Shulou(Shulou.com)11/24 Report--

Why is there no ChatGPT in China? What's the gap between us? How much money will it take to compete with it?

"the essence is focus and determination, whether or not you have achieved AI First." Li Di, CEO of Xiao Bing, said bluntly that from the technical level, the technology of the big model does not have many secrets, and the technology of the Chinese is no less than that of Europe and the United States. China even takes the lead in some aspects, such as localization practice technology, Chinese big model and so on.

"the big model behind AIGC is a game with a lot of capital, and it's really expensive to invest in it over a long period of time." Tian Taoyuan, a senior artificial intelligence expert, told Houchang Village No. 7 that before the emergence of ChatGPT, more companies would not choose to burn their money on an unknown opportunity. "GPT3.5 training costs US $3-4.6 million, which is just the cost of calculating power, not counting the cost of talent. A total of about 375 people in Open AI, annual salary expenses are US $200 million, and AI computing expenses are US $500 million, which requires strong capital support."

However, the stunning ChatGPT and the technological qualitative change of AIGC are projected onto the vast Chinese Internet, igniting the glory and dreams, passion and ambition of AI practitioners.

Chen Dong, an entrepreneur at AI, remembers that near the cold night of the Spring Festival, in the homes of friends in Beijing, they ate pistachios and melon seeds and talked all night. The topic all night was almost related to AIGC and ChatGPT, although many people were enthusiastic about whether they understood it or not.

An investor told reporters that last year, affected by the general environment, most of the AI entrepreneurial teams were worried about their survival, and the investment team responsible for the AI track at the end of the year was still worried about how to report to LP. In just three months, AIGC has become one of the hottest investment tracks in 2023. At present, this lively and contagious atmosphere made him feel dejected in 2022.

While being amazed by ChatGPT, the domestic industry is also thinking: what is the business model of AIGC? In which directions are there more opportunities? The crisis and opportunity of the Chinese version of ChatGPT?

Most importantly, which enterprises with their keen sense of smell and enterprising energy, seize this wave of technological dividends, seek the transformation of fate, and be able to stand out and lead the next era?

(the picture is a painting by Zhang Yun AI, provided by the interviewees.) the ecology of AIGC in China is emerging: large factories are engaged in infrastructure while small factories are doing innovation. At present, there is a ChatGPT trend on the Internet all over the world. The new chat robot model, launched by OpenAI, can not only answer human questions and admit their mistakes, but also write poems, program and even help college students finish their papers. It made Musk sigh "frighteningly good" and on behalf of AIGC shouted "sesame open" to the world.

In China, the market ecology around AIGC has come to the fore.

"AIGC is not a new concept." Chen Dong said that in many 2C applications in large factories, a lot of content has been generated by AI, and there have even been advertisements that have been using AIGC technology on a large scale. At least two years ago, there was a start-up team in this direction, and it was favored and invested by first-line US dollar funds.

In his view, AIGC is not only an opportunity for big companies, but also a lot of new start-ups. A kind of company is at the bottom, giving birth to breakthroughs in large models in new scenes, production and research such as DALL E, GPT and so on produce a lot of new models; in addition, in the application layer, there will also be many new companies, including generating text, images, videos, digital people and so on.

Under the great temptation to catch up with and surpass ChatGPT, the big factory took the lead in the "battle" to catch up with the gap between the basic and large models.

On February 7, Baidu confirmed that the name of the Baidu ChatGPT project is "Wen Xin Yiyan" and its English name is ERNIE Bot. It is expected to complete internal testing in March and open it to the public. In addition to Baidu, there are Tencent, Alibaba, Kuaishou, byte beating, JD.com, NetEase, 360, Shangtang, Meitu and other big companies have invested in the field of AIGC to participate in this new round of global AIGC technology competition.

"A good big model of artificial intelligence is very expensive." In the view of Tian Taoyuan, a senior artificial intelligence research expert, under this wave of pre-training and training models, it is actually a long-distance arms race of data, calculation, and financial resources.

According to him, the basic model GPT3.5 training costs US $3 million to US $4.6 million, which is still just the cost of calculation. excluding the cost of talents, there are about 375 people in Open AI, with annual salary expenses of US $200 million and computing expenses of US $500 million. " This is really expensive, and there is no commercial return for so many years, and the parameters of the basic model are a process of continuous optimization. With the acceleration of commercialization, we need continuous training and strong capital support. It is really impossible for small companies to use it. "

However, for the future business value of the basic big model and the application of the big model, practitioners in the industry have a consensus and expectations.

"the most noteworthy aspect of AIGC ecology is the ability to build large models, which will become an important growth point of AIGC's long-term value." AI entrepreneur "web3 Sky City owner" excitedly said, "this is a matter of trading short-term investment for a thousandfold return 10 years later."

On the other hand, in industry segments, the speed at which AIGC makes people rich is also keeping people's eyes open.

On October 19 last year, Jasper.Ai, a text-based AIGC project, completed a $125 million round of financing. After completing the financing, he joined the AIGC track Unicorn Club with a valuation of US $1.5 billion. Jasper.Ai 's announcement comes a day after Stability AI successfully raised $101 million, valuing it at $1 billion.

In the vertical technology scene of AIGC, according to the mode distinction, it can be divided into audio generation, text generation, image generation, video generation and other subdivision areas, all of which have been explored by Chinese enterprises.

Text generation: Xiao Bing, Glow, Tsinghua Nine songs, Caiyun Little Dream, Writing Cat, Lanzhou Technology, Meta-language Intelligence, etc.

Image generation: Baidu Wenxin, Dream Theater, TIAMAT, Dali AI, 6pen, Picture Universe, Artflow and so on.

Video generation: clipping of Douyin, cloud scissors of Kuaishou, Shadow Spectrum Technology, Blue Standard fission, Shangtang Zhiying, Movio of Shiyun Technology, etc.

Audio generation: NetEase Tianyin, StarX MusicX Lab, DeepMusic, etc.

Game AI:rct.Ai, NetEase Fuxi, etc.

Digital man: Qiyuan world, reflection sound, etc.

(NetEase combs with science and technology)

In January 2023, Wang Sijie, director of Qiyuan World Strategy, received several phone calls from friends in the investment and industrial circles. Two years ago, his company began to invest heavily in the production and research of Digital people and game AI, and most of his friends were asking about the industry environment and cash ability. After hanging up, Wang Sijie realized that someone might make the next move. "I feel moved."

From unicorns to large factories, from users to investors, from industry to academia, the innovation of AIGC in the application layer has aroused widespread concern.

Some data show that in terms of user growth data, the number of users tested internally by TIAMAT is 20, 000, the number of dream thieves is increasing by 50, 000 a day, and 500000 users have been accumulated in the two months since it was launched. Liang Wang, a partner in Shiyun Technology, also told reporters that its video generation product Movio has been online for less than a year and has achieved good user growth figures in overseas markets.

Behind the influx of capital and the layout of Internet giants is the huge economic value and market scale of AIGC.

Gartner predicts that by 2026, 30 per cent of global organizations will have products and services for digital people and metaspheres. McKinsey predicts that spending on global metacosmos could reach $5 trillion by 2030. What is more worthy of expectation is that AIGC is gradually exploring into the real industrial production chain and applied to many fields such as marketing, games, code, life science and so on.

Especially in the primary market, the pulse of investors is getting faster.

Summer, a partner at Mingxiang Capital, said bluntly that at the end of 2021, he and his colleagues were combing out the opportunities in the AIGC field and made an investment layout. AIGC excites him about three things: first, the upper limit of the user-in-the-loop and end-to-end data-driven model is extremely high, and model iteration + high-quality feedback can bring continuous improvement; second, the logic of content production and redistribution represented by Douyin and Xiao Hongshu may become the logic of users and AI system to jointly create personalized content; and third, a new generation of AI technology paradigm will have the opportunity to truly commercialize human mental work.

"the next era has begun."

At the end of last year, Xie Menghui, managing director of Zichun Capital, observed another increase in the speed of this wave of entrepreneurial trains. On one side, "the invested companies in the related areas of our previous investments, especially those based on the ChatGPT model for commercial derivatives, have been frequently invited to participate in various technology and investment forums."

To his surprise, this iteration of technology will promote the creation of new increments in business. "on the basis of the new productivity, the existing relations of production are bound to be reconstructed, the corresponding economic system and organizational model are bound to change, and the valuation of creativity will increase exponentially."

What is the business model of AIGC?

Chen Dong, an entrepreneur born in 1988, realized that he was in history. For many days in a row, the news that big companies had announced their entry into the chatGPT swept his attention. The entrepreneur's circle of friends was like a thermometer in the venture capital world. "people who used to do meta-universe, web3 or even NFT had to do AIGC." He and his friends lamented that it was crazy. "A big guy like Wang Huiwen issued an artificial intelligence entry manifesto on WeChat moments, saying that he would bring 50 million dollars into the group, regardless of position, salary and title, and asked to form a team."

Although Chen Dong has not yet found a clear direction, he has had coffee with investors. "they actually want to know more about the industry, especially how to commercialize it."

Since the open source of AI painting tool StableDiffusion, a large number of domestic websites and App focusing on AI painting have sprung up like bamboo shoots after a spring rain. Zhang Xun, an AI painter who is a heavy user of AI painting tools, will actively come to the door during the internal testing phase of some platforms to ask him for comments and Bug. "some front-end developers even come directly to ask, how to adjust in order to improve people's willingness to pay?"

(the picture is a painting by AI Zhang, provided by the interviewees.) for AIGC, there are still a series of really serious and disturbing questions in the industry and investment circles: what is AIGC's business model? Where are companies like open AI in China? How to make a product, not only universal, but also deep enough and wide enough?

Drawing lessons from the commercial experience of the United States, Summer roughly divides the field of AIGC into three types of business models: one is OpenAI, a company that focuses on growing models, which provides API, and companies that do scene segmentation applications need to call its API and pay by API. The second category of companies are companies that not only build large models, but also directly do vertical integration applications, such as Midjourney, which generates very popular images in AI, and the third kind of companies call the large model API of the first kind of companies. Based on a deep understanding of the subdivision of the scene, they focus on developing AI applications for specific scenarios, such as Jasper.

But whether a business model like OpenAI, which provides API, can really work in China is still a question mark.

Tian Taoyuan said that this direction not only has to bear huge labor costs and computing expenses in the R & D stage, but even if it does launch a large model similar to ChatGPT, data costs, algorithm tuning and AI ethical optimization remain high. " At present, every time ChatGPT answers a question, it costs at least one cent. To do business cash in this way, it is faced with high costs, massive demand for industry optimization, and multiple decision-making choices for science and technology strategic paths. "

At the same time, ChatGPT is similar to the first-generation iPhone, there is no successful popular model App (such as Angry Birds), it is still far from a stable commercial service, has not crossed the industrial red line. In his view, it is still in the early stage of moving from basic laboratory technology to popular industrial application services. "but at least it shows us the huge cross-domain service capability of the basic model, and finds a possible way for general artificial intelligence to succeed."

At the same time, there are very few companies that can really do the big model of AI in China.

Summer observation, anyone who can do the underlying model, whether it is to do a large language model or reinforcement learning, as long as the technical expertise, engineering product thinking as the underlying model of the company, financing is very good, in the past year or two VC have to rush to invest. "the capital premium may not be as exaggerated as in the United States, but there are still some companies that can achieve a valuation of $1 billion in a year or two."

Some investors confirmed to reporters that at present, many entrepreneurs are entering the vertical scene to solve the needs of subdivision, especially the game AI and Digital Life into these two comprehensive AIGC scenarios, there has been a mature commercial paradigm.

Wang Sijie and Qiyuan World team are now experiencing the busiest days since their start-up: due to the early entry, the accumulated capabilities of the whole system from algorithm to engineering have built large models of game AI and digital human AI with tens of billions of parameters. After the tuyere came, top game producers have taken the initiative to seek cooperation.

He prides himself on the fact that the training and reasoning costs of this vertical model can be an order of magnitude lower than that of the Open AI model of the same parameter scale. In Wang Sijie's view, first build a smaller vertical model (such as 10 billion parameters, 1 billion parameters), so that the data flywheel and model training can be well combined, and the vertical model may be more effective and cheaper than open AI in some areas. "this kind of imagination is more successful in China."

Xiao Bing CEO Li Di said bluntly that the biggest AIGC commercial return track is digital people. "its customer unit price will obviously be higher. We have done terminal content generation, and we feel that this is a vegetable market-like content market, and the unit price is very low. Until we bind the AI generation capacity to digital people, the average customer unit price has rapidly increased from 200000 to 3 million. Of course, the quality of your finished products should be high enough."

But for entrepreneurial teams that only use API calls to generate content, most of the respondents to this interview said, "the value is limited."

Li Di tried to send a warning signal that entrepreneurs who want to enter the AIGC track should be careful to use the tunable interfaces provided by large manufacturers and build their own vertical models as much as possible.

"the business model that API invokes doesn't make money, so even for development, big companies are bound to jump in and do what other ecological companies on the ecological chain do, first of all those who directly apply their technology to provide services." He judged the final result, "if you do it, you will be eaten, and there are no bones left."

Gap between China and the United States: not in technology, while focusing on being amazed by ChatGPT, the domestic industry is also reflecting: why is there no ChatGPT in China?

As a matter of fact, there was no lack of domestic companies with ambitions to become the "throne" of the Chinese version of ChatGPT. However, in the face of the possible future and the irrefutable reality, and in the face of the conflict between the huge investment and the performance statement, we chose to hesitate.

"the technology of the big model doesn't have many secrets." Li Di said bluntly that before the emergence of ChatGPT, Xiao Bing began to do AIGC-related things in 2017. from a technical level, the technology of the Chinese is no less than that of Europe and the United States. China even takes the lead in some aspects, such as localization practice technology, Chinese big model and so on.

The person in charge of Wenxin also told the reporter that compared with similar overseas products, the company actually has advantages at the technical level in terms of cross-modal generation ability, Chinese knowledge and understanding, and drawing efficiency. For example, he said, "in terms of Chinese knowledge and understanding, we will have a better understanding and characterization of the Chinese text, culture and style."

In fact, in China, a group of bravest young entrepreneurs with the wildest dreams have challenged overseas markets with their AI-generated products and matched their product capabilities with world-class listed companies.

"our main product now is called Movio, which is an AI video creation tool, mainly for the global market, starting from Europe and the United States." Liang Wang, a partner in Shiyun Technology, introduced to reporters the reasons for choosing to go out to sea. "first, the acceptance, willingness to pay and market volume of SaaS in Europe and the United States are much higher, and second, their expectations and standards will be much higher, just like participating in the world's top competition in the field of SaaS, we intend to choose the market with the highest standards and accept the eyes of the strictest users, so as to prove that the Chinese team can also make the world's leading SaaS products."

"for AIGC to really hit the ground, the threshold is not only the technical level, but also a flywheel that can be recycled from AI to users to data, and then to the product experience." From the feedback from overseas users, he judged, "in this direction, we are now in the lead in technology and product experience."

Compared with the teams at home and abroad, the biggest feeling of Liang Wang is that there are differences in their understanding and investment in innovation. For example, in the face of innovations that may not be rewarded in the short or even medium term, does the team still have the spirit and determination to do it anyway?

As the leader of the enterprise, Li Di is also deeply aware of the differences in the business and technological environment between China and North America, and most domestic enterprises do not have so much money to burn on an unknown prediction. "it takes a lot of money, it takes a lot of time, it's hard to have commercial value in the short term, and no one knows if this is feasible."

"the essence is focus and determination, whether or not you have achieved AI First." Li Di said bluntly that they were fully committed when they saw Open AI several years ago, which is the only difference. Open AI spent enough time and energy betting on a big model of violence and miracles, but their bet Xiao Bing really dared not bet, "between life and death, in a sense, if Open AI does not launch Chat GPT, they may be very worried today."

According to Tian Taoyuan, this incident gives you an inspiration: we attach great importance to artificial intelligence hardware (such as chips), but the basic software of artificial intelligence (such as large models, training framework, big data tool chain), is it long-term insufficient investment and insufficient attention? Are you determined to invest in long-term basic R & D?

Chinese investment and industry are more concerned about the performance of short-term business returns, that is, fast money. " In China, we all see which technology has reached the critical point of the transformation of commercial value. For many leading enterprises in science and technology or physical industries, AI does not pay enough attention to basic scientific research, lacks the ability of self-research and strategic determination of large models, and is oriented by the application of rapid output, while the research and development of technological algorithms must take a long time to accumulate. " Tian Taoyuan said.

However, the entry of some Chinese AI companies that are very good at engineering is bound to stir up the global AIGC market.

China's excellent AI start-ups have a very clear advantage, that is, they have a strong engineering ability, and it is possible to build a model with lower cost, higher iterative efficiency and better meet the requirements of clear optimization objectives than openAI.

At the same time, the whole application scene in China is rich enough. " With such a large population base and the initial foundation of industrial digitization, China has an advantage in the quantity of high-quality feedback, which is particularly important for the new generation of AI models. The AIGC company that we value should have better engineering ability and have its own unique understanding of the actual business scene in China. We think that only such a company has the best chance to get out. " Summer talk.

The crisis and opportunity of the Chinese version of ChatGPT with the explosion of ChatGPT, the technology born in San Francisco has detonated the global artificial intelligence industry. AI has almost entered the game, and the capital market at the top of the economic chain has also been shaken, and the flares indicating change have exploded one after another.

(the picture is a painting by AI Zhang, provided by the interviewees.) "China must have its own GPT-4, that is, the basic model (big model)."

Tian Taoyuan stressed, "China's technology entrepreneurship team should have a sense of crisis." Whether ChatGPT or basic model, the more developers, more diversified high-quality data, and usage scenarios, the more valuable the feedback and optimization of the basic model will be. "whoever does reinforcement learning first will have a first-mover advantage, which is a great motivation and pressure for us."

Lanchi Venture Capital affirmatively told reporters that the content itself is very regionalized and geographically, and it is also closely related to user data, so a higher standard of AIGC should have a deep understanding of the local culture. The Chinese market itself is big enough. I believe that there will be great opportunities for the Chinese market to produce content through AI.

On the other side of the opportunity is the biggest threshold for all ChatGPT-"security and compliance".

In 2016, Microsoft made a chat robot, which can respond to users in real time and modify its own answers, but it was taught to be racist 24 hours after it was released and was quickly removed from the shelves. "our lab also has its own conversation robot, which was afraid to release in the past." A head AI company founder told the "Houchang Village 7" column.

Talking about the specific landing, senior artificial intelligence researcher Tian Taoyuan put forward: "in the Language Models are Few-Shot Learners report released in July 2020, OpenAI officials reminded us of the capability boundaries and challenges of the language model behind ChatGPT:

1. Expression problems: semantic repetition, loss of coherence, self-contradiction in long paragraphs, and occasionally illogical

2. Lack of common sense in physics: it's hard for GPT-3 to answer, "will cheese melt if I put it in the fridge?"

3. Poor training efficiency and world model: there is a significant gap between GPT-3 and human language training efficiency, but the text size of GPT-3 during pre-training is much larger than that seen in human life. Based on additional information in the physical world, algorithm improvement may improve efficiency.

4. Uncertainty: GPT-3 's small-sample learning has uncertainty and fuzziness. In the future, on the basis of accurately understanding the principle of small-sample learning, we can "really start from scratch" to learn new tasks in model reasoning. (it is not clear whether human "learning from scratch" will be affected by past knowledge.)

5. The applicability of the large model is not good: it is not convenient to reason, the practicality is not good, and the large model is distilled to a manageable size. Large models such as GPT-3 contain a wide range of skills, but they are not needed in specific tasks. In principle, radical distillation "slimming" is a possible method, but it has not been tried on a large model with hundreds of billions of parameters. May encounter "decision-making is not easy to explain", "input prediction is not easy to calibrate", "training data bias" (data bias brings model bias, social and ethical impact). "

"I see that people like Open AI are now jumping out and taking the initiative to say they want to be regulated."

In the view of the founder of the above-mentioned AI head company, the final landing of this big model is bound to be subject to very strict supervision. "if left unchecked, it can greatly reduce the educational level of that generation of American students in a decade."

In his view, the real circle of ChatGPT in C is inseparable from the students who want to copy homework, which involves the ethical issues of AI. "in the United States, nine out of 10 students quickly learned about ChatGPT in the past few months and handed in their homework through it, which is an important driving force for it to be popular. It solves practical problems, but this solution is incorrect."

He is not interested in the so-called weather forecast in the capital market. During this period of time, there is no lack of investors to come over and propose to build a large model similar to ChatGPT, but he did not accept the olive branch. "I'm not particularly fond of stealing the limelight, I won't launch a thing for students to copy homework, no matter how much money they make."

He hopes that peers who choose to do similar ChatGPT applications will consider more about the self-discipline of AI ethics and predict in advance what students will do when they get it. What if it is rolled out quickly? "try your best to be safe and compliant, so that it can last for long and continue to exist."

On the issue of safety and compliance, the person in charge of Baidu Wenxin told reporters that he hoped to clarify the compliance rights and responsibilities of all parties, taking into account content security and business innovation and development, and suggested that landing rules should be formulated based on different subdivided application scenarios and product model characteristics. in the future, leading companies in the industry can be encouraged to take the initiative to share practical experience in innovative compliance and promote the formation of industry standards. To create a practical scheme of industrial compliance that can be used for reference.

There is a consensus in the industry that the discussion on ChatGPT will not stop in the next few years.

But what is clear is that the development of ChatGPT and its corresponding big model, a technological revolution in the future, may not be as frightening as you think, but it will release the innovation of the next few years.

In Tian Taoyuan's view, people are now obsessed with ChatGPT, looking at the whole future of artificial intelligence, which is just a stepping stone to move forward. "further up, there are more huge opportunities."

In his imagination, AI can promote the exploration of space flight, the exploration of new energy and the discovery of new materials. With its more thorough understanding of the laws of physics, it has a deeper and deeper understanding of people, environment and life, and can approach some unknown knowledge beyond the scope of human cognition, and even has the possibility of new knowledge.

In this great change which may be more subversive than the mobile Internet, for Chinese AI enterprises, there is still a long way to find, explore and climb.

(Chen Dong, an AI entrepreneur, is a pseudonym at the request of the interviewees.)

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.