Author: Xiaoyan

Editor: Xiaodi

Source: New Spark

If we rewind time by 5 years, we will find that there were only a few groups worldwide capable of developing large AI models. It's worth noting that the computing power required to train large models has only been available for the past three years, and the emergence of the phenomenon-level product ChatGPT has only occurred in the past year.

After all, the development of large models requires a high threshold, not only requiring powerful computing resources, but also requiring abundant data resources to support them. In the era of large models, it is inseparable from large computing power and large data.

However, in just one year, various companies and institutions specializing in large models have sprung up like mushrooms after rain. Everyone has started to compete, and then they have become embroiled in intense competition. Even the renowned AI scientist Kai-Fu Lee has stated that only 5 or 6 companies in the United States and China combined will survive in the large model industry.

We can't help but sigh at how the large model race has gone from "few competitors at the top" to severe competition in such a short time.

"Apart from the big players, the United States and China combined should be able to support around 5 or 6 companies."

On December 14, during a conversation related to AI, Kai-Fu Lee, the CEO of Sinovation Ventures, was asked a question: how many companies will ultimately survive in the large model race? In response, Kai-Fu Lee also spoke frankly, believing that apart from the big players, the United States and China combined should be able to support around five or six companies.

Kai-Fu Lee stated that it is becoming increasingly rare to develop huge pre-trained models. This channel cannot be said to be completely closed, but it will inevitably become more difficult. In the future, more opportunities will only emerge when new technologies are needed. At the same time, Kai-Fu Lee also outlined the future direction of AI: AI Infrastructure and AI applications.

Kai-Fu Lee explained, "The opportunity to develop AI applications is now, just like the era faced by mobile internet more than ten years ago. WeChat, which seized the opportunity early, succeeded. Of course, there are also subsequent successes like Douyin and Pinduoduo. But the earlier you enter the game, the greater the opportunity. Developers with real dreams and ambitions should focus on creating AI-First and AI-Native applications, as these applications are likely to become the greatest or most profitable applications of the AI 2.0 era."

Indeed, this is the case. The competition in the development of professional large model AI is fierce. When training their own large models, major companies also face significant challenges and limitations. On the one hand, training large models requires massive computing resources and time. If one is not a major player or a unicorn with strong capital support, this will inevitably become an insurmountable barrier. On the other hand, training data for large models requires a large amount of manual annotation and cleaning, a highly specialized, complex, and time-consuming process. More importantly, the application of large models faces challenges in data privacy and security. "How to protect user data and prevent misuse" is a problem that every large model developer needs to consider and urgently address. A slight misstep can lead to a quagmire.

Kai-Fu Lee leads Sinovation Ventures to release the first open-source bilingual large model "Yi".

Although emphasizing that the development of AI large models is turning into a "red ocean race," Kai-Fu Lee himself is the most aggressive player in the race.

As the chairman and CEO of Sinovation Ventures, he founded his own AI company, "Sinovation Ventures." Just last month, Sinovation Ventures released its first open-source bilingual large model "Yi."

The Yi series models released this time include two versions, 34B and 6B. On November 2, Sinovation Ventures uploaded these two parameters to Hugging Face. It is reported that Hugging Face is the world's most popular open-source community for large models and datasets, and is considered the GitHub of the large model field, with considerable authority in English language ability testing for large models.

According to the latest rankings provided by the Hugging Face English open-source community platform and the C-Eval Chinese evaluation, Yi-34B has climbed to the top of the pre-trained large language model and Chinese large model rankings on the C-Eval leaderboard. This is the only domestically produced model to have successfully topped the global open-source model rankings on Hugging Face to date.

At the same time, Sinovation Ventures' new round of financing has been very smooth, led by Alibaba Cloud. Currently, Sinovation Ventures is valued at over $1 billion, making it a unicorn. Kai-Fu Lee established the Sinovation Ventures team in March 2023 and began operations in June. In just 8 months, Sinovation Ventures has not only launched its core product, but has also become a "unicorn" valued at over $1 billion.

It's no wonder that in an industry filled with major players and unicorns, Kai-Fu Lee would lament the fierce competition in the large model race.

Yi, the leading global large model, aims to create more To C Super Apps.

It is understood that Yi-6B and Yi-3B represent data parameter sizes of 60 billion and 340 billion, respectively. According to Kai-Fu Lee, "34B is a golden size."

The biggest advantage of 34B is that it is not so small that there is no emergence or insufficient emergence, and it has already fully reached the threshold of emergence. At the same time, it is not too large, allowing for efficient single-card inference. The Yi-34B model leads in multiple benchmark tests globally, with training costs under the super Infra dropping by 40%, and simulated training costs at the scale of hundreds of billions dropping by as much as 50%.

The so-called AI Infrastructure, or AI Infra for short, mainly covers various underlying technical facilities for large model training and deployment, including processors, operating systems, storage systems, network infrastructure, cloud computing platforms, etc., and is an extremely critical "guarantee technology" behind model training.

AI Infra is a hard technical field that has received relatively little attention in the development of the large model industry, but it is also a very critical area. In Kai-Fu Lee's words, "People who have worked on large model Infra are rarer than those who work on algorithms."

For Sinovation Ventures, with Yi as the foundation, it is only possible to create more To C-end super applications. Yi is positioned as a general base and has officially launched on three major global open-source community platforms: Hugging Face, ModelScope, and GitHub. At the same time, various versions of the Yi series, including quantized versions, dialogue models, mathematical models, code models, multimodal models, etc., will soon be unveiled.

According to Kai-Fu Lee's logic, the technological barriers of AI will eventually be broken one by one. For AI companies, sustainable and profitable growth is the most important thing, and companies that cannot be commercialized will eventually be eliminated. In the AI 2.0 era, the biggest business opportunities will undoubtedly come from super applications, and they will be at the consumer level, such as super applications like Douyin and WeChat.

Currently, Sinovation Ventures has already initiated training of models with parameters exceeding 100 billion, and the multimodal large model team has also assembled more than 10 people, indicating that Sinovation Ventures will strive towards "consumer-level applications." This coincides with Kai-Fu Lee's belief that "AI Infra and AI applications are the future."

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。