alt Hacker News> Anyone who tried using Chinese models for serious programming
Perhaps, but that’s because they are just objectively worse for many tasks than GPT/Claude. I don’t see how that ties to censorship inside of China.
> censorship limits the amount of high quality training sets by limiting the amount of all training sets
Well again, I don’t really understand how is the lack of high quality Chinese datasets matter much if they have access to English/etc. datasets. That’s only an issue to their users inside of China.
> not have access to commercial-scale advanced nodes
Well the gap was way bigger a few years ago. If Chinese companies can at some point produce GPUs that are competitive cost wise i.e. they are willing to tolerate significantly lower margins than Nvidia (whose margins are obscene) that’s not a huge issue.
> I don’t see how that ties to censorship inside of China.
LLMs are parrots, when they are trained on English data they behave like English / western speakers which carries the __western__ values.
China would never want a shiny model deployed by some solemn state-own institute to pitch western values like democracy or freedom of speech.
So China is better off starting with Chinese content in the first place, but only to find out it shoots itself in the foot, because censorship __removes__ Chinese content from the Internet, and it __prevents__ truth-reflecting, i.e. quality, content from appearing.
When this is realized to be not sufficient, and they have to use English content, the English content needs to be censored, too.
So compared to other models trained on the uncensored set, the Chinese models are trained on the data sets that is likely smaller. And since high quality content is suppressed, there is likely less high quality training sets. Therefore, their models likely suck.
(Meta knows the importance of high quality training sets so that it started distilling its own employees who are smart.)
> produce GPUs that are competitive
The most recently known advanced process in China is some kind that requires multiple exposures in DUV machines. This already limits the cost to a minimum that's probably not economical.
For GPUs to be competitive, there's more to the hardware. Driver also matters. GPU drivers these days are compilers in disguise. How many high quality compiler projects have we heard of from China?
It's unlikely for both the hardware and the software of a Chinese GPU to deliver what's boasted in its marketing material.