ByteDance trains massive AI model in bid to rival Anthropic
Ars Technica Zijing Wu, Financial Times
ByteDance is quietly building a huge in-house AI push, skipping the shortcut of copying rival models. Founder Zhang Yiming says he's fine falling behind for now if it means eventually leading.
Based on reporting by Ars Technica, Zijing Wu, Financial Times — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
ByteDance doesn't talk about its AI work the way most Chinese tech giants do. While rivals loudly tout benchmark wins, the TikTok parent has kept its models mostly closed and its ambitions mostly quiet. Yet its SeeDance video-generation model ranks among the most advanced anywhere, and Doubao, its consumer chatbot, is the most-used AI app in China with 324 million monthly active users.
The scale of the bet behind those products is what stands out. Over the past three years ByteDance has poured more into AI than any of its Chinese peers, expanding data centers, hiring researchers, and pushing its cloud arm Volcano Engine deeper into enterprise AI sales. The company is even eyeing custom AI chips, a costly and technically brutal path that few outside the biggest labs attempt.
At the center of the effort is Seed, ByteDance's model-development team, led by Wu Yonghui, a former Google DeepMind scientist. Seed has grown to roughly 2,000 people spread across China and overseas, a mix of core researchers, infrastructure engineers, data labelers, and translators. What sets the team apart, according to people familiar with its approach, is a refusal to lean on distillation, the common shortcut of training a smaller model to mimic a larger one built by someone else. That independent-only rule has been in place for more than a year, and some inside the company think it's part of why ByteDance has moved slower than rivals.
Management, led by founder Zhang Yiming, sees that slowness as the price of building something that can actually surpass competitors rather than just shadow them. In an internal meeting roughly two weeks ago, first reported by Latepost and The Information, Zhang reportedly told the Seed team to aim for what he called world-leading model capabilities over the long run, and not to panic about lagging in the short term. ByteDance did not respond to a request for comment on the reporting.
My take — AI-written commentary, not fact-checked reporting
Betting everything on independent development while the rest of the field distills its way to fast wins is either visionary patience or a very expensive ego trip, and right now nobody outside ByteDance can tell which. Founders who tell teams to ignore short-term rankings are usually either sitting on something real or buying time to explain why they're behind, and history says both outcomes happen roughly as often as each other.
Read more about this at: Ars Technica