling from reasoning traces without logits makes models feel awful out of distribution. like actually brain damaged.)

@cloneofsimo What about the reverse? What past frontier models are current frontier OSS models ahead of? Two year gap would put existing OSS frontier at the level of 3.5 sonnet and gpt-4o

@cloneofsimo But this hypothetical doesn’t guide anything? They will distill and find way to distill. It is like claiming without OpenAI, the whole LLM revolution will be 3~4 years late. Probably true but not useful?

@peterom @cloneofsimo or that he forgets how weak models were two years ago.

probably both. people tend to forget how fast things are going, and china hawks tend not to use chinese models.

@cloneofsimo they're not even comparable, anthropic is alone

chinese models are unusable when you give them real agentic work

like I said, they're just now solving two of the biggest factors: M3 to me is probably the best sign of the pre-training progress so far, and K2.7 is the best sign of the post-training improvements...

but no one Chinese model is really the best of all that recent progress combined. I expect it'll be another generation before it all comes together fully.

@batcz @cloneofsimo What you're saying was true until recently, recent models generalise as well as e.g. Opus 4.5 in my extensive experience.

@cloneofsimo It was true only for gpt 3.5 & 4 Right now internal models for their labs might only have 6 months of delta at max.

@jmbollenbacher @cloneofsimo Either way:

@cloneofsimo Not very hot take: @clouneofsimo is stupid AF

@cloneofsimo wrong, its actually only 5 days ahead

@cloneofsimo Wrong

@cloneofsimo Do you remember that Opus 4.8 said it was Qwen / DeepSeek when asked in Chinese via API? 😂

@cloneofsimo Based on what exactly? There is no evidence of this…

@cloneofsimo that was true 2 years ago

@cloneofsimo Your speculation is disproven by facts