The door to listing is showing signs of a reboot, while Douyin's existing business has run out of imagination. Luckily, the oncoming Times Express ChatGPT is full of stories.
On February 9, some media broke the news that Douyin's artificial intelligence lab (AI Lab) is carrying out research and development similar to ChatGPT and AIGC (artificial intelligence generated content), and will provide technical support for the VR (virtual reality) department PICO in the future.
The relevant person in charge of PICO then responded that the relevant rumors were untrue, and PICO currently has no product planning to adopt similar ChatGPT technology.
ChatGPT is a large-scale language pre-training model launched by OpenAI, an American artificial intelligence technology company, which went live at the end of November last year. With its seemingly omniscient and omnipotent AI chat capabilities, ChatGPT attracted more than 100 million active users in just three months, prompting tech companies to rush to the same track.
In the United States, Microsoft, which has long bet on OpenAI, became the biggest winner of the global wave of ChatGPT, announcing at the end of January this year that it would invest another $10 billion in the company and plan to embed relevant functions in core products such as Office and Bing; Google, which is deeply aware of the epic challenge of the search engine, also hurriedly launched its benchmarking product Bard in early February.
Among domestic technology companies, BAT is vying to enter the game. Baidu announced on February 7 that it will launch a ChatGPT-like application "Wen Xin Yiyan", which is expected to complete the internal beta and open to the public in March; Tencent claims to have a layout in ChatGPT and AIGC; Ali also released the news that the Damo Academy version of ChatGPT is in closed beta. The second-tier small and medium-sized players are not far behind, and companies such as JD.com, Zhihu, Kuaishou, Xiaomi, and 360 have all been related to ChatGPT.
As the third pole of China's Internet, Douyin set up an artificial intelligence laboratory in 2016, and it is logical to catch up with this wave of ChatGPT. At the same time, PICO is the pioneer of Douyin's metaverse, if the two hot spots of the metaverse and ChatGPT are linked, even if no application scenarios can be found, it will help further develop the brand and increase sales.
But for Douyin, the value of ChatGPT and the general artificial intelligence behind it goes far beyond that. Whether it is for a short- and medium-term listing, or a long-term intergenerational competition in Internet technology, Douyin needs its own ChatGPT.
Douyin has been planning to go public for a long time, but the boots have not landed. In 2023, the warming of Chinese concept stocks and the Internet industry, especially the basic completion of rectification by Ant Group, as well as the resumption of Didi listing and new user registration, have injected new expectations into the entire industry. Douyin has completed a number of key actions such as group name change and CFO change last year, and the resumption of IPO this year is timely.
But the embarrassing thing is that Douyin is currently very large, but it is either an old business that has gradually entered a mature period, such as Douyin APP, today's headlines, etc., or a dou that cannot be supported, such as education, games, etc.; The group's efforts are also focused on businesses that can quickly generate cash flow, such as shelf e-commerce and life services. With the foundation of Douyin, although these businesses can contribute income, they are all businesses that can be seen at a glance, and lack the imagination space beyond the times, which is difficult to make investors excited.
If Douyin takes advantage of the unprecedented popularity of this wave of ChatGPT and comes up with its own general artificial intelligence roadmap, even if it is just a big pie for the time being, it is very likely to set off waves in the primary and secondary markets, boosting the declining valuation of the company. The soaring stock prices of Baidu, Zhihu and other companies have verified the capital market's pursuit of the concept of ChatGPT.
In the long run, AI technology represented by ChatGPT will be the key to Douyin's completion of the upgrade of the technical base.
Douyin has always used centralized algorithms for the distribution of graphic and video content, but now the efficiency of the algorithm is approaching the limit, which is reflected in the slowdown or even stagnation of major products such as Douyin and Toutiao. Considering the exhaustion of the dividend of new Internet users, Douyin can only make changes from the bottom to make the connection between the content pool and the user pool more accurate and efficient, and then restart the growth engine of the entire ecosystem.
The AIGC behind ChatGPT has exactly that potential. It uses artificial intelligence to complete the massive integration of existing non-standardized information, and realizes more refined personalized distribution in natural language chat. This is the information distribution path that Douyin has not yet dabbled in, and it is also a new engine that it needs to add to the recommendation algorithm.
Douyin's jump from machine recommendation to robot recommendation is not only the value at the business level, but also at the strategic level to resist the dimensionality reduction blow of artificial intelligence on machine algorithms. The wave of general artificial intelligence led by ChatGPT has arrived; And the best way to deal with the wave is to be a part of it.
A
Douyin noticed the technical trend similar to ChatGPT early on, but got up early and caught up late.
In July 2019, Li Hang, head of Douyin Artificial Intelligence Lab, said in a speech that there are good development opportunities for natural language dialogue, and four opportunities can be considered from the perspective of product design, including: task-driven dialogue robots, information robots, emotional robots, and game robots. Among them, the ideal form of "task-driven conversational bot" is very close to today's ChatGPT.
In the same month, OpenAI on the other side of the ocean achieved its first major breakthrough after the corporatization transformation, receiving a $1 billion investment from Microsoft, and also opened up the commercialization path with the help of Microsoft's Azure cloud computing platform.
It can be said that in the research of natural language large models, Douyin AI technicians are quite forward-looking, and their predictions of product forms are also consistent with future development. In terms of funding, although Douyin is not as good as Microsoft, it is already a Chinese Internet upstart, and it is not a problem to pay for AI research.
However, compared to OpenAI, which can be all in general artificial intelligence without distractions, the Douyin AI team has a much more complex KPI and increasingly tends to cooperate with the business unit. This ultimately leads to a vastly different growth curve between the two.
Judging from public information, Douyin has clear requirements for the landing of AI technology within the group. Its AI Lab's website claims that its research is focused on developing new technologies that serve byte-content platforms; Specific fields include natural language processing, data mining, computer vision, machine learning, etc., and are closely aligned with various business sectors of Douyin.
For example, in the area of natural language processing, the website lists several application scenarios: Byte Translator, which provides translation services for all ByteDance products; Xiaomingbot, an AI writing robot, a news writing robot involved in sports, finance, and current affairs; and search services for Toutiao and TikTok. In terms of computer vision, Douyin applies AI technology to content moderation, short video recommendation system, football game understanding and awkward dance machine.
OpenAI, by contrast, was born with the goal of general artificial intelligence, which is to enable machines to think like humans and engage in a variety of activities. If Douyin does AI to develop technology for a certain application scenario, then OpenAI is the opposite, it first tries to make AI as human-like as possible, and what it can do is completely left to the market and the public to explore and test.
As a result, although the AI department of Douyin has achieved a lot in recent years, it is concentrated within the group as a whole, and does not yet have the epoch-making technical power to break the circle. And in many cases, even Douyin itself is not proficient in the use of its own AI capabilities.
Taking natural language processing as an example, this is one of the key research objects of the Douyin AI team, and the group also tried to support Wukong Q&A around 2017, and invested 1 billion yuan to subsidize respondents for two consecutive years. However, in this unsuccessful Q&A community battle, Douyin AI did not play a significant role, and was only suspected to play a small role in the automatic generation of hot issues.
The limited application scenarios in turn drag down the improvement of technical strength. According to statistics from market research firm Statista, among the 7 companies with the most AI patents in the world in 2021, Tencent, Baidu and IBM are in the top three, while Douyin is not on the list.
On the other hand, OpenAI alternately stands at the top of the global artificial intelligence industry with Google. Since 2018, the GPT (Generative Pre-Training Transformer) model has undergone three iterations, of which the GPT-3 model has become today's ChatGPT after fine-tuning, which can generate almost any type of text, including chat, Q&A, papers, marketing plans, poems, scripts and even code, with nearly endless application possibilities.
In the latest round of funding, OpenAI was valued at $29 billion, equivalent to about 1/10 of TikTok, and the growth prospects are far from over. The ChatGPT fire may prompt Douyin to re-examine its AI strategy and give the AI team more time and space.
B
If Douyin focuses on general artificial intelligence, the direct beneficiary within the group will be Toutiao.
Since the development of today's headlines, the growth space has been extremely limited, and the person in charge has changed many times. But after all, it is the dragon of the entire Douyin, and it is also the second largest active user pool, and hundreds of millions of MAUs cannot give up easily.
In the past year or two, Douyin has tried to activate Toutiao from the supply side, focusing on the creator ecology, mainly to provide more traffic incentives and cash subsidies. Such activities may have a partial effect, but they will not really rewrite the overall situation in which today's headline growth has largely stagnated.
The fundamental problem is that today's headline users, after being "fed" by the platform's recommendation algorithm for many years, are immersed in the information cocoon, and it is increasingly difficult to become interested in content that goes beyond cognition and taste. This decline in demand-side dynamism can easily lead to a significant reduction in the effectiveness of any operational action.
To break the deadlock, Douyin must make changes on the demand side. It has previously tried to do Q&A, search, etc., all in the hope of allowing users to get information more actively, rather than simply waiting for the platform to cook. However, limited by various shortcomings, Douyin's attempts have not achieved significant results.
Today, the popularity of ChatGPT has allowed Douyin to see a new way to activate the Toutiao content pool and increase user activity.
Let's look at the content side first. Large natural language models require huge amounts of text data for training: OpenAI's GPT-1 has a data size of 4GB, GPT-2 increases to 40GB, and GPT-3 expands to 2TB. In general, the more data and the higher the quality, the stronger the final trained model.
For Douyin, this means that the content pool of today's headlines that has been precipitated for ten years suddenly has broad application prospects. As network information, many of these contents have lost their timeliness, but because many of them have better text and information value, they are still nourishing for large model training and continue to exert residual heat.
Let's look at the user side. Douyin's algorithm recommendations can provide insight into user preferences, but it is difficult to detect the high-level needs of humans to obtain structured information and system knowledge, which is one of the underlying reasons why it can't do Q&A.
But ChatGPT opens up a new paradigm for information acquisition: users design and ask questions, AI integrates content, and provides instant feedback in the form that users expect. It offers an experience that is very different from older tools such as algorithmic recommendations, search, and Q&A communities, while at the same time improving efficiency and accuracy. This personalized information integration and output capability based on general artificial intelligence is the ultimate key to improving the activity of graphic and text users on Douyin.
In the long run, large models such as ChatGPT are not only applied to text information, but also inevitably extended to rich media forms such as pictures, audio, and video. If Douyin makes its own ChatGPT, the beneficiaries will be the entire Douyin camp in addition to today's headlines.
Taking the "Graphics-to-Video" function as an example, Zhihu has previously launched the function of producing videos based on text, but the effect is not satisfactory; Douyin also already supports entering text and matching pictures to generate short videos, but there are not many uses of AI in it, and the quality of the finished product mainly depends on the creator himself.
Artificial intelligence promises to change that. OpenAI previously released DALL-E, a model based on text-generated images, and domestic companies such as Tencent and Baidu have also launched their own AI drawing tools. Douyin released an AI painting function based on an open-source model in 2022, which has been used by nearly 28 million people in total. For general artificial intelligence, it is only one step away from generating text to pictures to video.
For Douyin Group, adding general artificial intelligence, even if it is temporarily impossible to produce the same magnitude of ChatGPT, can bring new stories that the capital market craves, and then add chips to the listing. After all, this action itself shows that Douyin is not an ordinary company that only focuses on "fattening and thinning", and still retains the ambition and ability to explore the frontier of science and technology.
C
However, the research and development of general artificial intelligence takes a lot of time, and the annual capital investment is also hundreds of millions. This is not a small burden for Douyin, which is spending more and more cautiously.
According to industry estimates, the current single training hardware cost of large language models has exceeded 30 million yuan, and it is still climbing with the enlargement of the model volume. Coupled with other costs, the giants need to prepare tens of billions of dollars to finally decide the winner.
Even OpenAI, backed by Microsoft, has to balance model accuracy with cost. It was trained on GPT-3 only once; Despite the errors in the AI model, OpenAI decided it was too costly to train the model again.
As a Chinese company, Douyin also faces the problem of difficulty in obtaining advanced chips. At present, NVIDIA is the main chip supplier for the world's mainstream big prophecy model developers, and its customers include Microsoft, Google, Meta and so on. However, due to geopolitical factors, it is becoming more and more difficult for Chinese companies to buy the most advanced chips in the first time, and they are bound to encounter constraints in model training.
Like Baidu, Tencent, Ali and other companies, it is even more difficult for Douyin to make products on par with ChatGPT. But if you can kill a bloody path, the reward will be huge.
For Douyin, a large natural language model with independent intellectual property rights, in addition to bringing a leap to today's headlines and other sectors, can also output AI foundation capabilities and build a closed loop of business models. Whether it is content creation, customer management, code writing, repetitive work substitution, etc., a large number of application scenarios can be discovered.
Taking ChatGPT as an example, in addition to acting as a free chatbot, it is also integrated into many third-party products such as Microsoft Office as an AI component, and links small and medium-sized developers through APIs (Application Programming Interfaces), which explore more ways to play. In early February, ChatGPT also launched a paid version that allows users to pay a $20 monthly fee to access chatbots during peak hours.
Parent company OpenAI already generates a lot of revenue from it. The company expects revenue to reach $200 million in 2023 and more than $1 billion in 2024. If Douyin makes its own ChatGPT, after the high investment in the early stage, the road to commercial monetization will soon become wider and wider.
Developing its own general artificial intelligence without considering financial factors is also the only way for Douyin to avoid becoming a "dinosaur" of the times.
After the birth of ChatGPT, the entire technology industry saw the powerful power of AIGC, and the Internet content ecology may be rewritten, and Google and Baidu, which directly felt the threat, immediately announced similar products. Compared to the two companies, TikTok has completed the transition from graphics to video, and ChatGPT does not threaten it at least for now, but it does not mean that it can rest easy.
Douyin's housekeeping skill is the recommendation algorithm, which beats traditional portals and video websites by relying on the efficiency and experience improvement brought by "machines replacing humans". But recommendation algorithms also face a more powerful rival - AI, which replaces machines with "robots" that can more accurately grasp user needs and feedback in real time, thus making information distribution more accurate and efficient.
Just like the dimensionality reduction blow of algorithm recommendation to manual editing in those years, general artificial intelligence is also the existence of another dimension relative to algorithm recommendation. Now, with its debut through ChatGPT, it has shown its potential to disrupt the old landscape.
For Douyin, which holds the right to distribute information on the Internet, the wave of impact has not yet arrived, but it is already on the way. If you can't avoid the impact of the next era, then the best way to deal with it is to be part of the next era. Getting started with your own ChatGPT as soon as possible may be one of the things Douyin must do in 2023.
Resources:
Alphabet list, "Did the change of coach save today's headlines?" 》
Lei Feng.com, "Li Hang, head of ByteDance AI Lab: In the next 5 to 10 years, how should natural language dialogue seek breakthroughs? 》
Doing what AI does, "ChatGPT's Past and Present, and the Future"
Everyone is a product manager, "Deconstruction with network science, the logic behind Wukong's Q&A losing to Zhihu"
21st Century Business Herald, "Behind the expensive ChatGPT, the road to domestic GPU "nuggets" is long"