ByteDance is preparing to release an AI agent in the next few weeks, with the product form similar to Meta’s Muse.
Zhang Yiming personally spoke about the plan at a recent all-staff meeting, with an audience of researchers, engineers, and support staff from the AI team. The judgment he gave is quite calm: Byte has the ability to compete with Anthropic and OpenAI at the forefront, but it will not truly tie up until at least 2029 or 2030.
Some key information:
One is money. Byte will invest up to US$70 billion to build the Seedance video AI platform. Video generation is Byte's most confident direction at present, and the positioning of the new agent this time is also to "consolidate the leading position of AI video."
The second is the route. He explicitly does not advocate taking shortcuts through model distillation. Distillation is to use the output of other strong models to train your own model. The results are quick, but the ceiling is also set by others. Not taking this path means that it is more expensive and slower, and it also means that he really wants his own cutting-edge capabilities.
Third, Muse was selected as the benchmark object instead of ChatGPT. The characteristic of Muse is that it can continuously run tasks in the background and work across applications. Byte has a complete product line of Douyin, Jianying, Jiemeng, and Doubao. The intelligent agent can really mobilize more things than many rivals.
As for the time point in 2029, it shows that he admits that there is still a generation gap between the basic model and the American head laboratory. On the other hand, the user base of Doubao and the video capabilities of Jimeng and Seedance are Byte's most valuable bargaining chips in the past few years.
When the product comes out in a few weeks, we can see whether it is just another chat portal, or whether it can really help people turn videos from scripts into movies.