A developer on Hugging Face, a collaborative AI platform, found that Yi, the large language model released by Chinese AI venture 01.AI, used “exactly” the same structure as Meta’s Llama, but changed the names of two tensors. AI pioneer Kai-Fu Lee’s venture 01.AI, founded in June, was recently valued at $1 billion following a funding round led by Alibaba’s cloud unit. The developer called on 01.AI to republish Yi with the tensors aligned with Llama’s architecture. The Chinese AI unicorn, in response to the allegations, told Chinese outlets on Tuesday that Yi “draws on public results at the top level of the industry,” and its design is based on the proven structure of GPT. An open-source signatory of 01.AI, Richard Lin, published a statement under the Hugging Face post claiming the tensors’ renaming had been needed to meet “experimental requirements” during model training, and had not been switched back before release. “We’re going to change them back from Yi to Llama, just as you suggested,” wrote Lin. [Hugging Face; Jiemian, in Chinese]
Related
