2024-03-14 18:58:15
While several Silicon Valley giants tighten access to the codes of their generative AI models, this is the exact opposite approach that Kai-Fu Lee, a Taiwanese computer scientist and businessman based in Beijing, is trying to take. Her will ? Structuring an ecosystem based on a language model open source efficient, of which Chinese start-ups will be the first beneficiaries.
After a stint at Apple and Silicon Graphics, this globally recognized AI expert participated in the creation of Microsoft Research Asia, an organization which contributed to the training of a large number of leading Chinese engineers. Kai-Fu Lee also chaired Google’s search business in China.
Commercial competition between AI giants will gradually intensify. This will likely oppose ecosystems built around the exploitation of different language models available. According to Kai-Fu Lee, the ecosystem that emerges will be rooted in a language model open source small in size and easy to train. It will bring together a large number of engineers informed of the latest technological advances and capable of transforming these into marketable applications. These engineers will have been familiarized very early with the details of the language model they use. Trained in an institution that includes venture capital activity, they will be able to develop ideas with high commercial potential and likely to attract the most significant funding.
This is the environment he has sought to build for 15 years now.
Sinovation lays the foundations
In 2009, the engineer launched Sinovation, a venture capital company specializing in financing technological start-ups. During the first years, the firm operated in the United States and China. The intensification of the trade war between the two countries and the increasing difficulty in concluding agreements with American companies nevertheless led management to close its Silicon Valley offices in 2019. Today, Sinovation has 400 companies in its portfolio, representing a total of $3 billion in assets under management.
In addition to his financing activities, aware of the contributions of Microsoft Research Asia for the training and development of the skills of future engineers, the businessman created the Sinovation Ventures Institute for AI in 2016. This ensures the wide dissemination of scientific advances and accelerates their transformation into marketable applications.
The project is of major strategic interest for the development of AI in China: it contributes to the training of high-level engineers and finances the companies that these professionals are likely to launch.
01.AI, the open source bet
Last June, Kai-Fu Lee launched a new start-up, 01.AI, after having raised some 200 million dollars from BATX, a group of companies (Baidu, Alibaba, Tencent, Xiamoi) which are sometimes considered as Chinese GAFAM. The objective was to create the first “killer applications”, that is to say applications so convincing that they are adopted by hundreds of millions of users like TikTok.
And things didn’t drag on! In January, 01.AI launched a “multimodal” AI model nicknamed “Yi” based on 34 billion parameters, half as many as ChatGPT4. This modest size makes it easy to train. Throughout the world, many developers have already adopted Yi.
The 01.Ai language model is one of the so-called open source models, as is also, for example, LlaMA, the language of Meta. They are opposed to so-called proprietary models whose technical foundations are not shared. Contrary to what its name may suggest, Open IA which runs ChatGPT is one of the proprietary models, just like the tools developed by Google. The interests associated with openness are multiple. It facilitates access to language models for a greater number of researchers and engineers, contributes to better training of professionals and makes it possible to create a large ecosystem of developers responsible for extending Yi’s application areas.
A source of inspiration ?
China shows little concern for respecting individual freedoms. Laws governing respect for private life are less protective of individuals and it is easier in China than in the West to build a dataset large enough to train a language model. This despite a GDPR type law entry into force in 2021. Manyobservers thus attribute the competitiveness of Yi to the dataset on which it was trained. This model includes half as many parameters as GPT-4 and yet the performance of the two models is comparable.
France has no shortage of serious assets in the global AI industry. France also has a significant number of AI researchers and engineers of the highest order. Many of these experts are involved in creating neural networks, learning models or languages for Apple, AT&T, IBM, Google or Meta. The financiers are also there: Xavier Niel, Rodolphe Saadé and Eric Schmidt, for example, ensured the launch of Mistral AI and more recently, in November 2023, of the non-profit laboratory Kyutai. These two organizations have committed to the open source path and Mistral AI already offers a very competitive model.
Furthermore, these two projects have the financial resources necessary for their development. Kyutai has already lifted 300 million euros and Mistral AI has just completed a financing round of 385 million euros after having raised 105 million euros for its seed. It was valued at $2 billion at the end of last year. The financial resources of the two organizations are therefore, today, greater than those of 01.AI in China.
France therefore lacks neither the skills nor the sources of financing necessary to develop solutions ensuring our technological independence. However, it does not (yet?) have an ecosystem around models open source from Kyutai and Mistral AI. In China, Kai-Fu Lee has joined a research and training institute with a venture capital company dedicated to financing AI projects. A great idea and one that we should copy?
1710495839
#KaiFu #Lee #great #organizer #open #source #China #inspiration #France