Jia Yangqing, the former chief AI scientist of Ali, said again that the "magic change" in the open source field is not good enough.
On November 16, Jia Yangqing, the former chief AI scientist of Ali, complained that a new domestic model was actually using LLaMA architecture, and that only a few variable names had been changed in the code, causing a heated discussion on the Internet.
After being accused of responding to the big model and publishing the training model, Jia Yangqing's latest post talked about why the "magic change" in the open source field is not good.
Jia Yangqing pointed out that open source is a matter of mutual cooperation, and the essence is that everyone can seek common ground and respect differences, discuss design, run-in views, and finally form a rope to move forward. People who are new to open source and better engineering often have a tendency to say, "I'll take this code, change it by magic, and then I can customize it flexibly. Why not?" But as far as the community is concerned, this kind of thing absolutely does more harm than good, and finally forms a variety of split technical systems, which can not be maintained and collaborated, which is a very big problem for both creators and users.
He also said that the magic reform is cool for a moment, and if the community crematorium is to make a real achievement, it must really have the heart to play with the community.
On November 14, the company officially responded to this publicly. Richard Lin, director of the company's open source team, admitted that it was inappropriate to change the name of the Llama tensor and would change the name of the modified tensor from Yi to LLaMA as suggested by the outside world.
On November 15, zero ten thousand things officially made public an explanation of the Yi-34B training process. Kaifu Lee also forwarded the note on WeChat moments, saying: "the description of the training of the Yi-34B model of 010 everything also responds to the discussion of the model architecture in the past two days. The global large model architecture has gone from GPT2-- > Gopher-- > Chinchilla-- > Llama2- > Yi, and the industry has gradually formed a general standard for large models (just like being a mobile App developer will not create a new infrastructure other than iOS and Android). 01.AI started to benefit from open source, but also contribute to open source, learn modestly from the community, we will continue to make progress. "
Related readings:
"AI" responded to the Yi open source model "plagiarizing LLaMA" incident: admit that the practice is inappropriate and will change the name of the dispute tensor.
