Moment of Singularity, the Internet's last hurrah
August 9th, 2023

In 2023, the news of the failure of Tiger Fund, a well-known investment organization, to raise capital, quietly spread throughout the Internet.

In the past 10 years used to the wind mouth entrepreneurship, "investor winter" seems to be the first time. This and the new consumer, live with goods, meta-universe several wind mouth quietly fall, mergers and acquisitions and Chinese shares and other exit channels of the door half covered, all unfavorable factors are closely related to the venture capital market seems to be really cold down.

Startups are having a hard time financing, and the second venture of the big guys is not looking for a good direction. Wang Huiwen, who retired from Meituan, has been studying Web3 and meta-universe for a while. Wang Xiaochuan, who left after being acquired by Tencent to buy the company, tested the waters of AI medical. But everything shifted at the end of 2022, ChatGPT 3.5 was released, quickly allowing the market to form a consensus that the AGI (General Artificial Intelligence) era is here, and the whole industry began to run into the big model.

It is understood that Wang Xiaochuan, who was in a low-key venture at the time, has set up a company to do intelligent hardware. Intended to help hundreds of millions of people with sleep disorders, to create a smart pillow for snoring. When the big model boom rose in March, Wang Xiaochuan used 2 weeks to make a decision to put down this entrepreneurial project and lay out the big model.

Wang Xiaochuan found the former Sogou CTO Yang Hongtao to help take over the medical project, and the former Sogou COO Ru Liyun's shares in this company were cashed out to Yang Hongtao to follow Wang Xiaochuan's big model venture. Wang Xiaochuan took out a total of 50 million U.S. dollars to set up "Baichuan Intelligence", and invited Soul's technical talent to do the algorithm in charge, to accelerate to do the big model. Wang Huiwen's story, we are very familiar with, released a hero recruitment post on the table, set up a company outside the light years to do big models.

In the Internet big factory, the big model has also brought a turning point. There are big models of the project leader, a year ago because of the problem of promotion can not be, and thus proposed to leave. After 3 months time after the year, the CEO of the group became the general director of the big model, lifting the company's strength All in big model.

No one wants to miss this wave of the AGI era, and everyone believes that the singularity of generalized artificial intelligence is approaching after three ups and three downs in the development of AI. After all, in the AI boom, similar to ChatGPT and Midjourney and other dozens of people scale company, to create a valuation of about 4 billion U.S. dollars, the U.S. stock market value of the "seven giants" soared to 11 trillion U.S. dollars in a year, a surge of 60%. These exciting stories of explosive growth have once again stirred the domestic technology business market.

Domestic Internet makers, Robin Li, Zhang Yong, Zhang Yiming, Wang Xing and other bigwigs have been personally commanded, it can be said that in addition to Pinduoduo, has all entered the big model. As on July 19, the market value of Microsoft and NVIDIA increased by $175 billion, and Musk marveled when evaluating the related tweets, "Crazy times."

The soon-to-be-silent tech business market suddenly welcomed the stimulant of AI and sent the Internet into one last frenzy.

  1. A new dawn lit up in the trough

Li Ming is the CEO of a startup company with a team of more than 100 people. 2023 was the year he was most worried about financing.

At the beginning of the entrepreneurial process is very smooth, early to get well-known angel investment institutions of angel and A round of financing. "At that time, the industrial Internet was still a popular track, and it was not as flashy as many AI projects." Li Ming told AI Whale Selection, but in the middle of 2023, he slowly realized that the market was not right in the new round of financing launched.

Investment institutions are not only looking at data and stories, but also at revenue now. Li Ming, who was previously obsessed with productization, simply hasn't realized that the investment wind direction has changed. Plum Venture Partners founding partner Wu Shichun's speech, is now investing in projects "both (technology), and (data), but also (revenue)". No way, he began to look for FA institutions to help finance, and the financing rounds have also stepped back, seeking an A++.

"FA help to find more than 30 investment institutions, all failed." The lack of success in financing made Li Ming a little discouraged. But in June, he felt the power of the big model, so he internally mounted the industrialization business based on ChatGPT. "There is also no financing yet, but investors will take the initiative to find and communicate with each other, and the other party is obviously interested."

For Yuan Jinhui's first-class technology, the big model is also a lifesaver.2022, the company that makes AI deep learning frameworks has reached the point where financing is not going well and has to lay off employees to survive. Previously, the company was on the verge of a broken capital chain 3 times, all looking for angel investors, also at that time the CEO of the fast hand Cebu Hua borrow money.

"Doing things and Baidu's flying paddle, Huawei Sheng Si is almost the same, the most important thing is that at that time the business of the market large model training has not yet risen." First-class technology employees told AI Whale Selection, the company belongs to the time when there is money (2021) there is no business, and when there is business (2023) there is no money.

Just when Yuan Jinhui felt that the future was hopeless, the company also ushered in an acquisition opportunity in 2023.In April 2023, a VIP was welcomed in the first-class technology company in the Tsinghua Science and Technology Park, who was Wang Huiwen, the co-founder of Meituan, who had just announced that he had entered the big model.

The final acquisition price is okay, a laid-off employee of First Class Technology told AI Whale Selection, "It can be comparable to the valuation of the last round of investment by Gao Tiles Capital, and his own options are looking for a fall."

And Yuan Jinhui, who became the co-founder of Light Years Beyond, finally no longer has to worry about financing. Wang Huiwen's ability to raise funds is unrivaled in the current venture capital circle. According to the later acquisition agreement of Meituan, Light Years Beyond raised 2 billion yuan without a large model product.

Of course, investors who laid out earlier in this wave of action have successfully hunted unicorns.

Minimax was founded in November 2021, received an angel round of investment in January 2022, and the company's valuation reached the unicorn level in early 2023. Among the earliest four investment organizations, there was also Shanghai-based gaming company Miha Tour, reportedly because of family ties among the two founding executives. And according to Whale Selection, Wisdom Spectrum has also recently been raising capital at a valuation of RMB 10 billion.

Both companies were founded less than 2 years ago, but both have become unicorns, and the big modeling track is developing at an amazing speed.

The AGI boom is also a redemption for those old AI companies. Previously, the IoT listing story of Going Out and Asking has gone through several unsuccessful attempts. With the launch of the big model "sequence monkey" and the story of four AIGC products, although the big model is still careful not to public evaluation, but also let out the door to ask finally have a new story to tell, has been submitted to the Hong Kong stock listing application.

More big models and AIGC entrepreneurs are on the way, even on a startup camp, 60% of the projects are related to AI, with the advantages of light assets, high barriers and high ceilings, AGI has completely become the hottest track of the moment.

  1. Take the dream of AGI to its peak

If it is said that 2023 is the "first year" of big model entrepreneurship. Then the "source year" of the earliest entry of the Internet big factory into the big model can be traced back to 2019.

Ali started the layout of the big model in September 2019, and in April 2021, the PLUG big model was released. As early as before ChaTGPT 3.0 came out, there have been a number of large models with trillions of parameters in China, which are the M6 of Dharma Institute and Huawei Cloud's Pangu large model and Wisdom Source's Wudao 2.0.Compared with ChaTGPT, although the model parameters have surpassed the model parameters, but the abundance of the data is not the same, and the effect is not comparable, and in the opinion of Zhang Cong of Dharma Institute, the large models in China have started to catch up with the late set, and the most important thing is that they have not done two things. The most important thing is not to do two things.

The first thing is not to do alignment. At that time, Ali had a lot of large and small models, mainly did not do the training results alignment. "You see now ChatGPT can do poetry will chat, very much like human intelligence, in fact, is aligned with human values." Zhang Cong said, these need to reasoning results of human adjustment, rather than using the logic of the machine to do.

The second did not go for high-quality datasets, ChatGPT early use of university professors in the Philippines for data labeling, the country is to use secondary school students to do the labeling, the problem of corpus is also very much affect the results. In Zhang Cong's opinion, the fine-tuned Chat model of Llama 2, announced on July 19, was trained on 1 million human labeled data, and the total number of training tokens increased by 40%, which is an all-around improvement compared to Llama. "So big models are not inventions that vigorously produce miracles, but carefully engineered creations."

And in contrast, the domestic AI industry also faces many other factors that interfere. At that time, there were two main teams in the Dharma Institute to do the big model, one is the machine intelligence team led by Jin Rong, Si Luo is responsible for AliciMind; one is the natural language laboratory led by Zhou Jingren, in which Yang Hongxia is responsible for the big model M6.

In the evaluation at the end of 2022, the results of the M6 big model were slightly superior, and the two were eventually integrated into the current Tongyi big model. "In fact, the Dharma Institute big model team is only twenty or thirty people, mainly his pre-training, are placed in the Ali cloud." Zhang Cong told AI Whale Selection, but now Tongyi is an important project of the group, involving more than 600 people, now a lot of resources are tilted to the big model, the group CEO asked about the progress of technology once every 2 weeks.

For Baidu, this wave of AGI boom, but their own AI era that was predicted to come from 2016, naturally will not miss.

February 7 this year in the internal formal project, March 16 officially released. During this period rose directly to the Baidu Group's highest priority project, Robin Li personally supervise the war, CTO Dr. Wang Haifeng directly marshal, then Baidu Yangquan supercomputing center is dedicated to large model training.

Baidu algorithm engineer Zhao Hui told AI Whale Selection, Baidu Natural Language Processing Department has always been in the research of NLP and other technologies, the chief scientist Wu Hua has also been the leader, the department has hundreds of people. Baidu's ERNIE2.0 after the switch to Wenxin big model, "before it was doing Baidu brain, now it is said to be a big model Wenxin it."

Doing things are similar, of course, there are also differences. Zhao Hui mentioned that in the past, Baidu would do a lot of vertical search Rank, is to reorder the search results based on human clicks. After the emergence of the big model, these capabilities are precipitated in the algorithm of the big model, which is also conducive to giving more accurate answers.

For Baidu, the big model to promote the next generation of search qualitative change has been written into Robin Li's OKR. however, for the ecology, Baidu's Wenxin big model is based on the bert model, "including Zhiyuan's GLM is an independent technology route, and the international GPT is not the same." A Baidu cloud personnel told AI Whale Selection, this point in fact do not have to worry, Wenxin Qianfan what type of model are available, GPT2, 3, 4 is also very different.

And back to Yang Hongxia, who left from Ali, after she went overseas, she was also tapped by ByteDance to be the head of research and development of the North American big model. Zhang Yiming has been studying whether the big model will be open source or closed source, so he didn't ask to focus on catching up. "There will be a real breakthrough by the end of the year." Yang Hongxia told AI Whale Select.

Taken together, ByteDance should be the company that is more compatible with the big model in business after Baidu. A headhunter told AI Whale Select that although the big model is not in a hurry, it is still quite aggressive in the AIGC field. For example, Tiktok is doing advertising creative business AIGC, the director position is given a budget of 100-150W, and the requirement is to lead the team after 88.

So far, the Internet big factories, in addition to Poundland, have all entered the big model. The enthusiasm of the big factories to enter the game, even more than the year's O2O and live broadcast.

  1. Watershed suddenly appeared that night

In June, at the Sohu Building in Beijing, Lightyear Beyond, the most financed large model company, was feverishly starting up.

The original first-class technology Oneflow deep learning framework is still looking to continue to do, but by the big model business has drawn a lot of people. But on June 23, suddenly someone revealed on social media that Wang Huiwen was sick, when the company also went to seek confirmation, and got the news that there was no such thing. But on the night of the 25th, the United States suddenly announced that co-founder Wang Huiwen was hospitalized for depression, resigned from the company director of the matter, its business venture light years away from the company facing out of the news.

For a time, light years beyond can not do, Wang Huiwen early run away news, become some people's speculation. Whale Choice got the news from the investor circle that Wang Huiwen's condition was indeed very serious. In the end, Wang Huiwen's brother who slept on the upper bunk, Wang Xing, founder of Meituan, helped take over Light Years Beyond.

Big model really not work? Everyone sprouted this question. During that period, it happened that well-known investor Zhu Xiaohu and Cheetah founder Fu Sheng also argued in the circle of friends, whether the large model industry has a bubble. Zhu Xiaohu extremely bearish market swarming, do the status quo of the general large model, that the vast majority will die at the end of the year.

Light years away from the active changes, also whether the corroboration of Zhu Xiaohu's remarks?

From the information obtained by AI Whale Selection, the acquisition of Lightyear Beyond's Meituan, currently does not stop the footsteps of the big model. Not only exclusively invested hundreds of millions of dollars in Wisdom Spectrum AI, at the moment is still recruiting big model project director, giving an annual salary of up to 3 million dollars, and even set up a technical research institute in the United States. Earn hard money of the United States group, and do not want to lag behind in this wave of technology, especially in the hungry one clear to access the big model of Tongyi, there is business competition Ctrip, has also been launched after the big model.

But for the domestic market, the generalized big model has really too much. Incomplete statistics, in just less than 8 months, there have been more than 85 big models released, many of which have become the concept of cash for listed companies.

Wind data show that in 2023, 24 "AIGC concept shares" has been a total of 67 holdings, the wave of divorce of major shareholders is also amazing. 2023 year-to-date, nearly ten AI plate company's major shareholders family was exposed to divorce. Much attention is paid to the A-share AI company Kunlun Wanwei, recently occurred 11% of the shares of Ms. Li Qiong (founder Zhou Yahui's ex-wife), plans to reduce 3% of the shares (roughly 1.3 billion yuan), and then interest-bearing loan to the company. According to people who know the inside story to the whale selection society, feel the AGI dividends of Kunlun Wanwei, not only do the big model, recently also intensively set up a team, go all out to do the benchmark Microsoft Copilot.

Listed companies utilize AGI to speculate on concepts and leave the market with cash. Large model startups are involved in the process until they die.

Zhang Yang, an investor who recently set up the AIGC fund, told AI Whale Selection that with the arrival of the open source, free and powerful Llama 2, many large model companies are bound to face financing difficulties in the second half of the year.

Now everything has a clue, in July 11, Baichuan Intelligence launched ten billion level parameters of the big model Baichuan-13B, not only announced open source, also free commercial. Although the parameter scale of Baichuan-13B is not large, but based on accurate Chinese corpus training, Baichuan is often ranked as the head of the large model in the ten billion scale parameters.

Baichuan-13B's free strategy has greatly impacted the domestic big model paid market. At present, Zhiyuan AI just announced on the 14th, the enterprise registration was authorized to allow free commercial use of ChatGLM-6B and ChatGLM2-6B.

After more and more big models are open source and free, the death elimination game of big models is officially opened. A CTO of a large model-based startup company told AI Whale Selection that Zhiyuan's large model from the beginning of the private domain deployment to 20 million yuan, to the beginning of the year to call the price of 1.8 million to 300,000 optional packages, and then now free, the industry is changing very fast. Fu Sheng believes that this is the market from the big model parameters of the fight, into the ecological scale of the fight.

Internet big factory is not worried about ecological construction, due to the internal model is very much, there are also free and paid points, the most important big model is still closed source and paid form. Startups to build the ecology is more difficult, many startups do big models have spent all the effort, do ecology is inevitably not enough power. It is understood that MiniMax is currently a startup company, one of the few companies to adhere to the public cloud, do MaaS model of the large model.

Clove CTO Fan Kai described this wave of open source free tide, like the water plant (big model) free to the user's home, so that each family manually a faucet, those closed-source tap, it is best that your water invincible good taste, we are willing to pay to go to you that.

4.AGI development into a fork in the road

After the watershed appeared, the former chairman of the technical committee of the Jingdong Group, now the founder of the title of the far technology told AI Whale Selection, when the startup competition has been and develops 3 factions.

One faction is to adhere to the full self-research big model, this faction are strength players. This faction is mainly Baidu, Ali, byte and other Internet big manufacturers, as well as wisdom spectrum, MiniMAX, articulated far and other startups. But these power players are also divided into two types of enterprises.

The first category is to insist on doing self-developed generalized big models, benchmarking ChatGPT and constantly catching up with ChatGPT's iteration speed.

In the opinion of Chen Yu, Managing Partner of Yunqi Capital, the general large model is the way to go, and the development of vertical large models is limited. "Because for the general large model, the vertical field does not need to be retrained, the general large model can do industry deepening through the vector database, but the vertical large model is difficult to intelligently emerge."

From the current point of view, there is a dream of surely still have to do the general big model, after all, made it can become the next Internet big factory.ChatGPT in the field of collaborative office, e-commerce, code generation, auxiliary design and other areas have shown this disruptive potential.

The second category is to recognize the reality of focusing on landing, adhere to the vertical model, this faction, including the last to reach a unified point of view of Zhu Xiaohu and Fu Sheng, both believe that the vertical model will be more industrial applicability.

Generic large model is generally in the hundreds of billions of parameters above, while the vertical large model is in the tens of billions or 7 billion scale or so. Similar to the big model products ProductGPT and Cao Zhi of the big model products of Daguan Data, the parameters are in the tens of billions of scale.

Articulate far technology is not a big model of the parameter school, "we have a general big model of the base。

Subscribe to Lenson
Receive the latest updates directly to your inbox.
Nft graphic
Mint this entry as an NFT to add it to your collection.
Verification
This entry has been permanently stored onchain and signed by its creator.
More from Lenson

Skeleton

Skeleton

Skeleton