Kai-fu Lee's large-scale model company Zero One Wanwu is online, and dozens of core members are in place

Original source: Qubit

Image source: Generated by Unbounded AI‌

Who is the most competitive large-scale model entrepreneurial dream team in China?

Zero 10,000 things.

That's right, at least starting today, this name will open a new stage of being noticed, discussed, and tracked.

There is no other reason - this is the AI2.0 company incubated by Dr. Kaifu Lee who "rolled up his sleeves" and led the team personally. In late March this year, Kai-fu Lee, Chairman and CEO of Sinovation Ventures, officially announced the entry of the large model under the name of Project AI2.0... and 01Wanwu is the official company name of this start-up company, the official website 01. AI is also online today.

Qubit learned that after nearly 3 months of preparations, 010,000 things have also made a series of progress, including:

**Company: **The headquarters is registered in Beijing, but it recruits top talents from all over the world.

**Route: **Self-developed large model. Self-built platform, build thousands of GPU clusters for training, tuning and reasoning, increase the amount of effective data parameters, and the density of high-quality data used to build continuous, controllable, and scientific training capabilities.

**Progress: **The internal test of the model with a scale of 10 billion parameters has been realized within three months, and it is expanding to a scale of 30 billion to 70 billion parameters.

It is reported that within just 3 months, 01Wanwu has assembled a team of dozens of core members, focusing on large-scale model technology, artificial intelligence algorithms, natural language processing, system architecture, computing power architecture, Data security, product development and other fields. Among them, the members of joined joint creation team include former vice president of Alibaba, former vice president of Baidu, former head of algorithm of Didi/Baidu, former executive of Google China, former vice president of Microsoft/SAP/Cisco, The algorithm and product team backgrounds are from major domestic and foreign manufacturers.

Since returning to China to found Microsoft Research Asia 25 years ago, Mr. Kaifu has become the most famous mentor and Bole in China's computer industry, objectively boosting the development of China's Internet and AI.

Now, standing at the turbulent threshold of large-scale models and the AI2.0 trend, former students, former departments and admirers once again follow and gather.

01Wanwu is the seventh company** incubated at the top of Innovation Works AI Engineering Institute** founded by Kai-Fu Lee. This special approach combines fund investment and peak incubation to accelerate the creation of a new AI 2.0 entrepreneurial ecosystem.

010,000 things, self-developed general model

According to public industrial and commercial information, Beijing Zero One Wanwu Technology Co., Ltd. (hereinafter referred to as Zero One Wanwu) was established on May 16, and it was exactly two months before Lee Kai-fu’s circle of friends announced that he would personally enter the large-scale model business.

**What is the explanation of "zero and one thousand things"? **

Zero one is not difficult to understand, the underlying rules of the computer are binary, and 01 is the native language of engineers. However, after further hearing about the qubit, in Kaifu's view, 01-01 also represents the entire digital world, from zero to one, and even all things in the universe. can" ambition.

Industry and commerce information also shows that the business scope of 01Wanwu covers artificial intelligence mass innovation service platform, artificial intelligence basic resources and technology platform, engineering and technology research and experimental development, data processing services, etc.

This range is also consistent with Kaifu's thinking and judgment on AI2.0. After ChatGPT drove the upsurge of large models, Kaifu Lee publicly stated that the era of AI 2.0, which is based on the large model of the base, will set off multiple technologies, platforms and applications. level of revolution. Just as Windows drove the popularization of PCs, and Android gave birth to the ecology of the mobile Internet, AI2.0 will create a platform opportunity ten times larger than that of the mobile Internet. It will rewrite existing software, user interfaces and applications, and will also create a new Approve the application of AI-first and give birth to an AI-led business model.

More details about the business of 01Wanwu are also disclosed:

The concept is AI-first, the driving force is technical vision, backed by the excellent Chinese engineering background, the breakthrough point is the large model of the base, and the coverage includes technology , platform to application multiple levels.

More specifically, the large-scale model entrepreneurial route chosen by 01Wanwu is self-developed large-scale model.

And it has already kicked off-the company has started to build an experimental platform, built a GPU cluster with thousands of cards, for training, tuning and reasoning. In terms of data, the focus is on increasing the amount of effective parameters and using high-quality data density.

From the announcement to the appearance, in the past three months, 01Wanwu has completed the internal testing of the model with a scale of 10 billion parameters, and is expanding to a scale of 30-70 billion parameters. In the future, it plans to launch an open source model version.

It is reported that after launching the general-purpose large model, the next step of 01Wanwu is to launch a multi-modal model that combines pictures, videos, and 3D**.

The entry-level Super App is also part of the plan, and platform middleware and developer tools will also be released gradually.

At the same time, there is also the "seven major module technology routes" of 01Wanwu in the AI 2.0 era.

In short, 01Wanwu believes that in order to make a large model platform, it is necessary to recruit top scientists and engineers in seven directions, namely:

• Define the pre-training framework (Pre-train) technology of model capability ceiling
• Post-train technology to improve model realization ability
• AI Infra technology supporting efficient model training
• More multi-media multi-modal model technology
• Practical training scale technology
• Platform middleware and tool chain technology
• Inference and service deployment technology

Based on these 7 technical modules, 01Wanwu recruited a technical team in a targeted manner.

Qubit learned that the team currently has dozens of core members in place. The backgrounds of the advanced algorithm and product teams are from major manufacturers at home and abroad. They focus on large-scale model technology, artificial intelligence algorithms, natural language processing, system architecture, Computing power architecture, data security, product development and other fields.

There are elite soldiers as well as strong generals. Judging from the lineup of the Lianchuang team that has been confirmed to join, it looks like an "AI 2.0 dream team":

Former Alibaba Vice President, Former Baidu Vice President, Former Didi/Baidu Chief Algorithm Leader, Former Google China Executive, Former Microsoft/SAP/Cisco/Vice President...

It can be said that the core team has been deeply involved in the construction of the world's largest software platforms.

China's large-scale model evolution is divided into 3 stages, and the vision of the spire can make the best use of it

Since the start of the spire hatching of Zero One and Everything, Teacher Kaifu has also started the intensive preparation mode, which is very busy and low-key.

But the qubit also learned its latest thinking and judgment on the evolution of the large model. In Li Kaifu's view, the evolution of China's large language model will not be completed in one step, but will go through three stages from L1 to L3.

At this stage, most companies in China are in the L1 stage. The next step is to cross the "emergence" threshold and become a mainstream L2 language model, and finally catch up with L3 global first-tier companies and become a powerful AI 2.0 "technical infrastructure" to achieve high-quality, high performance goals.

The core reason why 01Wanwu chooses the road of self-developed large models lies in the technical realization of open source models and small models, which absolutely cannot achieve the same performance as the self-developed models of overseas major manufacturers.

Judging from the actual experience at this stage, many open source models that claim to be similar to ChatGPT are limited to simple dialogues, and are limited by the structure and size of the model itself in tasks such as long text generation and reasoning, and cannot adapt to slightly generalized or complex tasks. .

In addition, within 01Wanwu, it is also determined that from the perspective of medium and long-term development, excessive reliance on models means that only the results can be known, and the process is not grasped, "knowing what is not knowing why." The ceiling and boundaries of one's own ability are defined by others, and the ceiling of ability to achieve a variety of tasks will soon be encountered. Domestic and foreign conditions, cultures, user habits, and laws and regulations are also different. Overseas open source technologies are subject to risks such as shutdown, switching to high fees, or restricting regional access at any time.

Only by fully combining Chinese language data can the development of AI2.0 be independently controlled.

However, 01Wanwu also revealed that it plans to launch an open-source model version in the future to attract outstanding talents to jointly improve the model and promote the global AI 2.0 developer ecosystem. Therefore, large-model startups should embrace open source, but not overly rely on the open source model. Only by establishing our own technological advantages as soon as possible and building a strong ecological moat can we truly seize the historical opportunity of AI2.0.

Li Kaifu believes that the diversification of large-scale model companies will give birth to a wave of innovative ecology from AI 2.0 models, AI 2.0 infrastructure, and new AI 2.0 applications. "We predict that once China has a truly original, high-quality large-scale model, High-quality and creative applications will flourish—just like in the era of mobile Internet and AI 1.0.”

In the global market, there will be a few large-scale companies with abundant resources to support the demand for computing power, which will drive the prosperity and development of various downstream applications, gradually improve the new ecology of China's AI 2.0, explore new business models, and create huge economic value. and social value. "If the market is calm or if there are only a few players, then we really need to worry about being left behind."

Kai-fu Lee also believes that the AI2.0 market is huge and can accommodate giants, small and medium-sized enterprises, and start-up companies competing in three levels of companies.

But "the technical threshold for large models is very high, and it is very necessary to integrate an efficient team with both scientific research capabilities and development capabilities, formulate and implement the company's technology and product strategies, and implement them consistently from top to bottom."

From an objective point of view, China may be a few years behind in R&D breakthroughs, but it is good at technological iteration and has excellent engineering capabilities and execution speed. The future can be expected. Compared with big companies, start-up companies with top technology and able to focus on mastering large-scale models have the advantages of concentrated resources, flexible strategies, tough battles, and strong implementation. On the contrary, they have a better chance of leading the industry, just like the OpenAI company behind ChatGPT Same.

And this may also be the core reason why Kai-fu Lee chose to enter the game in person and incubate all things at the top of the tower—**The entrepreneurship of the self-developed large model base requires not only technology and talents, but also the integration of resources at the top and the cognitive vision of the top . **

It is only when Kai-fu Lee bows himself into the game that Zero One Wanwu appears so different. Among the many innovation echelons, 01Wanwu has significant advantages and competitive characteristics. Kai-fu Lee has more than 40 years of international AI experience, and personally led the organization. Top AI experts from around the world joined in one after another. The core team currently formed has been deeply involved in the construction of the world's largest software platforms, and has a deep understanding of the platform's successful methodology.

"AI is the interpretation of human learning process, the quantification of human thinking process, the clarification of human behavior, and the understanding of human intelligence... AI is the last mile for human beings to know and understand themselves. I hope to join this new blooming , a promising future scientific field".

This is Kai-fu Lee's application letter for a Ph.D. at Carnegie Mellon University in 1983. As one of the pioneers in the field of artificial intelligence, he has been active in the frontline of artificial intelligence research and industry for 40 years.

In 1998, Kai-fu Lee founded the Microsoft China Research Institute (the predecessor of Microsoft Asia Research Institute MSRA), which has become the "AI University" that trains the most artificial intelligence talents and outputs the highest quality talents in China. The young talents at that time are still shining in the tide of artificial intelligence.

It is also this time-tested result that has made Dr. Kai-Fu Lee gradually become the AI Pioneer passed on by word of mouth. It is also this time-tested result that has made more people who are full of faith in AI 2.0 and AGI Young engineers respond to the call to seize opportunities in the technological era and change the world.