Kai-Fu Lee's New Move: 01.AI Open-Sources Large AI Models, Leading Globally
-
Recently, 01.AI, an AI company led by Kai-Fu Lee, Chairman and CEO of Sinovation Ventures, officially open-sourced two pre-trained large models, Yi-34B and Yi-6B, on its official website. This marks the company's first major release since its establishment in July and represents China's first domestically developed model to top the global open-source large model rankings.
The Yi series models released by 01.AI include two versions: Yi-34B (34 billion parameters) and Yi-6B (6 billion parameters). Both versions were made available on the open-source community Hugging Face on November 2. As of November 5, Yi-34B ranked first on the English pre-trained open-source model leaderboard with an average score of 70.72, outperforming large-scale models such as LLaMA2-70B and Falcon-180B.
On the Chinese large model leaderboard C-Eval, Yi-34B also surpassed all global open-source models, becoming the "double champion" of global open-source large models. It is the first and only domestically developed model to successfully top the Hugging Face global open-source model rankings.
The outstanding performance of the Yi series models is attributed to 01.AI's technological innovations and optimizations. First, Yi boasts the world's longest 200K context window, capable of processing approximately 400,000 characters of ultra-long text input. In comparison, OpenAI's GPT-4 has a context window of only 32K, handling about 25,000 characters.
The context window is one of the key indicators of a large model's comprehensive computational capabilities, crucial for understanding and generating context-specific text. Language models with longer windows can process richer knowledge base information and generate more coherent and accurate text.
Second, Yi retained as much general capability as possible during the pre-training phase, without excessive focus on math and code data. As a result, it excels in metrics such as commonsense reasoning, reading comprehension, and knowledge reasoning, aligning closely with Hugging Face's evaluations.
Finally, Yi implemented strict data selection and cleaning processes, filtering 3T of high-quality data from over 100T of raw data. This dataset includes bilingual Chinese-English corpora, enabling Yi to better meet the needs of the Chinese market.
01.AI's AI strategy reflects Kai-Fu Lee's forward-looking judgment and strategic deployment for the AI 2.0 era. Lee believes that the biggest business opportunities in AI 2.0 will emerge in consumer-facing (To C) super applications, with large models serving as the core technology to achieve this goal.
Large models refer to AI models capable of processing massive datasets, offering strong versatility and creativity. They can perform tasks across domains, media, and languages, making them the cornerstone of AI 2.0. Lee has stated, 'People with experience in large model infrastructure are rarer than algorithm experts.' Thus, 01.AI's strategy centers on large models, targeting AI 2.0 to build China's leading large-model technology team.
What progress has 01.AI made in its AI strategy? How do its new products impact the company's commercialization and offerings? The team comprises top-tier AI talent from global giants like Google, Microsoft, Alibaba, Baidu, ByteDance, and Tencent, continuously attracting elite Chinese AI professionals worldwide.
Members of 01.AI's algorithm and model team include experts whose papers have been cited by GPT-4, award-winning Microsoft researchers, and super engineers honored with Alibaba's CEO Special Award. Collectively, they have published over 100 large-model-related papers at prestigious conferences like ICLR, NeurIPS, CVPR, and ICCV.
The Yi series models, 01.AI's latest product, represent the team's concentrated efforts over the past six months. With steady progress and world-class research and engineering capabilities, they have delivered a globally competitive breakthrough. The open-sourcing of the Yi models not only showcases 01.AI's technical prowess but also opens vast opportunities for commercialization and product development.
Kai-Fu Lee stated that 01.AI will leverage the Yi series models to create more To C super applications, inviting the developer community to collaborate on building an ecosystem for the open-source Yi models and jointly shaping the super applications of the AI 2.0 era.
01.AI's AI strategy has also gained recognition and support from the capital market. Reports indicate that the company has completed a new funding round led by Alibaba Cloud, with its valuation exceeding $1 billion, securing its place as a unicorn.
This marks 01.AI's first funding round since its establishment in July this year and represents Alibaba Cloud's inaugural investment in the AI 2.0 domain. Alibaba Cloud's investment not only provides 01.AI with ample financial backing but also grants access to its powerful cloud computing platform and computational resources, facilitating the acceleration of large model development and applications.
Kai-Fu Lee revealed that 01.AI will continue to invest heavily in AI, with the next focus areas being multimodal large models and consumer-grade super applications based on large models. Lee believes that AI 2.0 is the most significant technological revolution in history, with the greatest potential to change the world lying in platforms and technologies. Similar to Microsoft Office in the PC era and WeChat, TikTok, and Meituan in the mobile internet era, the highest probability of explosive commercial growth lies in ToC applications. 01.AI's vision is to create another WeChat or TikTok in the AI 2.0 era.