字节跳动训练万亿级 AI 模型，意图挑战 Anthropic

英文标题：ByteDance trains massive AI model in bid to rival Anthropic

TikTok 母公司字节跳动正在训练一款拥有10万亿参数的大型 AI 模型，以期在能力上挑战 Anthropic 等领先实验室。字节跳动旗下的视频生成模型 SeeDance 已跻身全球最先进之列，面向消费者的 Doubao 月活跃用户数达到3.24亿，位居中国第一。过去三年，字节跳动大举投资数据中心、AI 研究人员、企业云服务 Volcano Engine 及定制 AI 芯片。由前 Google DeepMind 科学家吴永辉领导的 Seed 团队坚持自主研发、避免模型蒸馏，管理层认为这是最终超越竞争对手的必要路径。

打开原文相关 2 · 机会 0

中文

字节跳动是爆款短视频平台 TikTok 的母公司。由于其模型大多处于闭源状态，与许多中国同行不同，字节跳动在 AI 开发方面一直保持低调。其最新的 SeeDance 模型已跻身全球最先进的视频生成模型之列，而其面向消费者的旗舰模型 Doubao 则以3.24亿月活跃用户数成为中国最受欢迎的模型。

过去三年，字节跳动在 AI 领域的投入比其他任何一家中国科技巨头都更为激进，持续扩建数据中心网络并招募研究人员。该公司还加大了对云服务部门 Volcano Engine 的投入，该部门向企业销售 AI 解决方案。字节跳动也有开发定制 AI 芯片的雄心。

其模型开发团队 Seed 由前 Google DeepMind 科学家吴永辉领导，在中国及海外约有2,000名成员。该团队包括核心研究人员、基础设施工程师、数据标注团队以及翻译人员。

据其中一位知情人士透露，其模型开发还采取了一种更为独立的路径，不涉及从其他实验室的现有模型中进行“蒸馏”（“distilling”）。这一做法已实施一年多，一些人认为，这导致其相较于竞争对手的发展速度更慢。

模型蒸馏是指通过训练较小的模型复制较大“教师”模型的知识与输出，将大型、复杂的 AI 模型压缩为更小、更快模型的过程。

由创始人张一鸣领导的字节跳动管理层认为，只有独立研发才能打造出性能超越竞争对手的模型。

据该知情人士称，张一鸣两周前在一次内部会议上重申了这一立场。他告诉 Seed 团队，应以长期实现“世界领先的模型能力”为目标，不必过度担心短期内落后于竞争对手。

中国媒体 Latepost 和 The Information 此前首次报道了张一鸣在近期内部会议上的相关言论。

字节跳动未回应置评请求。

English / 原文

ByteDance, the parent of viral video platform TikTok, has kept a low profile in its AI development as its models are mostly closed, unlike many of its Chinese peers. Its latest SeeDance model ranks among the most advanced globally in video generation, while its flagship consumer-facing model Doubao is the most popular in China with 324 million monthly active users.

Over the past three years, ByteDance has invested in AI more aggressively than any of the other Chinese tech giants, building out its network of data centers and hiring researchers. It has doubled down on its cloud unit, Volcano Engine, which sells AI solutions to enterprises. ByteDance also has ambitions to develop custom AI chips.

Seed, its model development team led by former Google DeepMind scientist Wu Yonghui, has about 2,000 members in China and overseas. The team includes core researchers, infrastructure engineers, a data labeling team, and translators.

Its model development has also implemented a more independent approach that does not involve “distilling” existing models from other labs, according to one of the people. This approach has been in place for more than a year, which some believe has led to its slower development versus rivals.

Model distillation is the process of compressing a large, complex AI model into a smaller, faster one by training the smaller model to copy the knowledge and outputs of the larger teacher.

ByteDance’s management, led by founder Zhang Yiming, believes only independent development can result in a model that outperforms rivals.

Zhang reiterated his stance in an internal meeting two weeks ago, where he told the Seed team to target “world-leading model capabilities” in the long run without getting too worried about falling behind in the near term, according to the person.

Chinese media Latepost and The Information first reported Zhang’s comments from the recent internal meeting.

ByteDance did not respond to a request for comment.