On February 22, Baidu released its 2022 financial report. It was supposed to be a stage to showcase the achievements of the past year, but “Wen Xin Yi Yan” took the lead and stood at the C position.
Including Li Yanhong, no one seems to care about Baidu’s performance in 2022, but all eyes are on the same line with Wenxin, and they all predict that it will subvert all walks of life after its release. Robin Li even boosted morale in an internal letter, saying that Baidu will integrate its mainstream business with Wenxin, and that Baidu is “standing on top of the wave.”
Unexpected, reasonable.
During the months when ChatGPT was booming, 80% of the popularity was contributed by domestic followers. As a representative of followers, Wen Xinyiyan, although the product is still in a hurry to develop, has already been pushed to the forefront.
In addition to Baidu, major manufacturers are also scrambling to express their determination. They not only regard ChatGPT as the first outlet in 2023, but also a career worthy of All in for the rest of their lives.
A promise is a promise, and whether it can be fulfilled is another matter. Everyone regards ChatGPT as “the hope of the whole village”, but hope does not belong to everyone.
This had to turn to the other side of Cyclonus.
The reason why ChatGPT is so popular is very simple: Compared with the so-called “next generation Internet” such as Metaverse, WEB 3.0, VR, etc., ChatGPT not only shows grounded application, but also opens up infinite imagination space , and generally only the latter can be seen in the past.
But at the same time, the ChatGPT tuyere also exhibits other atypical characteristics. For example, the extremely high capital and resource threshold corresponding to the large language model technology, as well as the continuous investment in the future, are destined to be the exclusive game of giants from the root.
The clear threshold determines that this is not a poker table that can be played casually. While speculators have nowhere to hide, those who enter the game must also master the correct posture of chasing the wind.
All the responses, the application scenario is activated
Overnight, almost all major Internet and technology companies gathered at the venue.
Baidu, Alibaba, Tencent, JD.com, Bytedance, Kuaishou, 360, Netease, Xunfei, Huawei. ． ． ． ． ． Although they each focus on different businesses, the major manufacturers have reached a consensus on the technical capabilities of AIGC represented by ChatGPT.
Objectively speaking, some of the “reprints” of major domestic manufacturers tend to be inspired by ChatGPT, and the application scenarios are activated. Because some companies have already made the necessary technical reserves in the past, the emergence of ChatGPT points out the direction for the implementation of technology.
As a practitioner, Li Di, CEO of Microsoft Xiaoice, holds a similar view, “The most important thing is not ChatGPT, but the rise of a new paradigm of large language model technology represented by ChatGPT.”
The rise is reflected in two aspects:
One is to throw money at real money.
Wang Huiwen, the No. 2 person in Meituan, is the most high-profile person. He invested 50 million U.S. dollars and claimed to build China’s OpenAI. Don’t worry about money.”
Zhou Hongyi also spoke with money. He said bluntly, “If a company can’t take the bus of ChatGPT, it will probably be eliminated.” At the same time, it told investors that 360 said that it has more than 20 billion funds to support 360’s research and investment in ChatGPT-like technology.
In addition, although Baidu, Ali, Tencent, ByteDance, Huawei and other entrants have not disclosed specific funding plans for ChatGPT, they have found some references.
For example, Li Yanhong recently disclosed a set of data. In the past ten years, Baidu has invested more than 100 billion yuan in AI, and the annual R&D accounted for more than 15% of its revenue. This makes Baidu have a full-stack artificial intelligence Layout, including the large language model “Wenxin”.
The second is the expansion of the business level.
The main business of each company is different, and the application scenarios of ChatGPT-like technology are naturally different. For example, Bytedance, combined with news and information services, AIGC technology is mainly used for content creation. NetEase uses related technologies for education, and its main business scenarios are AI translation.
And as major manufacturers advance step by step, ChatGPT appears in more business scenarios.
In the media industry, according to statistics, dozens of media platforms such as Pengpai News, Upstream News, Elephant News, and Daily Economic News have announced access to Baidu’s “Wen Xin Yi Yan”; in the tourism market, Ctrip has introduced corresponding technologies for smart travel Planning, travel content, etc.; car tracks, natural dialogue capabilities can be used to improve car-machine interaction.
Some people in the industry call it “wherever you need to move”. If it is further expanded to the scope of AIGC, text generation, image generation, video generation, audio generation, digital human and other technical subdivisions, the corresponding application scenarios will be further expanded.
Therefore, the explosion of ChatGPT is just the beginning, and more penetration is still on the way.
Exclusive games for giants, startups have no chance
The outlet is there, everyone can try it, whether they can play well is another matter.
It can be seen from the origin of ChatGPT that the big language model is born to “hate the poor and love the rich”.
According to public information, the basic model GPT-3.5 needs to cost 3 million to 4.6 million US dollars for one training, which is obviously not affordable for ordinary enterprises. Moreover, with the improvement of model capabilities and continuous optimization of parameters, the cost of training will continue to increase. Today, ChatGPT costs as much as $10 million to train once.
Even Wang Huiwen, who shouted “don’t worry about money”, the 50 million US dollars he took out was only enough for a few training sessions. Li Di also publicly stated that “If you consider the cost of ChatGPT, you will burn 300 million yuan every day, and you will burn more than 100 billion yuan a year.”
Another person in the industry emphasized, “In addition to being expensive, model training has another feature, that is, there is no return in a short period of time. Even if there is commercial realization, it is necessary to continue training according to actual application scenarios and invest more funds. Startups simply can’t afford it.”
The data shows that before the ChatGPT fire, high computing, labor costs and weak commercialization capabilities formed a huge gap, and the financial situation of its main company, OpenAI, was not good. Taking 2022 as an example, OpenAI will have a total revenue of US$36 million and a loss of more than US$500 million.
Under such circumstances, the most basic guarantee for the successful launch of ChatGPT lies in capital transfusion.
Fortune magazine recently reported that Microsoft invested a total of $10 billion in OpenAI. Prior to this, OpenAI had received a total of $4 billion in investment in more than seven years since its establishment.
The funding threshold blocks the path for startups to reproduce ChatGPT. In addition, the resources that startups can use for model training are also restricted.
Take data as an example. Massive data is an essential resource for the development of artificial intelligence. The quality and quantity of data fed to the model determine the upper limit of its capabilities to a certain extent.
Looking back on the growth history of ChatGPT, in the GPT-2 stage, the training text of the model includes 8 million Reddit forum posts, a total of 40GB of data, and the corresponding parameters are 1.5 billion; to GPT-3, on the basis of the previous, OpenAI scientists put the previous In 12 years, hundreds of billions of words of English data, such as news reports, posts, full-text books, and various web pages collected from 60 million domain names, were input into the model for training, consuming tens of millions of dollars in computing resources.
In the context of the Chinese Internet, data resources are naturally in the hands of Internet giants such as BAT. At the same time, they also have the “money ability” that startup companies cannot match. This means that the re-enactment of ChatGPT is a game exclusive to giants, and the probability of a startup company winning in it is infinitely close to 0.
Atypical outlets, speculators have nowhere to hide
According to past experience, the trend will inevitably attract a group of speculators in order to catch a wave of popularity or make a wave of quick money, and ChatGPT is no exception.
For example, in the recent Metaverse, Facebook’s parent company will be renamed Meta in 2021. Domestic companies are scrambling to enter the Metaverse track. Metaverse apps emerge in endlessly, and the concept of Metaverse is applied to various industries.
The wind has been blowing for more than a year, and Metaverse is still at the conceptual level. The leader Meta has fallen into reality with massive layoffs and plummeting stock prices. Zuckerberg said recently that he is optimistic about the company’s prospects in the next 5 to 10 years, and admitted that he had mistakenly believed in 2021 that the fiery trend of the metaverse would continue.
The big brother who took the lead is still like this, and all kinds of chaos are emerging in an endless stream of followers. There is no need for the Internet big positions who bet on the metaverse to persist. It just so happens that ChatGPT is hot now, and there will be no shortage of money in the future, so it will undoubtedly become an excellent choice for speculators.
As a result, many companies that participated in the Metaverse in the last round began to turn around and tell new stories about ChatGPT.
However, combined with the characteristics of the large language model technology itself, what ChatGPT provoked is an atypical outlet, and speculators may find it difficult to find a place to stay.
After all, most of the speculators get together are low-investment outlets, while ChatGPT is a real “asset-heavy” route.
In the past, many hot spots that flowed from Silicon Valley did not have high requirements for players in terms of funds, and domestic companies did not need to invest too much in copying.
The audio social software Clubhouse in 2021 is a typical example. It also became popular overnight, with an astonishing growth rate of users, and won the platform of big names such as Musk. Players who are chasing the trend in China can catch this express train only by launching the chat room function that focuses on voice social interaction.
The Dialogue Bar app launched by Inke, with the aura of “China’s version of Clubhouse”, was taken off the shelves after a short-lived flash in the pan. Although it failed to bet on the trend, but only at the cost of losing a beta version of the APP, it really gained a sense of presence in the Internet industry and social circles. For Inke, which has been silent for a long time, it is more cost-effective than marketing expenditures.
According to the access thresholds of capital, talents, resources, etc. mentioned above, it is almost impossible to use the routine of running away in the past to copy ChatGPT. Without a large language model, ChatGPT is out of the question.
Therefore, it is only necessary to use a simple indicator of sufficient funds to clearly identify who is following the hype.
For example, a company like Secoo, which has been exposed to bankruptcy and run away, has already been unable to take care of its original financial gap, so how can the funds be used to support model training? Another example is HKUST Xunfei, which said two months ago that it has no plans related to ChatGPT, recently announced that it has Confidence in achieving similar technological progress, and the financial report shows that the company’s annual revenue is about 20 billion yuan, which is far from enough even if it is fully invested in research and development.
In the same way, 360, which claims that 20 billion funds can be used for ChatGPT-like technology research, cannot escape the soul torture of whether the funds are sufficient. Moreover, referring to 360’s performance in live broadcasting, games, smartphones, IoT, new energy vehicles, metaverse and other projects in the past, it is still doubtful whether it will really bet all its wealth on an uncertain result.
Under the atypical outlet like ChatGPT, no matter how determined the company is, it will directly show its original shape if it fails to pass the hard threshold of funds, and speculators who want to catch the wave of enthusiasm and run away have nowhere to hide.
Giants eat meat, can small factories drink soup?
As a follower of the trend, companies express their belief in ChatGPT in different ways. For example, Zhou Hongyi has publicly stated more than once that “if a company cannot catch the ChatGPT bus, it is likely to be eliminated.”
The fact before us is that, except for giants with strong reserves, other companies are struggling to even get started in the field of large language models. As Wu Hequan, an academician of the Chinese Academy of Engineering, said, “There are still a few companies in the world that can support such a large-scale computing power.”
Therefore, the above two views seem to be fundamentally in conflict. But in fact, there is a middle way hidden in the middle,Taking the bus of ChatGPT does not mean that you have to train a large language model yourself.
Combining the past development experience of other industries, companies that have entered the game will increase their research and development of ChatGPT-like technologies. After the products mature, they will open up their capabilities to the outside world. They can use this to realize commercialization and fill the corresponding R&D investment. The companies that cooperate with it can introduce large language model technology into actual business with a lower threshold, and enjoy this wave of technological dividends in another way.
This once again points out that ChatGPT has different characteristics from other outlets, that is, it has more specific and applicable usage scenarios.
Take Baidu as an example. It has been focusing on the AI field for many years, and its related technical reserves are leading in the industry. After ChatGPT became popular, Wenxin Yiyan big language model was pushed to the front. Although it has not been officially launched yet, the capabilities of this model have already won the support of partners in many fields. The industry supports Wenxin Yiyan. Expectations are also placed on Baidu’s new search experience.
On the other hand, some relatively floating concept outlets before, such as Metaverse, Baidu has also tried Metaverse business, such as the launch of Metaverse APP Xiyang, which has been online for more than a year now, and the screen in the app can still be seen through molds, freezes, etc. Low-level technical issues, and because the application experience cannot win user recognition, so far all Metaverse APPs have not been widely popularized.
For speculators who are interested in the popularity of ChatGPT, they actually have the opportunity to make a new choice. Instead of waiting for the bubble to burst, it is better to stand on the shoulders of giants to get a real ticket.
Sam Altman, the founder of OpenAI, also expressed a similar view. He believes that it is expected that there will be several large-scale basic models in the future, and developers will develop AI applications based on these basic models. But the current situation is still that a certain company develops a large language model and then opens the API for others to use.
For start-ups and other companies with insufficient reserves, they can neither afford huge investment nor need to start from scratch. The most cost-effective and correct path is to use the basic models that have undergone a lot of calculation and training, and then carry out targeted training according to their specific business applications.
Sam Altman believes that the 1% training done by these companies is crucial for the practical application of large language model technology. And, “these startups are going to be very successful.”
Please note that this article is from Xiong Chumo. For reprinting, please contact the original author for authorization
Original title: ChatGPT Dissuades Entrepreneurs