ByteDance launches Seed Edge and doubles down on AGI Research
There were rumors on the market in the second half 2024 that ByteDance (19459033) was planning to set up a “large-scale research institute.” Subsequently ByteDance replied by saying: “There is a plan to strengthen the research on large models but no decision has yet been made to create an independent institution.”
The AGI research team at ByteDance is finally known to have emerged recently:
ByteDance launched a new research project in late January under the codename “Seed Edge.” The goal is to conduct more long-term, fundamental AGI frontiers research than pretraining and large-model iteration. Seed Edge has identified five major research directions.
The current large-scale modelling field in China has an advantage in that it relies on excellent engineering and learning capabilities to keep up with the latest achievements in the industry. In certain directions, models can be developed which are able to match or even approach the performance of leaders, with less resources and greater efficiency.
The establishment of Seed Edge sends a completely different message. It indicates that ByteDance – a Chinese internet giant – is trying to explore original AI innovations: using more talent and resources to explore uncharted boundaries of intelligence towards approaching the future AGI.
ByteDance’s most valuable technology and business direction is AI.
Zhang Yiming is a founder of ByteDance and he has told us that he attaches a high value to the investment in AI research. He reads papers, pays attention to the technical details, chats to top AI researchers and encourages ByteDance AI Research team to explore and study basic topics.
We understand that ByteDance in Singapore has a dedicated team of researchers to help Zhang Yiming understand cutting-edge technology and discuss research plans. Feng Jiashi is a former professor of the National University of Singapore, who is now a research at ByteDance. He has been coaching Zhang Yiming regularly since 2023.
A second investor told us that, by the year 2023, Zhang Yiming began to meet with important AI authors, including doctoral candidates who hadn’t yet graduated.
ByteDance is known to have been considering the establishment of a research organization for some time. After forming a feasible strategy, internal communication, and obtaining opinions, took only one week.
A person close to ByteDance stated that the company was among the Chinese large corporations most willing to adapt with an extremely rapid action speed when faced with the AI transformation.
We learned that Seed Edge’s five major research areas are relatively long-term, and will not pursue immediate results like iterating model.
According to reports, Seed Edge is first going to operate as a virtual team to explore research directions that are more uncertain.
In the past year, ByteDance demonstrated its ability to keep pace with advanced models and produce product at an user scale by releasing large models and AI apps such as Doubao App.
ByteDance has released the Doubao 1.5 Pro Model yesterday (January 22), a month after iterating the Doubao Basic model. It claims that it can achieve performance comparable with top-tier large dense models for pre-training, but with smaller activation parameter.
SEE OTHERWISE: ByteDance launches Doubao-1.5 Pro Model with 50% Gross Margins
ByteDance Doubao models now cover different directions, such as language, speech, vision, or performance, with different versions focusing on efficiency or performance. They are also comparable to leading models in terms of capabilities. Doubao’s daily active users have exceeded tens and millions, making it China’s largest large-scale application for models.
We understand that ByteDance’s large model research group has high goals to not only keep up with the latest models in the industry, but also to develop the next-generation leading models.
ByteDance hopes to continuously improve intelligence levels through Seed Edge.
This is a “blurry” objective compared to models that are constantly updated. The challenges are not just due to a lack of data and computing power, but also to fundamental goal judgements and path choices.
ByteDance began investing in AI fundamental technologies even before the formal launch of Seed Edge. According to our preliminary estimates, ByteDance’s research team will publish over 100 AI related papers by 2024.
There is still a big gap between ByteDance’s number of papers and those of large tech companies overseas: Google has selected over 120 papers at NeurIPS while Microsoft has also selected more than 100.
ByteDance is a company that conducts cutting-edge exploration. Its obvious advantage lies in being ‘rich and resourceful’, but it’s more important to see if resources can be converted into talent density. This will provide talents with innovation and space for research, and slowly form a soil where more outstanding talents can grow.
ByteDance is a very attractive AI talent in China.
The company has taken a number of actions to build up its talent reserve. In May of last year, Doubao’s large-scale model team introduced the “Top Seed Talent Program”which aims to bring together top AI researchers among current and recent PhD students in order to tackle world-class AI issues. This team has been selected to present 57 papers at international top conferences, including ICLR, CVPR and NeurIPS. They have also released open-source software projects that have received over a million downloads, and popular projects on GitHub.
The “Seed edge” program, if implemented on a long-term basis, is expected to drive AI research from a short-term approach to gradual progress in both the short and long term. This will further close the gap between the advanced levels of the international AI industry.
