Do large AI models have first-mover or late-mover advantages? This is a question that partners at an early-stage investment institution have been thinking about.

“Everyone is catching up with advanced large-scale foreign models. For the six AI tigers, it mainly depends on whose hematopoietic ability is strong. You can wait until the technology is equal before copying, but if this is the case, their first-mover advantage will also be lost. No more,” the partner said.

If large AI models cannot continue to keep up with the progress of the latest models, or lag behind open source models, then how can these large AI models survive, transform, or be acquired? Is it a start-up company with a latecomer advantage? Are you more able to cross the river by feeling the stones?

Google researchers once proposed in a blog that when free, unrestricted open source models are of equal quality to closed source models, people will not pay for restricted models, while open source models and closed source models The gap between models is closing rapidly.

The perception of the flywheel effect that existed in the field of generative AI has begun to waver. Previously, David Cahn of Sequoia America pointed out in his article that large model companies that hope to obtain more data through more users have found that this has not proven effective.

For founders of AI large model unicorns, how to survive will be a topic that will continue from 2024 to 2025. This is a problem that founders of large model unicorns always need to face. . Entering 2025, the elimination round of large model unicorns begins.

On January 6, a rumor began to appear on social media that "01 Wanwu was disbanded and Ka and the pre-training team were sold to Alibaba." Immediately, Li Kaifu, founder and CEO of Lingyiwu, refuted the rumors:

According to "Smart Emergence": Lingyiwu's financing with some local governments and state-owned assets is not progressing smoothly, " The money in the account will not last more than a year."

Before Sequoia American partner David Cahn once calculated in the article "AI's $600 Billion Problem" that AI companies still have a $600 billion gap between their investment in training large models and generating revenue. When AI can convert cost items into asset items is a question that investors have always wondered about.

If you want to catch up with the next generation GPT model and the pre-training Scaling Law continues to be effective, some analysts have previously predicted that a single 100,000-card cluster will require a capital expenditure of US$4 billion. This does not include other investments such as power supply. Consider it.

Baidu’s net profit in 2023 will be 20.315 billion yuan. It is difficult to participate in the competition of the next model, not to mention other large AI model startups that do not have core cash cow business. It is difficult to invest in this resource-intensive competition.

Not long ago, there were rumors in the market that the AI ​​Six Little Dragons have been launched today.In the first half of the year, I stopped doing pre-training and then switched to post-training. However, according to the author's various verifications, the AI ​​Six Little Dragons are actually still doing pre-training, but the focus may be different.

According to an early investor privately stated that Kimi actually focuses more on Post-training‌ (‌Post-training‌ refers to the process of optimizing model performance through further training based on the pre-trained model). Because the latter is more effective, as long as the engineered product is easy to use. Wang Xiaochuan has previously privately expressed his approval of the post-training approach.

Overseas media information has published an article saying that GPT improvement is slowing down, and the AI ​​industry is shifting its focus to improving the model after initial training.

The emergence of GPT 01 opens up another possibility of Scaling Law for everyone. By performing RL in the post-training process, the model's capabilities in reasoning and mathematics can be improved. GPT 01 may only require 1%-10% of pre-training computing power in terms of post training computing power, and its inference computing power is ten times that of GPT 40.

For AI large model startups, the focus on post-training is also due to computing power limitations and cost considerations.

According to sources close to SMIC, domestic demand for chips will actually decline in 2024. "Only a small number of large manufacturers like Tencent and Byte are making normal purchases in 2024. In the past, many medium-sized customers did not make purchases this year. In fact, the six AI tigers do not purchase much and rent more."

Morgan Stanley Data shows: Starting in July 2023, the rental cost of H100 on the AWS cloud has dropped from US$8.5 at the beginning of the year to less than US$5. According to data on GPU computing marketplace GPUlist.ai, the median hourly rental price for an H100 is just $2.27. The domestic GPU chip leasing situation is similar.

Compared with high capital expenditure and tight cash flow, leasing is cheap and fast for AI startups, allowing them to invest more money in research and development.

At the same time, due to competition from domestic Alibaba Cloud, Volcano Engine, Google Cloud, Amazon Cloud and other parties, this wave of entrepreneurs have also enjoyed a wave of good discounts.

In fact, the Six AI Tigers spent nearly half of the financing on investment and recruitment.

A headhunter in the AI ​​field said that the founders of the six AI tigers in 2024 have very different plans on how to spend their money. Minimax is more inclined to go overseas and has joined Google and Facebook. Dark Side of the Moon has focused on the Kimi scene from overseas business and has recruited a lot of product managers.

Since Kimi was launched on various social platforms such as Bilibili and Xiaohongshu in March 2024, it has brought a new wave of growth traffic and directly raised the launch price from 3 yuan.At nearly 30 yuan, the number of visits in March was 1.261 million, a month-on-month increase of nearly 3 times. MAU once reached 5.897 million, a month-on-month increase of nearly 100,000. After that, the cost of launching again has also increased, and the number of new user registrations and retention rates have also increased. Not as good as expected. “There’s almost nothing we can do, but the traffic effect is not that good.”

Wu Wei, the founder of Feifan Industrial Research, analyzed that the ROI of many AI products cannot be corrected at present, “You can’t spend money to buy volume. Making a product with a large DAU to sell ads is advertising logic, while AI products are more about payment logic.” In his view, Kimi’s investment actions in the first half of 2024 are more for the purpose. Tell a story, such as tagging the chatbot with the largest traffic. When there is no way to achieve profitability, at least the product data needs to allow investors to see the company's future potential. This is the logic behind the investment of some AI companies.

However, investors soon woke up and looked not only at MAU, but also at 30-day retention and 90-day retention. According to the August data of the 2024 AI product list, the average daily average of domestic AI APPs in the past three months None of the top ten users have been used for more than 10 days.

For example, Minimax’s virtual chat product Talkie. According to overseas media reports, Talkie’s ROI has been corrected. According to data from the AI ​​Product List, Talkie’s MAUs are 2.51 million, and it has the highest retention among domestic AI overseas apps. The average daily usage time in the past three months is 73 minutes. Its main revenue models are advertising, subscriptions and in-app purchases.

According to previous overseas media reports, Minimax’s current net sales in 2024 are expected to reach approximately US$70 million, most of which will come from Talkie. Minimax’s Conch AI has also recently launched a paid version.

When going overseas to do TO C, running through the payment model has become some routine operations for large AI models.

From the perspective of recruitment needs, Zhipu, which is firmly committed to TOB business, is mainly expanding its business resources. It is reported that Zhipu currently has thousands of employees.

However, rapid expansion and high-profile attacks have also brought about personnel instability. According to a former business practitioner of Zhipu, the standard set by Zhipu for business personnel is sales of 1 million yuan in six months. This is also the reason why many business personnel of Zhipu have high mobility in the later period.

In order to survive, the Six AI Tigers, who were still struggling with technology last year, are quickly looking for landing scenarios and products.

Previously, the person in charge of Microsoft Xiaoice said in a public forum that Xiaoice has invested a lot of technology in AI field, but in the final consumption scenario, the profit margin is very low. The huge investment in technology is difficult to realize when it is actually implemented. This is a problem faced by the first generation of AI companies.

The four AI visual tigers who have experienced the last wave of market testing will be more restrained and more focused on technology investment.Focus on the implementation scenario, but financial pressure forced them to change quickly. "In 2023, large AI model companies will use applications to train models, but in 2024 it will completely change, and they will use models to train applications." said the above-mentioned early investor.

Of course it is not easy to find scenarios and products that can be monetized.

According to Feifan Industrial Research Data, the current OpenAI ARR is US$4.552 billion, with revenue mainly coming from membership subscriptions and APIs.

However, membership subscription and API are currently not feasible for large domestic models. The former is due to domestic payment habits, while the latter is mainly only useful to the world's leading technology leaders.

In the view of Wuyuan Capital partner Meng Xing, selling APIs is not a long-term business model. Unless it is the first in the industry and maintains an absolute lead, there will be no chance after third place. .

In fact, there has been a wave of profiteering parties in startup companies. In order to attract companies to call APIs, some large model companies will give away an additional portion of Tokens. In the past, when Tokens were relatively expensive, some entrepreneurs would even register with different companies to earn money.

Informed people believe that many startups are now more willing to go to cloud manufacturers. "Some cloud vendors may give you cloud-related vouchers of different amounts, such as 50,000 US dollars or 100,000 US dollars. You can consume both cloud resources and token resources to meet the more diverse needs of startups."

In addition, unlike SaaS services, for customers with large models, there is almost no cost for model switching, which further intensifies the price war.

Then the only way left is to develop overseas TOC, adopt a subscription model or help industry customers develop TOB solutions.

At present, among the six AI dragons, only Zhipu is firmly following the TOB route. Starting from the beginning of 2024, Zhipu will begin to recruit GR and BD. "Zhipu is very good at serving TOB customers." A person familiar with the matter said that even if the money is not much, if you have any needs for large models, they will directly send very expensive engineers to help customers implement them.

At present, the TOB customers mainly served by large models are concentrated in finance, government affairs, operators, universities and other industries. Clients of Zhipu’s external publicity cooperation include Huatai Securities, Mengniu, SAIC, etc.

From extraordinary data, as of September 30, 2024, Zhipu has won 11 public biddings, ranking fourth after iFlytek, China Telecom, and China Mobile. In 2020, there were 5 winning bids for Zhipu. Public data since 2024 show that the winning amount of Zhipu in the bidding was 24.98 million yuan. Among the top ten, only Zhipu is a startup company, while the others are large companies or listed companies that have been established for many years.

“The marketing of large models in the TOB field is a price war, with emphasis on delivery.” A colleague close to Zhipu Business said.

There is no such thing as landingIt is so easy. For all companies doing TOB business, they need to go through a lot of troubles in the not-so-sexy business process before they can cross the river called "delivery". Customers who have cooperated with Zhipu before have said that there is still a gap between its promises and the final results.

In the TO C field, the Six AI Tigers are also constantly trying various new products. Meng Xing believes that this is somewhat similar to the previous development stage of self-driving startups. In the early stage, the founders continued to receive financing because they were technology experts. But when it comes to the realization stage, what is more needed is products that can be monetized.

Is AI a productivity revolution or an interaction revolution? This is also a question Wu Wei has been thinking about. "If it is the former, it means that there will be no new platform and no new traffic dividends. In fact, it is more of an opportunity for existing Internet companies and mobile phone hardware manufacturers."

Foreign data shows that currently, real users There are still a small number of AI products, and most of them are concentrated in existing scenarios.

If it is an interactive revolution, then more opportunities for AI large-model startups lie in AI native.

Investors who pay attention to the field of AI believe that although models are critical, interaction and user emotions are also important. "What many professors are doing now is just a large model with a shell. It is not actually a product. The product should be combined with the demand scenario and can be used in a fool-like manner."

The opportunities for AI native that we have seen so far have been eliminated. What is verified is chatbot, programming, and digital human companionship. The domestic ecology is different from that of foreign countries. Domestic major manufacturers prefer to do All in One. From underlying models to front-end applications, platform businesses or applications with hundreds of millions of DAU are all within the scope of major manufacturers.

The most obvious big opportunities often face fiercer competition. For example, in the field of Chatbot, Qianwanji’s Doubao currently has 5.99 million MAUs. Zhipu’s Zhipu Qingying is not its focus. The strategic level of Conch AI’s chat tool has also declined internally. Currently, only Kimi is still there. Stay on that course.

For six For Xiaolong, more opportunities may be hidden in subdivisions and ecology.

For the AI ​​Six Little Dragons, they are looking for more vertical and subdivided directions. For example, Baichuan has focused more on the medical vertical model, and on the other hand, they have all adopted incubation + investment. +Self-developed model to build a broad ecosystem.

According to an early investor, Minimax is acquiring some small companies, mainly acquiring products, because they are worried about Talkie’s ability to continue to produce hit products in the future. On the other hand, they will also incubate some startup companies, which focus more on Token usage. It is reported that Minimax's investment in AIPPT is also a sign of the huge usage behind it.

Zhipu also raised two funds in 2024. Among them, Zhipu AI, Xinglian Zhaoji, and Jingcheng Yanbei jointly invested 260 million yuan to establish a new investment fund, Xinglian Dingsen. At the same time, Xinlian Capital, which has a close relationship with it, also established Hainan Sanya Lianxing Shangzhi No. 1 Equity Investment Fund. In addition, Zhipu Z plans to join forces with ecological partners to launch a large model entrepreneurial fund with a total amount of 1 billion yuan. The main investment directions include large model algorithms, underlying operators, chip optimization, industry large models and super applications, etc. So far, Zhipu has invested in 13 companies, and it invested in 8 companies last year.

In addition, it is reported that Baichuan also has a fund of its own. At the same time, Innovation Works, which is inextricably linked to Zero-One Everything, also recently announced its cooperation with Geek Park for incubation.

In the first half of 2024, Google acquired C.AI for US$2.5 billion. Later, the company’s founder and 30 core employees joined Google. This acquisition method mainly focused on this team. Similar ones are Microsoft acquired Inflection and Amazon acquired Adept.

Some legal experts analyzed that this kind of acquisition can not only obtain an excellent team, but also does not bear the company's past liabilities and some legal risks.

In the TOB and TOC modes, each of the six AI dragons has its own focus, except for zero, one, and all things. Zero One's journey in terms of TOC was not smooth. It is reported that the ARR of POP.AI is expected to be between 10 and 20 million yuan, and the product manager of POP.AI has resigned not long ago. On the ToB side, Zero One is facing direct competition from Wen Xinyiyan and Zhipu, and its position is relatively awkward.

In fact, the external performance of competition among large AI models is ecological competition, and its core is talent competition.

At present, domestic large AI models are still at least half a year behind foreign countries, at the level of GPT 3.5. "Because of foreign talent density and continuous innovation, they can always be half a year ahead of you. This is the real barrier." Wu Wei said.

A common understanding in the field of large AI models is that talent density is greater than data quality than model architecture.

Scarce talent is the most important. Kunlun Wanwei founder Zhou Yahui said on social media that Byte’s 2024 AI strategy scores full marks and the organization has strong ability to quickly iterate. A video generation industry practitioner was surprised that Byte was able to launch a video generation model with good results in less than a year.

Byte’s talent strategy is inseparable from the great efforts that can produce miracles.

A Byte insider revealed that really talented people either start their own businesses or are poached by Byte.

According to late reports, Li Xiangang, the co-founder of Zero One Thing, was also revealed to have resigned and returned to Shell. Ming Chaoping, the former product leader of Noisee, founded the AI ​​Coding company. Hong Tao, the co-founder of Baichuan Intelligence, has also resigned. Minimax Lianchuang Song Yachen came out to found the company.AI+3D company Vast.

But more people are moving to Byte. Byte has frequently contacted people at Alibaba’s P9 and P10 levels this year. What is currently known is that Zhou Chang, the person in charge of Alibaba Tongyi Qianwen’s large model (P9), has joined, and has previously launched a series of multi-modal models. Earlier, Huang Wenhao, the person in charge of model pre-training, and Qin Yujia, a core member of Wall-Facing Intelligence, joined the Byte Flow team.

In addition to not having to consider survival issues, Byte also has obvious advantages in traffic and data. According to industry insiders, the Douyin platform mainly prioritizes its own products for streaming, while other products are mainly non-mainstream products. In terms of data, Byte has multiple social platforms with MAUs exceeding one billion. In terms of funds, Douyin Advertising revenue last year reached 400 billion yuan. In terms of talent, Byte is also sparing no effort to recruit people. What Byte lacks is only time.

For the six AI dragons, what is lacking is not only time, but also funds, computing power, and traffic.

However, as the above-mentioned early investment institutional partners said, we can wait half a year to see who among these six companies will have obvious problems and who will be more aggressive.

In 2025, let the bullets fly for a while.