In actual fact, among the many few lengthy Chinese language tokens in GPT-4o that aren’t both pornography or playing nonsense, two are “socialism with Chinese language traits” and “Folks’s Republic of China.” The presence of those phrases suggests {that a} important a part of the coaching information truly is from Chinese language state media writings, the place formal, lengthy expressions are extraordinarily widespread.
OpenAI has traditionally been very tight-lipped in regards to the information it makes use of to coach its fashions, and it in all probability won’t ever inform us how a lot of its Chinese language coaching database is state media and the way a lot is spam. (OpenAI didn’t reply to MIT Expertise Evaluate’s detailed questions despatched on Friday.)
However it isn’t the one firm scuffling with this downside. Folks inside China who work in its AI business agree there’s a scarcity of high quality Chinese language textual content information units for coaching LLMs. One purpose is that the Chinese language web was, and largely stays, divided up by large firms like Tencent and ByteDance. They personal many of the social platforms and aren’t going to share their information with rivals or third events to coach LLMs.
In actual fact, that is additionally why search engines like google, together with Google, kinda suck on the subject of looking in Chinese language. Since WeChat content material can solely be searched on WeChat, and content material on Douyin (the Chinese language TikTok) can solely be searched on Douyin, this information shouldn’t be accessible to a third-party search engine, not to mention an LLM. However these are the platforms the place precise human conversations are occurring, as a substitute of some spam web site that retains making an attempt to attract you into on-line playing.
The shortage of high quality coaching information is a a lot larger downside than the failure to filter out the porn and normal nonsense in GPT-4o’s token-training information. If there isn’t an present information set, AI firms should put in important work to determine, supply, and curate their very own information units and filter out inappropriate or biased content material.
It doesn’t appear OpenAI did that, which in equity makes some sense, given that individuals in China can’t use its AI fashions anyway.
Nonetheless, there are various folks dwelling exterior China who wish to use AI providers in Chinese language. And so they deserve a product that works correctly as a lot as audio system of another language do.
How can we clear up the issue of the shortage of fine Chinese language LLM coaching information? Inform me your concept at zeyi@technologyreview.com.