The whole point of OpenAI and Anthropic is that only the USA has the resources to train frontier models. The fact that China, under massive sanctions on chips, in training models on par with OpenAI and Anthropic (many of which are released with open weights) is really bad news for the AI bubble.
I think it’s pretty well known that a lot or all of the Chinese AI models are trained on ChatGPT/Claude so to be fair to OpenAI/Anthropic, they are probably the two companies that are mostly behind pushing AI capabilities forwards at the moment. It’s not 100% proven but there is a lot of evidence. That being said, I don’t feel bad for OpenAI/Anthropic personally due to how the original models were trained.
Not really, that’s just the data side of the equation. Both kimi and glm used deepseek’s architechture, deepseek has been kinda frontier in efficient/cost effect intelligence (acc to artificial analysis).
So the answer is much more nuanced than “they just copied chatgpt”.
Yeah I don’t think they just copied the other models end of story. If that was true it would be impossible for Kimi K3 to be in the same realm as Fable/GPT-5.6 since they all released around the same time. The Chinese companies are clearly also bringing new things to the table. But I think it’s highly likely most of the Chinese models have been at one point trained on at least one of the US models.
The whole point of OpenAI and Anthropic is that only the USA has the resources to train frontier models. The fact that China, under massive sanctions on chips, in training models on par with OpenAI and Anthropic (many of which are released with open weights) is really bad news for the AI bubble.
I think it’s pretty well known that a lot or all of the Chinese AI models are trained on ChatGPT/Claude so to be fair to OpenAI/Anthropic, they are probably the two companies that are mostly behind pushing AI capabilities forwards at the moment. It’s not 100% proven but there is a lot of evidence. That being said, I don’t feel bad for OpenAI/Anthropic personally due to how the original models were trained.
Not really, that’s just the data side of the equation. Both kimi and glm used deepseek’s architechture, deepseek has been kinda frontier in efficient/cost effect intelligence (acc to artificial analysis).
So the answer is much more nuanced than “they just copied chatgpt”.
Yeah I don’t think they just copied the other models end of story. If that was true it would be impossible for Kimi K3 to be in the same realm as Fable/GPT-5.6 since they all released around the same time. The Chinese companies are clearly also bringing new things to the table. But I think it’s highly likely most of the Chinese models have been at one point trained on at least one of the US models.
Again I don’t care that they did this.