Key Takeaways
ByteDance’s Doubao family of AI models recently got two new members, PixelDance and Seaweed, which generate video clips from text prompts.
But looking further back, the platform’s origin story is mired in controversy. Before it became Doubao, “Project Seed” was accused of secretly using OpenAI’s technology to build a rival chatbot.
After ByteDance released the Doubao chatbot in August 2023, an exposé in The Verge reported that Project Seed had abused the company’s access to the GPT-4 API.
As laid out in OpenAI’s terms of service, the output from any of its models can’t be used to build competing AI systems.
Nevertheless, internal documents referred to by The Verge indicate that ByteDance used OpenAI outputs “during nearly every phase of development” of its own large language model (LLM).
Employees were even instructed to use “data desensitization” techniques to obscure the practice, which apparently led to ByteDance repeatedly maxing out its API access.
In the wake of the revelations, OpenAI suspended ByteDance’s account pending further investigation. Responding to the accusations, the Chinese firm admitted that a small group of engineers initially used OpenAI’s API but claimed that the data was removed from their model later in the development process.
Despite the controversy surrounding Project Seed, ByteDance has continued to develop new Doubao AI tools.
Since debuting the first Doubao LLM in 2023, ByteDance has expanded the platform to include different model sizes and specializations.
From a pureplay chatbot, Doubao has grown to incorporate image generation, voice capabilities, and, in the latest update, text-to-video functionality.
Unveiled at an event in Shenzhen on Monday, Sept. 24, Doubao-PixelDance can produce high-fidelity 10-second videos, while the Doubao-Seaweed model can generate clips of up to 30 seconds. Both models are expected to be released in October.
The announcement of ByteDance’s new video generators comes just a week after Chinese Big Tech rival Alibaba debuted its own take on the technology, the Tongyi Wanxiang Video Model.
Both firms have focused their attention on the domestic market, positioning their products as uniquely suitable for Chinese applications.
For example, Alibaba claimed its model is “optimized for […] Chinese aesthetics”. Meanwhile, Tan Dai, the president of ByteDance cloud unit Volcano Engine, noted that the new platform leveraged videos accumulated by the Chinese version of TikTok, Douyin.
The decision to focus on the Chinese market reflects the business environment in which Alibaba and ByteDance operate. But it also means they don’t have to compete with industry leader OpenAI on the international stage.
While rival developers are closing the gap, OpenAI’s currently unreleased Sora model remains the most capable video generator in terms of sheer fidelity and photorealism.
As the sector becomes more competitive, clandestine practices like stealing trade secrets are rising. With OpenAI’s leading position in the global AI market, Project Seed is unlikely to be the last attempt to misappropriate its technology.