Not willing to just be an AI application factory, ByteDance makes up a big model
Source: light cone intelligence
Text | Hao Xin
Editor | Wang Yisu
Light Cone Intelligence gets the latest exclusive news. Following the first launch in August 2023, in April this year, ByteDance’s big language model Lark Big Model is about to usher in an important version of the upgrade.
The update frequency of eight months is really rare in the large model market. In such a large model with internal volume, bytes seem to be somewhat different, and the sprint is as rapid as a hundred rivers of intelligence. In the early stage, a large model was released and upgraded on average once a month.
But looking back on the past year, Byte didn’t want to lie flat.Different from the thinking of large model companies at home and abroad that "the large model is used first, and then the product is applied", byte selection, which is known as "application factory", bets on AI native applications hard, accurately and quickly on the basis of having a lark large model.
First, from the organizational level, in the second half of 2023, Byte began to integrate the big model team and other business groups, and set up a new department, Flow, which focused on AI innovation business. The department also became the vanguard of Byte in AI technology research and product development.
Another is like a huge manufacturing machine. In more than half a year, Byte frantically delivered more than a dozen AI products to domestic and foreign markets. At the same time, in the process of product development, a number of basic technology research results of image generation and video generation have been accumulated and published.
Nowadays, Byte is like coming back after a full circle in the product and market fields, and looking at the basic big model itself from a more diversified perspective.
As a upstart of the Internet, Byte has not experienced the wave of AI labs in 2016. Instead, it has combined AI algorithm technology with graphics and video to create today’s headlines and explosive applications in Tik Tok.
According to the latest disclosure, Byte’s Q3 revenue in 2023 was $30.9 billion, which has surpassed Tencent. After the new king ascended the throne, the problem in front of Byte is also very clear. How to make use of his own advantages to make up for the missed lessons and catch up with the new wave of AI?
AI application manufacturing factory, with eleven products launched in half a year.
"Strengthen the sense of crisis, always start a business, and escape the mediocre gravity." At the beginning of 2024, Byte CEO Liang Rubo set a goal for the whole year.
To say that the most suitable entrepreneurial attribute mentioned by Liang Rubo is the Flow established by Byte last year.
In August last year, the lark model was officially released, and at the same time, it was announced that it began to test the AI dialogue product "bean bag" externally. Bytes immediately went to the battlefield of the next application, and the lark model gradually disappeared in the overwhelming news and turned into the "base" behind a series of AI products such as bean bag.
In September, the newly established Flow adapter became the main force. According to the publicly reported information, Byte has deployed a large number of senior management talents at the helm. Zhu Wenjia, the head of the big model team, is also responsible for the business line of Flow. Hong Dingkun, vice president of Byte Technology, is responsible for the technical line of Flow. Zhu Jun, vice president of Byte Products and Strategy, is responsible for the product line of this department. At the same time, Qi Yuanjun, vice president of products of Feishu, has joined us.
Under the big advance of bytes, it has realized the situation of multiple applications and multi-flowering. According to incomplete statistics of light cone intelligence,From last August to now, in more than half a year, Byte has tested and launched eleven AI application products at home and abroad, and an image product Picpic has not yet been launched, among which eight products are led by the Flow team.
From the perspective of product types, the direction of byte selection is mainly concentrated in four main directions: Chatbot, virtual role, Agent and image, which basically covers the application-level entrepreneurial direction of last year’s fire.For example, there is an efficient product ChatGPT in the field of Chatbot, a considerable number of users in the field of virtual characters, GPTs introduced by OpenAI in the direction of Agent, and so on.
Attacking from many directions and blossoming everywhere, this picture seems to have returned to the eve of Tik Tok’s birth overnight. When entering the AI application track, Byte once again adopted the strategy of "internal horse racing". In China, the capability was provided through the Lark model, and in foreign countries, the service was provided based on GPT. Many times, overseas is used as a test field, and similar products are launched in China first, so as to run market and user data and prepare for domestic launch.
In addition, bytes are also very clear about their own advantages. Tiktok and Tiktok, two large traffic pools, have naturally become sharp tools for byte drainage and innovation. According to the light cone intelligent observation, Byte has specially set up a "live room with goods" for bean bags, introducing the functions of bean bags to users who enter the live room, and promoting that the App can be downloaded for free. In addition, Byte also invited a large number of talented people from Tik Tok to be the platform of bean bags, and implanted new functions of bean bags in the jokes.
Perhaps it is because of the traction of traffic that the bean bag released late has surpassed Baidu’s ERNIE Bot in terms of awareness and monthly activity. According to reports, it is revealed by some sources that the monthly activity of bean bags has increased to 2 million in December last year, and doubled on this basis in January 2024. The monthly average daily activity of bean bags has already surpassed that of ERNIE Bot.
From the perspective of byte’s own business, in addition to Flow, flying books, clipping, byte Singapore company, huge engine, and vigorous education are also testing water and launching AI tools and products. Since the advent of Sora, the AI video track has rekindled the war, and the byte clipping business has also been placed with high hopes. Zhang Nan resigned as the CEO of Tik Tok Group to lead the clipping team. From the current point of view, the clipping has been launched with the functions of AI cloning timbre, AI drawing, AI drawing, AI generating oral broadcast and so on.
It may only be a matter of time before Byte launches its products on AI video. In terms of technical reserves, Byte has accumulated the video generation model MagicVideo-V2, video editor Boximator, and video generation research PixelDance, and has dug up counterpart talents from Google’s video generation model team.
It has been reported that Byte is secretly developing a number of products in the field of AI big model, including multi-modal digital human products, AI video products and AI video products.
Byte is not eager to change the old business, but uses the ability of a single point to radiate to internal horse racing. On the one hand, it is to test the market, and on the other hand, it is also exploring how to embed AI into the original business flow.
For example, the "smart partner" was introduced in the flying book business line, and the traditional workflow was changed by the technology of Agent, and the functions of content creation, content summary and data analysis in the office scene were realized, so as to reduce costs and increase efficiency for individuals and enterprises. Where to play the ability of Agent, where to use the ability of literary drawing, and which scenes to call the ability of dialogue reasoning, these should be tested during the landing process.
Xie Xin, CEO of Feishu, once said at the press conference that the ability of AI will definitely become very strong in the future, and great changes will take place in all walks of life. However, at present, the ability of AI is still very limited, which may not make every task as expected. "What is more important now is to make yourself AI Ready first."
Return to the main battlefield and make up the big model.
Liang Rubo reflected at the annual meeting at the end of 2023, "Bytes are not as sensitive to technology as startups, and GPT was not discussed until 2023. The big model startups that have done well in the industry were founded from 2018 to 2021. "
The implication of Liang Rubo is that the big model of bytes is slow.
In March last year, Baidu released ERNIE Bot, and then Huawei and Ali quickly followed suit to release the big model. It was not until mid-August that the big model of Tik Tok lark arrived late.
The news about the big model team of Byte can be traced back to January last year. 36Kr reported that Byte formed the first big model team at that time, including the big model team of language and the big model team of pictures. Among them, the language big model team is led by the byte search department, and the picture big model team is led by the intelligent creation team under the product development and engineering architecture department.
At that time, ChatGPT and Midjourney had already exploded. Perhaps it is to see these two types of products with different paths behind them, and then consider how to transform a series of products.From the initial formation of the team, Byte chose technology and products to walk on two legs. But the big model is the base of most AI applications. If you want to develop products, you must first have a big model.
Lark big model just assumed such a role. As soon as the big model landed, Bytes quickly started the research and development of AI applications at the same time, but the immature performance of the big model also affected the landing of AI applications to some extent.
First of all, from the time line, the products with the same function went online earlier in foreign countries than in China. For example, the domestic opening time of buttons on GPTs-like platforms is two months behind that of foreign countries. Even after the domestic buttons are launched, many users in China choose the foreign Coze because they can directly call GPT-4 turbo.
The ability of the model will also be reflected in the use effect of the product. For example, the overseas version of CapCut has recently launched a text-generated video function, but with user feedback, this function is not satisfactory in terms of video clarity, understanding of prompts and waiting time for generation.
This led to the killing of bytes in the product battlefield and had to go back and make up the big model.
However, for bytes, it doesn’t make much sense to benchmark OpenAI, and thinking about the big model route that suits you is the way out.
Judging from the public information, the force point of bytes in the AI direction is still concentrated in the fields of images and videos.As for the big model, Byte has introduced the Lark, a universal big language model, and BuboGPT；, a multi-modal big model that supports text, image and audio. In the direction of image vision, the MagicVideo-V2 video generation model launched last year has caused a wave of heat at home and abroad, which can make people in still pictures move. The research after bytes continues to extend in the direction of video, including how to control the action of characters by inputting text and how to improve the dynamic effect of video.
From this point of view, bytes still learn from the practice of OpenAI, that is, outside the GPT big model, all kinds of single-point capabilities are pulled to the full-Whisper model in voice direction, DALL·E series in image direction and Sora in video direction.
Because the breakthrough of single-point ability also depends on the ability of the underlying big model, especially Sora provides an idea that the framework of Transformer big model can be combined with the image generation model, which means that the reasoning and understanding ability of the big model will affect the final video generation logic. Therefore, on the basic big model, all big model companies, including bytes, still can’t be ignored.
In addition to complementing the ability in the video model, Byte has also made great efforts in team talent allocation.At the beginning of the establishment, Zhu Wenjia, the former head of TikTok technology in Singapore, was transferred to lead the large model team, and later he was also responsible for the business line of Flow. A vanguard department, Flow, has assembled the vice presidents of technology, products and strategy, and flying books. Recently, it was revealed that Jiang Lu, the research leader of Google video generation model VideoPoet, joined the intelligent creation team. It is reported that the idea of VideoPoet is very similar to that of Sora world model.
After Sora burst into flames, many people compared it with the cut film personally led by Zhang Nan, the former CEO of Tik Tok, but in fact, Zhang Nan still paid more attention to the product level. The real byte version of Sora must be born in the team led by these scientists and technical leaders.
Advertising & cloud business, AI has a greater impact on bytes than expected.
Last year, Zhang Yiming, the founder of ByteDance, spent all his energy on AI, which is of great significance to the biggest beneficiary who benefited from the previous generation of AI technology (recommended algorithm).
What AIGC ultimately generates is content, which is naturally a change in the form of content production. Compared with other companies’ original business attributes such as e-commerce, search and social networking, Tik Tok’s gene itself is the content. Therefore, the strategic significance of this big model wave to Byte may be far greater than that of other companies.
Although ChatGPT-4 has just been born for one year, the big model and AIGC technology have only taken the first step, but the imagination about the future business growth of the company may just be opened.
Take Baidu as an example, its just-released annual financial report data for 2023 shows that AI has brought practical benefits to Baidu. In the year of All in model, Baidu search, advertising and other old businesses have been revitalized, and the once depressed cloud computing business has also seen new growth momentum.
In 2023, Baidu’s core revenue was 103.465 billion yuan, and the net profit attributable to Baidu’s core was 27.4 billion yuan, a year-on-year increase of 38%; The big model is bringing more and more business income to Baidu. In the fourth quarter, the revenue growth brought by the big model alone reached 660 million yuan, and the revenue of Baidu AI Cloud also reached 8.4 billion yuan. According to Morgan Stanley’s forecast, Baidu’s advertising revenue is expected to achieve a year-on-year growth of 7% in 2024.
Although for bytes, this growth is far from our own eyes, but many business lines still have certain reference and reference value.
According to the light cone intelligence, influenced by the big model wave, the revenue growth of Volcano Engine, a byte cloud computing business, is also considerable. Thanks to a large number of NVIDIA’s GPUs, many large model startups are willing to take the initiative to join the ecology of Volcano Engine, thus boosting the growth of Volcano Engine.
Although the first wave earned the first bucket of gold by selling computing power, for Volcano Engine, customers who pay for cloud computing business again by using the algorithms of these big model companies in the future are more attractive.
For the advertising business on which Bytes depend, the influence of the big model has not yet been exerted. At present, several companies, including Baidu and Netease Youdao, have mentioned that large models can promote the conversion of their advertisements. This is undoubtedly good news for Tik Tok and today’s headlines.
In order to improve the marketing efficiency, Tik Tok’s marketing platform Giant Engine also released the automatic technology brands UBMax on January 23rd, based on three scenarios: application download, clue retention and e-commerce drainage.
In addition, Volcano Engine has also launched a very byte-featured product-Volcano Engine Intelligent Creation Cloud, which is an intelligent SaaS platform to generate videos in batches, mainly helping e-commerce sellers to generate commodity materials in batches. Although there are many companies making such products in the market at present, and some of them are deep partners of Tik Tok, for companies with both technology and scenarios, they can only do it in bytes.
To sum up, on the whole, Byte’s investment in AI tends to be conservative, paying more attention to products that can generate value for actual business, while its investment in cutting-edge technologies has just started.
This is also related to Byte’s focus strategy in the past year. In the past year, Byte has almost retired or cut all business lines unrelated to core business, such as Pico and games.
When the core business is strong, development can cover up all problems. Byte’s quarterly revenue can still maintain a year-on-year growth of more than 40%, which has envied all other domestic companies, but for Byte and Zhang Yiming, there is still a dream.
