(Yicai) Nov. 6 -- The artificial intelligence firm 01.AI that AI pioneer Kaifu Lee set up to develop ChatGPT-like large language models is now worth more than USD1 billion and released its first open-source base model, tech news outlet 36Kr reported.
Yi, which can be used in both Chinese and English, was released today after Beijing-based 01.AI completed a new funding round led by Alibaba Cloud, the report said.
Before the latest fundraiser, 01.AI's debt was in the tens of millions of US dollars due to model training and computing power costs, Lee said in the report. Foundational models need vast resources to be able to answer prompts with text or code while further models with more specific tasks can be built on top of them.
On Nov. 2, the startup uploaded two base models with six billion and 34 billion parameters on Hugging Face, an open-source AI community. As of yesterday, Yi-34B was top of the two leaderboards of pre-trained LLMs and C-Eval, a Chinese evaluation suite.
Yi was mainly trained by crawling the web and databases of publicly available text, Lee said, explaining that the problem with the training data lies in its high repetition rate and poor quality. The team selected three terabytes of data from 100 terabytes of public text and Yi was trained on more English inputs than Chinese ones due to the low quality of the latter source.
Similar to OpenAI, the developer of ChatGPT, 01.AI aims to take AI mainstream. In the era of AI, the biggest business opportunity will appear in business-to-customer super apps, Lee said, adding that 01.AI aims to create a platform as diverse and big as WeChat and Douyin, the Chinese version of TikTok.
As of July, 01.AI had attracted dozens of key members from industry giants such as Alibaba Group Holding, Baidu, Google, and Microsoft after Lee started looking for talent in March. The computer science expert founded venture capitalist firm Sinovation Ventures in 2009 after being Microsoft’s global vice president, as well as Google’s GVP and China president.
Editor: Emmi Laine