How to enclose land, build houses, and attract investment in the era of large models?Interview with

Mondo Technology Updated on 2024-01-30

Smart stuff.

Author |Cheng Qian.

Edit |Desert Shadow.

In the 2023 global large-scale model competition, GPT-4 deserves to be ranked.

1. With a perfect score of 100 points, Li Tao, chairman and CEO of APUS, gave his own large model a score of 80-85.

As one of the successful enterprises in China's Internet industry that rely on mobile applications to go overseas, APUS is also one of the earliest enterprises in China to enter the large-scale model track, and it is also the first batch of enterprises to launch large-scale models with 100 billion parameters.

In April this year, when APUS released the APUS large model with a parameter scale of 100 billion parameters, domestic large models continued to emerge like mushrooms after a rainNowadays, large-scale model companies have begun to fight for applications and services, and OpenAI, the top model of large-scale models, also threw out blockbuster updates at the first developer conference, including updating general large-scale models and allowing users to customize GPT ......Behind the continuous increase in large-scale model competitions, more and more new challenges are facing domestic large-scale model players.

Since December last year, APUS has been deeply involved in the large model for a year, and APUS, which has a first-mover advantage, has accumulated more than 2 billion users for its 200+ products and services launched before, plus the cheap computing power at the beginning of the year, and has obtained rich computing resources. Today, based on the APUS large model, APUS has begun to build an ecosystem and make applications, and has been implemented in the fields of smart healthcare, smart e-commerce, and smart government affairs.

When we look at the entire industry, what new challenges will domestic large-scale model enterprises face?The entry point for building an ecology and providing services for large models is in the first placeWhat does OpenAI's inaugural developer conference mean for the industry?

With these questions, Zhidong had a face-to-face communication with Li Tao, chairman and CEO of APUS, in the APUS office in Wangjing, Beijing, trying to open a new chapter in the application of large models of domestic players from the reason and confidence of this Internet company All in AI.

Li Tao believes that on the surface, there is only a gap of 15 points and 25 points between the APUS large model and GPT-4, but it takes N years of hard work behind this. This includes not only the first-mover advantage of APUS being the first to enter the game, but also the in-depth thinking on the first scenario of the landing and application of the large model and how to align the values of the large model. APUS, which has experienced almost every link in the development of the domestic large-scale model industry, has also seen many industry truths.

Li Tao, Chairman and CEO of APUS.

In April this year, the multi-modal APUS model with a scale of 100 billion parameters was unveiled, and text, image, and audio models were distilled for more subdivided user needs, and large models of medical, e-commerce, network information, education, manufacturing and other industries were launched for the industry.

There is a key reason behind thisLayout earlyWhether it is in contact with large models or in the application of large models, APUS is ahead of the curve.

In 2014, when APUS was just established, Li Tao decided to "go to sea". APUS's first product, APUS Launcher, has gained more than 10 million users worldwide within a month of its launch, and the number of users has reached 100 million in half a year.

However, one of the key things that Internet companies need to consider is that they must accept the impact of the clash of civilizations in the real world. When a number of APUS apps were banned overseas in 2018 due to geopolitical influences, Li Tao realized that they had to "walk on two legs", making both tools and content, and making some apps that users relied heavily on.

This also made APUS and AI bond, and quickly followed up step by step.

In 2019, Li Tao felt the powerful ability of GPT-3. When he traveled to Silicon Valley, OpenAI CEO Sam Altman demonstrated the application of GPT-3 in games for the group, which was cool, but they didn't think about the application of large models in other scenarios at the time.

The turning point was on November 30, 2022, when many young Americans poured into ChatGPT. "The day ChatGPT went live was an extremely ordinary day in China, but after that, we were inevitably caught up in the technology wave and wanted to be the leader of the technology wave. Li Tao said.

Therefore, after the release of ChatGPT in November, APUS quickly followed up with the large model as soon as possible: in December 2022, it was decided that the technical route would all be turned to the large model, and in April 2023, the entire company was all in AI. These two turning points have established APUS's first-mover advantage in the field of large models.

Panoramic architecture of the APUS large model (Source: APUS).

Li Tao said that most domestic companies actually made big models in this month. Therefore, from December to April, APUS has enough time to develop and purchase cheap computing resources to support the development of large models. He added that the computing power bought for 50 million yuan at that time may now be worth more than 100 million yuan.

After getting the computing power and developing the large model by itself, APUS decided to create a unique open source method with Chinese characteristics, that is, APUS provides funds and computing power to allow universities and research institutions to do their own projects. Li Tao said that more than 50% of APUS's own computing power is provided to third parties.

At present, APUS has cooperated with Shenzhen University, Nanyang Technological University and Tsinghua University. In this way, open source cooperation is finally through intellectual property sharing, so that more researchers can participate in the research and development of large models.

At the same time, APUS has a wealth of data. Li Tao revealed that the cumulative number of users of APUS's products and services has exceeded 2 billion. A major threshold for the development of domestic large models is data, the global corpus of Chinese corpus data may not exceed 3%, do global business APUS in the corpus has an inherent advantage, global users use APUS applications every day, will feedback, proofread data, improve the quality of large model data.

Some products of APUS (Source: APUS).

In July this year, APUS went all out to build an ecology and provide services in the ecology and application of large models.

In Li Tao's view, there are four main stages in the development of domestic large models.

The first stage is that the large model is widely recognized by the public. In March 2022, OpenAI released the InstructGPT large model, and at the end of November, it released ChatGPT for the public.

The focus of the second stage is computing power. "Running a large model without computing power is all nonsense. Li Tao said that until the third quarter, computing power is an eternal topic.

At the same time, everyone also realizes that the direct application of large models in production and life will face many problems, and cannot provide users with good services. Therefore, the idea of AI agent was put forward in the third quarter of this year. The industry has begun to use the large model as a low-level support similar to the operating system level, allowing it to evolve and apply on top of it.

In the fourth quarter, more and more AI applications appeared, and making applications and building ecosystems became the focus of the large model industry. Li Tao said that this can be compared to the process of enclosing land, building houses, and attracting investment for enterprises.

He believes that from the perspective of the evolution of the large model itself, enterprises need to solve four things, algorithms, computing power, data, and application ecology. The first three things are infrastructure, and the last thing plays a decisive role. "Many scientists are not good at doing ecology and application, so they can only put their research on the shelf. He said.

In April, after the release of the APUS model, Li Tao determined the company's first mission, which is to systematically transform more than 200 software and applications with the large model. At present, the APUS model has been applied in the fields of smart healthcare, smart e-commerce, and smart government affairs.

In terms of smart healthcare, AI applications such as smart triage and follow-up have been developed and launched based on the APUS medical model. After assessment, the accuracy of answering questions in the examination for licensed doctors reached more than 85%. Li Tao said that in the future, the application of AI in these fields will be more accurate than that of doctors with four to five years of work experience.

In addition, there is also the field of smart e-commerce, Li Tao said, they learned that some models need 200 yuan to shoot a ** for a piece of clothing, and 100 pieces of clothing and 20 models need to spend about 400,000 yuan. Based on the APUS e-commerce model, the AI innovative e-commerce "marketing service" application eShop can be built to achieve functions such as commercial shooting within 1 minute, model customization, and post-product effect change at any time, and this cost can be pressed to thousands or tens of thousands.

For end users, APUS has launched a variety of AI products such as Intelligent Q&A Master, Stick Figure, and Ink Dyeing, and has upgraded the original applications with AI.

For the first scenario of AI application, Li Tao believes that small B and big C will be the first scenario of AI application, that is, bloggers or KOLs (key opinion leaders) who bring goods on social platforms such as Xiaohongshu and Douyin with the Internet as a platform. The reason is that these users are very sensitive to efficiency, and AI can help them solve the problem of human input, and even help them manage a company.

In addition, finance is also a typical case, he explained, in the short term, the financial field has high labor costs and high output value, and AI applications can quickly reflect the value.

Compared with the world, the development characteristics of the domestic large-scale model industry can be summarized as "the same frequency is not synchronized". Li Tao said that the development of domestic large-scale models is probably a few months behind. In his opinion, on November 7, the first developer conference of openai, a popular fried chicken abroad, presented the latest development trend of the large model industry.

The first is GPT-4 Turbo, which provides the underlying capabilities, which is stronger, cheaper, lower development cost, and fresher than GPT-4, which can empower the application and ecology of large models.

Looking ahead, GPT-4 Turbo will "monopolize" all kinds of data, and once the data flywheel rolls, the data will grow exponentially, thus making GPT smarter.

The second major release content is GPT Store, Li Tao believes that this is actually equivalent to OpenAI starting to build an ecosystem. With the release of GPT Store, OpenAI will also unify all AI applications and services into a single entrance, which is very similar to a search engine.

At the same time, OpenAI has blocked the developer entrance, and users use the ** application in the GPT Store as a plug-in, providing services in its background, "which may mean that all developer brands will cease to exist in the future."

The third big release was GPTS, he explained, "The essence of GPTS is the agent we talked about today, so that everyone can train their own agent. In other words, this agent can be the object of uploading or sustenance of the user's data and emotions, and become the entrance to the user's communication. As a result, users may have a strong dependence on their own agents in the future.

Customize the dedicated version of ChatGPT through GPTS

Once OpenAI achieves the monopoly of technology, data, and traffic entrance, this may challenge the existing social operation model. Therefore, Li Tao believes that in the current development of the large-scale model industry, people's first concern should be that in the next 10 to 20 years, there will be a challenge to the existing civilization caused by AI with powerful energy controlled by a small number of people.

Looking forward to the future development of the large model industry, Li Tao believes that large economies around the world must have their own large models and incorporate them into the existing social framework, in addition, each country and civilization must have its own AI agent to conform to its own values.

Therefore, there are two decisive strategies for APUS in the future, the first is to make its large model technology leading enough, and the second is to maximize the ecology and application of the large model, "these two can be kept at one end".

The rapid development of AI technology goes hand in hand with the battle for security and the commercialization of technology.

At the end of November, OpenAI's seizure of power became the focus of the global tech community, and a major technological breakthrough may be one of the key wishes for the CEO to be fired, reflecting the growing contradiction between the two concepts of security and technology commercialization.

At the same time, on the one hand, a number of large-scale model companies have officially announced their technological progress, and on the other hand, countries have also held high the banner of AI supervision and have begun to draft and introduce corresponding policies.

For APUS, Li Tao said that their logic is particularly simple, that is, to "customize a large model for China".

First of all, it needs to be clear that the large model made by APUS must be in line with China's values, and secondly, it must be leading enough and can really provide convenience for people's production and life. He added that it is not necessary to compare the capabilities of the model, but more important is how to make more services, applications, and a better ecology, so that the large model can provide value.

Another key to measuring is how large models can reduce the burden on people, which is also a responsibility that they must bear. On the other hand, this process will allow the large model to achieve self-restraint, making the large model smarter and smarter while making people more aspect.

He talked about his views on OpenAI's seizure of power, which is essentially a clash of ideas and values, and everyone hopes to build their own ideal utopia through OpenAI.

The overall process of the global large-scale model industry is basically the same, but from the perspective of large-scale model capabilities, there is still a certain gap in China. In the second half of this year, more and more AI applications and services appeared, and it has become a consensus to provide services and build ecosystems.

Enterprises that do the underlying large model cannot do services, and need to provide a platform for more participantsIt also needs rich domestic scenarios to provide opportunities for the landing and application of large models.

As one of the earliest companies in China to enter the large model, APUS has accumulated computing power, data resources, and self-developed algorithms, coupled with its products have more than 2 billion users, which can amplify the value of the data flywheel, thereby making the APUS large model smarter and building richer applications and services on top of it.

At the end of the day, Li Tao believes that the ultimate goal of AI is to become smarter and reduce the burden on people.

Related Pages