ByteDance Launches Seed Edge, Doubling Down on AGI Research

In the second half of 2024, there were rumors in the market that ByteDance was planning to establish a ‘large-scale model research institute.’ Subsequently, ByteDance responded by stating: ‘There is a long-term plan to strengthen research on large models, but no decision has been made to establish an independent institution.’

It is understood that recently, the AGI research team of ByteDance has finally emerged:

In late January, ByteDance officially established a research project with the codename ‘Seed Edge,’ with the core goal of conducting more long-term and fundamental AGI frontier research than pre-training and large model iteration. Seed Edge has outlined five major research directions.

One advantage of the current large-scale modeling field in China is its reliance on excellent learning and engineering capabilities, quickly keeping up with the latest industry achievements: in relatively certain directions, developing models that approach or even match the performance of leaders with fewer resources and higher efficiency.

The establishment of Seed Edge, on the other hand, sends a different signal, indicating that ByteDance, a Chinese internet technology giant, is attempting to explore original AI innovation: using more resources and talent to explore uncharted intelligent boundaries towards approaching the future of AGI.

AI is currently the most valued technology and business direction for ByteDance.

People close to ByteDance have told us that the founder of ByteDance, Zhang Yiming, attaches great importance to strengthening AI research investment. He reads papers himself, pays attention to technical details, chats with top AI researchers, encourages the ByteDance AI research team to explore and study fundamental topics.

We understand that in Singapore, ByteDance has a dedicated research team to assist Zhang Yiming in understanding cutting-edge technologies and discussing research plans. One of them is Feng Jiashi, a former professor at the National University of Singapore who is now a researcher at ByteDance. Starting from 2023, he has been coaching Zhang Yiming frequently.

An investor also told us that by the end of 2023, he found that Zhang Yiming had started visiting important authors of AI papers one-on-one, including doctoral students who had not yet graduated.

It is understood that ByteDance has long been considering establishing a research organization. However, after forming a feasible plan, internal communication and soliciting opinions only took one week.

Someone close to ByteDance said that when facing the transformation brought about by AI: ‘ByteDance is among the Chinese large companies most willing to adjust with extremely fast action speed.’

We learned that Seed Edge’s five major research directions are relatively long-term and will not pursue quick results like iterating models.

According to reports, Seed Edge will first operate in the form of a virtual project team to explore research directions with greater uncertainties.

Over the past year, ByteDance has demonstrated its ability to keep up with advanced models and produce products at a user scale through the intensive release of large models and AI applications such as Doubao App.

Just yesterday (January 22), one month after iterating on the Doubao basic model, ByteDance released the Doubao 1.5 Pro model that integrates multimodal capabilities, claiming it can achieve performance comparable to top-tier large dense pre-training models with smaller activation parameters.

SEE ALSO: ByteDance Launches Doubao-1.5-pro Model with 50% Gross Margin

By now, ByteDance’s Doubao series models have covered different directions such as language, vision, and speech, with different versions emphasizing performance or efficiency. In terms of capabilities, they are also on par with leading models. The daily active users of the Doubao app have exceeded tens of millions, making it China’s largest large-scale model application by far surpassing the second place.

We understand that the high goals set by ByteDance’s large model research team are not only to keep up with the most advanced models in the industry but also to develop next-generation leading models.

The establishment of Seed Edge means that ByteDance’s ambitions for AI go beyond just models and products; they also hope to gain innovative capabilities for continuously improving intelligence levels.

Compared to constantly updating models, this is a more “blurry” goal. The challenges it faces are not only due to insufficient data and computing power but also fundamental goal judgments and path choices.

Before the formal establishment of Seed Edge, ByteDance had already started investing in research on AI fundamental technologies. According to our preliminary statistics, ByteDance’s research team is expected to publish over 100 AI-related papers by 2024.

In terms of the number of papers, there is still a gap between ByteDance and large overseas tech companies: at NeurIPS, Google has selected more than 120 papers, while Microsoft also has over 100.

Conducting cutting-edge exploration, ByteDance’s obvious advantage is being ‘rich and resourceful,’ but more importantly, whether resources can be transformed into talent density, providing talents with innovation and research space, and gradually forming a soil for nurturing more outstanding talents.

From the current situation alone, ByteDance has a strong appeal to AI talents in China.

The company has taken many actions in talent reserve. In May last year, the Doubao large-scale model team launched the “Top Seed Talent Program”, mainly targeting top AI research talents among current and recent doctoral students to jointly tackle world-class AI topics. Over the past few months, this team has had 57 papers selected for international top conferences such as ICLR, CVPR, NeurIPS, and has also released open-source projects with over a million downloads and popular projects on GitHub.

If the “Seed Edge” program can be implemented in the long term, it is expected to drive AI industry research from short-term layout to gradual advancement in both short and long terms, further narrowing the gap with international advanced levels.