After ChatGPT came into existence, the conversation between human beings has become "technological".

From as far away as a classmate who is breastfeeding a child in the 18th line, he asks, "I heard that ChatGPT can write its own scripts, will you be replaced in the future?" From an investor who has handled tens of millions of transactions in the primary market to an investor who asks, "Can your product be combined with ChatGPT?" Then, at a young age, they saw an advertisement in their circle of friends and asked about the ChatGPT course to improve their future workplace competitiveness, asking, "Can I learn how to keep me from losing my job without raising my salary?"

Some people embrace ChatGPT, some people fear it, some people still do not know why it continues to be hot, some people rely on the call ChatGPT big model to get $125 million in financing, and some people take advantage of ChatGPT to make the company's ARR (Annual Recurring Revenue) to $10 million.

So is the myth of ChatGPT's "wealth creation" true? With ChatGPT not open to mainland China and Hong Kong, how can domestic entrepreneurs ride on this new wave of entrepreneurial opportunities? Will China have a latecomer advantage in the battle of the big models?

ChatGPT ignites the AI boom

ChatGPT is not the end game, it's the beginning of the detonation of AI.

Once ChatGPT was launched, users around the world rushed to it, and within two months, the number of users exceeded 100 million, setting a record for the fastest application to break 100 million on the Internet, a figure that took TikTok nine months, Facebook two years, and Instagram 30 months.

A devastating third technological revolution kicked off. The philosopher of science Kuhn once proposed that the development of science is not a smooth linear development model, but will be like walking steps, every other stage on a step, just as ChatGPT was introduced, pulling artificial intelligence on a new ladder.

OpenAI CEO Sam Altman predicted that AI big model technology, will become the biggest technology platform in the future after the mobile Internet, and the development of chatbots as the interface, together with multimodal models such as images, music, and text, will give birth to world-class large-scale enterprises.

Microsoft CEO Satya Nadella, the enabler behind this, also came out and said that the development of AI at this stage can already be described as "exponential".

The retired Masayoshi Son also publicly said that will end the relatively dormant state all in artificial intelligence. "I am most interested in the artificial intelligence revolution, I believe that humans will be surpassed by computers or artificial intelligence, we hope to become the leader of the artificial intelligence revolution!" Also said that he is a heavy user of ChatGPT and also talks with OpenAI CEO Sam Altman frequently.

ChatGPT is not the full picture of AI, but just a form of generative AI. The prediction from Sequoia Capital is that the wave of innovation in generative AI applications inspired by large models will be trillions of dollars in economic value.

The primary market is also awash in generative AI startups. On the foreign side, according to PitchBook data, venture capital investment in AIGC has grown by more than 400% in the more than two years since GPT-3 was released, reaching a staggering $2.1 billion this year.

Domestically, between November 2022 and April 2023, AIGC financing reached RMB 2.24 billion, equivalent to 41.4% of the total financing in the past two years.

After YUE, a startup gas pedal under Sequoia China, opened registration on February 28, there were 300-400 AIGC themes among 2,000 registered projects; on June 3, the Spring 2023 Startup Camp Roadshow of Qiji Creation Forum, 41 of the 60 participating companies were AI themes, doubling compared to last year, and 39 of them were related to large models. This shows that the attention of the primary capital market to the AIGC track has increased steeply.

Overall, the number of domestic generative AI startup projects and scenarios are far less abundant than those abroad.

An overseas AI entrepreneur told Titanium Media Venture Capitalist, "First, it is more convenient for overseas entrepreneurs to call ChatGPT interface, which is easy to fine-tune and develop new products; second, the open atmosphere of technology innovation in Silicon Valley allows the sky to be the limit, and product applications in overseas markets are more diverse; third, the atmosphere of open source exchange community similar to Reddit is strong. People will put their understanding of ChatGPT, and new applications up for public discussion and re-creation."

On the contrary, in China, ChatGPT is not open for registration in mainland China and Hong Kong, so how can entrepreneurs and major manufacturers get on the "third technology revolution" train?

"Money for big models, no money for application layer"

Entrepreneurship is not only about dreams, but also about the amount of money in your pocket. Building a big model arithmetic, data, algorithms, talent, capital reserves a can not be missing, but this is not the average small and medium-sized startups can afford the cost.

For example, arithmetic resources, data show that the total arithmetic power consumption of ChatGPT is about 3640PF-days (with one thousand trillion calculations per second, it needs to calculate 3640 days). It is said that 10,000 Nvidia A100 chips are the threshold of arithmetic power to make good AI big models. At present, there are no more than 5 companies with more than 10,000 GPUs in our manufacturers, among which, there is at most 1 company with 10,000 Nvidia A100 GPUs.

For example, according to the report "How much arithmetic power is needed for ChatGPT" by Guosheng Securities, it is estimated that the cost of one training of GPT-3 is about $1.4 million, and for some larger LLMs (Large Language Models), the training cost is between $2 million and $12 million. With ChatGPT averaging 13 million unique visitors in January, the initial investment cost is about $800 million, with daily electricity costs in the $50,000 range.

Costly as it may be, the Chinese version of ChatGPT is on a mission. "China will definitely have its own big language model, which is determined by the current situation." Xu Siqing said to Titanium Media Ventures.

China's big manufacturers have already fist-pumped into the state of the 100-mode war. Baidu, Ali, Huawei, Tencent, 360, Shang Tang, Jingdong, KDDI, Byte Jumping and other giant manufacturers have been involved in the research and development of large models in conjunction with their own business and strategic layout, and strive to be the first "China ChatGPT", which is recorded in the history of China's artificial intelligence.

Leonis Capital said that China seems to be the only country that develops its own model infrastructure except for the US and UK monopoly. First, four of the top 10 developers in the world in terms of number of models are Chinese companies, including tech giants such as Baidu and Alibaba, top universities such as Tsinghua University, and government-sponsored labs such as the Beijing Institute of Artificial Intelligence (BAAI); second, in addition to model training infrastructure, Chinese developers are becoming independent at the hardware level, and in response to the threat of U.S. sanctions Chinese companies are increasingly turning to homegrown GPUs, such as Huawei's Ascend 910; third, while Chinese models are about a year behind the top Western models, it is only a matter of time before they catch up.

Xu Siqing agrees: "Open AI is riding high, and their technological breakthroughs have given them an absolute lead in the AIGC field. Chinese big models are rapidly following, following the path explored by Open AI, rapidly iterating and catching up; American companies are standing more in the more cutting-edge exploration field, which is determined by the characteristics of the two countries' companies and the resources at hand."

"We also have a view that the one who gets the talent gets the model, and the one who gets the model gets the world. The models here include pendant models trained for specific scenarios, both ToC and ToB models trained and deployed for specialized scenarios. We are more optimistic about this field than big models, for the simple reason that 98% of the computing time spent on training a model is spent on pre-training, but the next three steps of supervised finetuning, reward modelling and reinforcement learning only use a small percentage of computing time. " Xu Siqing said.

"One does not have ample funds, two does not have the cutting-edge technology talent pool of small and medium-sized enterprises, early to explore opportunities in the application layer, to create some first-mover advantage, is a good choice." A Silicon Valley investor said to Titanium Media Venture Capitalist.

"I think China's big model spell arithmetic is not suitable for general entrepreneurs, understand artificial intelligence entrepreneurs have a great opportunity, because do not have to start from the most basic model brutal arithmetic, than the accumulation of knowledge, team engineering capabilities and iteration speed, on the shallow surface layer of product innovation, such entrepreneurs is a descending blow. Of course, technology is important, industry experience and the choice of application scenarios are equally important. Our seed round investment in HiDream.ai (wisdom of the future) project founder, the former vice president of Jingdong, the Canadian Academy of Engineering foreign academician Mei Tao is based on this consideration." Xu Siqing said.

In addition, Xu Siqing also pointed out that the breakthrough of AIGC technology brings three levels of opportunities to Chinese entrepreneurs. First, it is the opportunity of LLM, big factory or national team; second, it is the opportunity to do superficial application or product-driven application with the help of big model, no matter To B or TO C, there are very many opportunities, characterized by the difficulty of judging which ones are easy to win, as difficult as foreseeing which one must be successful at the beginning of APP emergence; third, it is to establish the pendant model to solve the application level problem, which is in ToB and ToC There is a very wide range of prospects, because the threshold is high and the certainty is relatively stronger. They are characterized by strong modeling ability, and can work on the model to solve the problems that cannot be solved by product innovation.

Finally, he emphasized that there will also be many opportunities for infrastructure-based services that accompany various language models and applications, such as private deployment, cross-cloud, and cross-model service tools.

Foreign big models, AI applications, as much as the carnival

In Silicon Valley, Open AI wants to stand alone, Google, Meta, Amazon can not promise.

Recently, Amazon, which has been invisible, released the Amazon version of ChatGPT, a new model named Amazon Titan. The Titan series of models are divided into two kinds, one is a text model for content generation, and the other is an embedding model that can create vector embedding for creating efficient search functions, etc.

Meta CEO Mark Zuckerberg, not according to the rules, directly hit the "low price", with open source declared war on Open AI "hegemony". On June 16, according to The Information, Zuckerberg is considering commercializing a new version of the open source large language model LLaMA, allowing other companies to use LLaMA for free to develop related AI software and allowing developers to use these products for commercial purposes.

OpenAI has announced plans to create an epic LLM "app store" that will open up all ChatGPT applications, allowing the platform to be bi-directional, allowing developers to shelve their products built on OpenAI technology, and companies to use commercially available LLMs on demand, completely capturing the AI model The platform can realize two-way connection, developers can put up their products based on OpenAI technology, and enterprises can use commercially available LLM on demand, so as to completely seize the AI models.

"The big model competition in the United States must be blossoming, and the degree of openness will become higher and higher, the history of IT development constantly proves this. From Unix to Linux, from iPhone to Android are examples. The leading open power is not enough, and the vendors who are slightly in a passive position will use openness to construct their own ecology, which is a very positive phenomenon for the industry." Xu Siqing told Titanium Media Venture Capitalist.

Looking at the application layer again, there are not a few cases that have gotten commercial validation. Stable Diffusion, a free open source startup, got $100 million in financing in just one month, and its valuation soared to $1 billion; Jasper.ai, a SaaS company that builds applications based on GPT-3 API, was founded 18 months ago and received $125 million in Series A financing, valued at $1.5 billion; Copy.ai, which helps companies write promotional copy. ARR of $10 million in only 2 years.

"There are differences in overseas projects compared to domestic projects. First, the ability of overseas underlying large models is more mature (OpenAI API); second, the commercialization environment is better, users are willing to pay for high-quality single-point tools, which is also the result of experiencing the development of SaaS industry education; third, the ecology of overseas projects is better, there is no squeeze similar to the domestic Internet majors, and there is exploration of new products and new business models in both C- and B-sides. " Shen Huaxin, head of AI research at Yun Cui Capital, said to Titanium Media.

According to PitchBook statistics, the investment circle invested a total of $1.37 billion (equivalent to about RMB 9.369 billion) into generative AI companies in 2022, almost as much as the sum of the past five years.

After ChatGPT came into existence, the conversation between human beings has become "technological".

From as far away as a classmate who is breastfeeding a child in the 18th line, he asks, "I heard that ChatGPT can write its own scripts, will you be replaced in the future?" From an investor who has handled tens of millions of transactions in the primary market to an investor who asks, "Can your product be combined with ChatGPT?" Then, at a young age, they saw an advertisement in their circle of friends and asked about the ChatGPT course to improve their future workplace competitiveness, asking, "Can I learn how to keep me from losing my job without raising my salary?"

Some people embrace ChatGPT, some people fear it, some people still do not know why it continues to be hot, some people rely on the call ChatGPT big model to get $125 million in financing, and some people take advantage of ChatGPT to make the company's ARR (Annual Recurring Revenue) to $10 million.

So is the myth of ChatGPT's "wealth creation" true? With ChatGPT not open to mainland China and Hong Kong, how can domestic entrepreneurs ride on this new wave of entrepreneurial opportunities? Will China have a latecomer advantage in the battle of the big models?

ChatGPT ignites the AI boom

ChatGPT is not the end game, it's the beginning of the detonation of AI.

Once ChatGPT was launched, users around the world rushed to it, and within two months, the number of users exceeded 100 million, setting a record for the fastest application to break 100 million on the Internet, a figure that took TikTok nine months, Facebook two years, and Instagram 30 months.

A devastating third technological revolution kicked off. The philosopher of science Kuhn once proposed that the development of science is not a smooth linear development model, but will be like walking steps, every other stage on a step, just as ChatGPT was introduced, pulling artificial intelligence on a new ladder.

OpenAI CEO Sam Altman predicted that AI big model technology, will become the biggest technology platform in the future after the mobile Internet, and the development of chatbots as the interface, together with multimodal models such as images, music, and text, will give birth to world-class large-scale enterprises.

Microsoft CEO Satya Nadella, the enabler behind this, also came out and said that the development of AI at this stage can already be described as "exponential".

The retired Masayoshi Son also publicly said that will end the relatively dormant state all in artificial intelligence. "I am most interested in the artificial intelligence revolution, I believe that humans will be surpassed by computers or artificial intelligence, we hope to become the leader of the artificial intelligence revolution!" Also said that he is a heavy user of ChatGPT and also talks with OpenAI CEO Sam Altman frequently.

ChatGPT is not the full picture of AI, but just a form of generative AI. The prediction from Sequoia Capital is that the wave of innovation in generative AI applications inspired by large models will be trillions of dollars in economic value.

The primary market is also awash in generative AI startups. On the foreign side, according to PitchBook data, venture capital investment in AIGC has grown by more than 400% in the more than two years since GPT-3 was released, reaching a staggering $2.1 billion this year.

Domestically, between November 2022 and April 2023, AIGC financing reached RMB 2.24 billion, equivalent to 41.4% of the total financing in the past two years.

After YUE, a startup gas pedal under Sequoia China, opened registration on February 28, there were 300-400 AIGC themes among 2,000 registered projects; on June 3, the Spring 2023 Startup Camp Roadshow of Qiji Creation Forum, 41 of the 60 participating companies were AI themes, doubling compared to last year, and 39 of them were related to large models. This shows that the attention of the primary capital market to the AIGC track has increased steeply.

Overall, the number of domestic generative AI startup projects and scenarios are far less abundant than those abroad.

An overseas AI entrepreneur told Titanium Media Venture Capitalist, "First, it is more convenient for overseas entrepreneurs to call ChatGPT interface, which is easy to fine-tune and develop new products; second, the open atmosphere of technology innovation in Silicon Valley allows the sky to be the limit, and product applications in overseas markets are more diverse; third, the atmosphere of open source exchange community similar to Reddit is strong. People will put their understanding of ChatGPT, and new applications up for public discussion and re-creation."

On the contrary, in China, ChatGPT is not open for registration in mainland China and Hong Kong, so how can entrepreneurs and major manufacturers get on the "third technology revolution" train?

"Money for big models, no money for application layer"

Entrepreneurship is not only about dreams, but also about the amount of money in your pocket. Building a big model arithmetic, data, algorithms, talent, capital reserves a can not be missing, but this is not the average small and medium-sized startups can afford the cost.

For example, arithmetic resources, data show that the total arithmetic power consumption of ChatGPT is about 3640PF-days (with one thousand trillion calculations per second, it needs to calculate 3640 days). It is said that 10,000 Nvidia A100 chips are the threshold of arithmetic power to make good AI big models. At present, there are no more than 5 companies with more than 10,000 GPUs in our manufacturers, among which, there is at most 1 company with 10,000 Nvidia A100 GPUs.

For example, according to the report "How much arithmetic power is needed for ChatGPT" by Guosheng Securities, it is estimated that the cost of one training of GPT-3 is about $1.4 million, and for some larger LLMs (Large Language Models), the training cost is between $2 million and $12 million. With ChatGPT averaging 13 million unique visitors in January, the initial investment cost is about $800 million, with daily electricity costs in the $50,000 range.

Costly as it may be, the Chinese version of ChatGPT is on a mission. "China will definitely have its own big language model, which is determined by the current situation." Xu Siqing said to Titanium Media Ventures.

China's big manufacturers have already fist-pumped into the state of the 100-mode war. Baidu, Ali, Huawei, Tencent, 360, Shang Tang, Jingdong, KDDI, Byte Jumping and other giant manufacturers have been involved in the research and development of large models in conjunction with their own business and strategic layout, and strive to be the first "China ChatGPT", which is recorded in the history of China's artificial intelligence.

Leonis Capital said that China seems to be the only country that develops its own model infrastructure except for the US and UK monopoly. First, four of the top 10 developers in the world in terms of number of models are Chinese companies, including tech giants such as Baidu and Alibaba, top universities such as Tsinghua University, and government-sponsored labs such as the Beijing Institute of Artificial Intelligence (BAAI); second, in addition to model training infrastructure, Chinese developers are becoming independent at the hardware level, and in response to the threat of U.S. sanctions Chinese companies are increasingly turning to homegrown GPUs, such as Huawei's Ascend 910; third, while Chinese models are about a year behind the top Western models, it is only a matter of time before they catch up.

Xu Siqing agrees: "Open AI is riding high, and their technological breakthroughs have given them an absolute lead in the AIGC field. Chinese big models are rapidly following, following the path explored by Open AI, rapidly iterating and catching up; American companies are standing more in the more cutting-edge exploration field, which is determined by the characteristics of the two countries' companies and the resources at hand."

"We also have a view that the one who gets the talent gets the model, and the one who gets the model gets the world. The models here include pendant models trained for specific scenarios, both ToC and ToB models trained and deployed for specialized scenarios. We are more optimistic about this field than big models, for the simple reason that 98% of the computing time spent on training a model is spent on pre-training, but the next three steps of supervised finetuning, reward modelling and reinforcement learning only use a small percentage of computing time. " Xu Siqing said.

"One does not have ample funds, two does not have the cutting-edge technology talent pool of small and medium-sized enterprises, early to explore opportunities in the application layer, to create some first-mover advantage, is a good choice." A Silicon Valley investor said to Titanium Media Venture Capitalist.

"I think China's big model spell arithmetic is not suitable for general entrepreneurs, understand artificial intelligence entrepreneurs have a great opportunity, because do not have to start from the most basic model brutal arithmetic, than the accumulation of knowledge, team engineering capabilities and iteration speed, on the shallow surface layer of product innovation, such entrepreneurs is a descending blow. Of course, technology is important, industry experience and the choice of application scenarios are equally important. Our seed round investment in HiDream.ai (wisdom of the future) project founder, the former vice president of Jingdong, the Canadian Academy of Engineering foreign academician Mei Tao is based on this consideration." Xu Siqing said.

In addition, Xu Siqing also pointed out that the breakthrough of AIGC technology brings three levels of opportunities to Chinese entrepreneurs. First, it is the opportunity of LLM, big factory or national team; second, it is the opportunity to do superficial application or product-driven application with the help of big model, no matter To B or TO C, there are very many opportunities, characterized by the difficulty of judging which ones are easy to win, as difficult as foreseeing which one must be successful at the beginning of APP emergence; third, it is to establish the pendant model to solve the application level problem, which is in ToB and ToC There is a very wide range of prospects, because the threshold is high and the certainty is relatively stronger. They are characterized by strong modeling ability, and can work on the model to solve the problems that cannot be solved by product innovation.

Finally, he emphasized that there will also be many opportunities for infrastructure-based services that accompany various language models and applications, such as private deployment, cross-cloud, and cross-model service tools.

Foreign big models, AI applications, as much as the carnival

In Silicon Valley, Open AI wants to stand alone, Google, Meta, Amazon can not promise.

Recently, Amazon, which has been invisible, released the Amazon version of ChatGPT, a new model named Amazon Titan. The Titan series of models are divided into two kinds, one is a text model for content generation, and the other is an embedding model that can create vector embedding for creating efficient search functions, etc.

Meta CEO Mark Zuckerberg, not according to the rules, directly hit the "low price", with open source declared war on Open AI "hegemony". On June 16, according to The Information, Zuckerberg is considering commercializing a new version of the open source large language model LLaMA, allowing other companies to use LLaMA for free to develop related AI software and allowing developers to use these products for commercial purposes.

OpenAI has announced plans to create an epic LLM "app store" that will open up all ChatGPT applications, allowing the platform to be bi-directional, allowing developers to shelve their products built on OpenAI technology, and companies to use commercially available LLMs on demand, completely capturing the AI model The platform can realize two-way connection, developers can put up their products based on OpenAI technology, and enterprises can use commercially available LLM on demand, so as to completely seize the AI models.

"The big model competition in the United States must be blossoming, and the degree of openness will become higher and higher, the history of IT development constantly proves this. From Unix to Linux, from iPhone to Android are examples. The leading open power is not enough, and the vendors who are slightly in a passive position will use openness to construct their own ecology, which is a very positive phenomenon for the industry." Xu Siqing told Titanium Media Venture Capitalist.

Looking at the application layer again, there are not a few cases that have gotten commercial validation. Stable Diffusion, a free open source startup, got $100 million in financing in just one month, and its valuation soared to $1 billion; Jasper.ai, a SaaS company that builds applications based on GPT-3 API, was founded 18 months ago and received $125 million in Series A financing, valued at $1.5 billion; Copy.ai, which helps companies write promotional copy. ARR of $10 million in only 2 years.

"There are differences in overseas projects compared to domestic projects. First, the ability of overseas underlying large models is more mature (OpenAI API); second, the commercialization environment is better, users are willing to pay for high-quality single-point tools, which is also the result of experiencing the development of SaaS industry education; third, the ecology of overseas projects is better, there is no squeeze similar to the domestic Internet majors, and there is exploration of new products and new business models in both C- and B-sides. " Shen Huaxin, head of AI research at Yun Cui Capital, said to Titanium Media.

According to PitchBook statistics, the investment circle invested a total of $1.37 billion (equivalent to about RMB 9.369 billion) into generative AI companies in 2022, almost as much as the sum of the past five years.