Zhang Yiming: Seed's training data is several orders of magnitude less; ByteDance won't be able to compete with OpenAI and Anthropic until at least 2029.

Beating AI News Flash: Zhang Yiming recently gave a timeline at an all-hands meeting of the Seed team: ByteDance will not be able to compete with OpenAI and Anthropic in the frontier AI field until 2029 to 2030 at the earliest.

Zhang Yiming specifically pointed out that the pre-training dataset used by the Seed large language model is several orders of magnitude smaller than that of top competitors.

ByteDance has actually already been dealt a lot of good cards in the AI field. Doubao is the AI chat application with the largest user base in China, and Seedance has also ranked among the world's top video generation models. It is precisely the general-purpose language model that has not kept up. When Seed2.1 was released in June this year, the team itself admitted that the model still has obvious shortcomings in highly difficult open-ended tasks and frontier scientific research problems.

ByteDance is prepared to invest enormous funds for this. In May this year, the company discussed raising full-year capital expenditure to as much as $70 billion, mainly for building AI data centers and other infrastructure. In August, it was also reported that Seed is pre-training a large model with up to 10 trillion parameters, a scale exceeding all publicly disclosed large models in China.

Zhang Yiming is also unwilling to take shortcuts through distillation. So-called distillation means using answers generated by leading models such as Claude to train one's own model in turn. In August this year, he clearly stated that this method can indeed improve performance in the short term, but what is learned is ultimately others' existing capabilities, and at best one can only keep approaching them, making it very difficult to truly surpass them. Internally, Seed also uses API detection and training data filtering to restrict the outputs of third-party models from entering the training set.

In order to catch up, ByteDance is also continuously recruiting and retaining talent. Seed currently has about 1,000 R&D staff, led by Wu Yonghui, whom Zhang Yiming poached from Google DeepMind last year. ByteDance has also specially launched virtual shares linked to the Doubao business, "Doubao shares," and eligible employees can also choose to receive their year-end bonuses as Doubao shares to share in the gains brought by business growth.

ByteDance plans to launch a personal AI assistant similar to Meta Muse in the coming weeks, allowing users to interact with AI through voice and avatars and have it complete tasks on their behalf. Doubao's internal personal assistant project, codenamed "Spell," began internal testing as early as April, and was later not officially launched due to reasons such as the release of Doubao phones. After Meta Muse went viral, ByteDance accelerated progress and was rushing development before the National Day holiday, planning to further integrate Spell with Doubao.