字号·· | 护眼
彭博社

前字节跳动实习生创办的2亿美元AI实验室挑战李飞飞Ex-ByteDance Intern’s $200 Million AI Lab Takes On Fei-Fei Li

点「原文对照」整页切到原文,或双击某段只看那段的原文。

田可语 来源:田可语Tian Keyu Source: Tian Keyu

彭博社讯 前字节跳动实习生已获得包括五源资本(5Y Capital)在内的风险投资巨头支持,为其一家处于保密状态的初创公司提供资金,以在世界模型这一新兴领域向李飞飞等人工智能先驱发起挑战。

A former ByteDance Ltd. intern has secured funding from venture heavyweights including 5Y Capital for a stealth startup to take on AI pioneers such as Fei-Fei Li in the nascent realm of world models.

田可宇的研究实验室目前还没有名字或产品,但它已从包括月之暗面(Moonshot)背后的投资者五源资本和IDG资本在内的投资者那里筹集了近3000万美元,投资后其公司估值达到2亿美元。田可宇去年夏天在北京大学获得硕士学位,过去一年里,他一直致力于攻克由李飞飞和杨立昆(Yann LeCun)大力倡导的领域:模拟现实世界物理学和力学的人工智能模型,以支持从机器人技术、互动视频到自动驾驶等各种应用。

Tian Keyu’s research lab doesn’t yet have a name or a product, but it’s raised close to $30 million from investors including Moonshot-backer 5Y and IDG Capital, valuing his company at $200 million after the investment. Tian completed his master’s degree last summer at Peking University and has spent the past year tackling a field championed by Li and Yann LeCun: AI models that simulate real-world physics and mechanics to power applications ranging from robotics to interactive video and autonomous driving.

在如此初级阶段就进行押注,表明投资者中出现了一种日益增长的趋势,即发掘并投资于有前途的开发人员。李飞飞的世界实验室(World Labs)上个月在一次数十亿美元的交易中被超微半导体公司(AMD)收购,AMD首席执行官苏姿丰表示,此举是为了“获取世界级的人才”。据《金融时报》报道,大语言模型全新且更便宜的替代品Jev背后的开发商也吸引了投资者的兴趣,估值超过100亿美元。

The bet at such an embryonic stage illustrates a growing trend among investors toward identifying and investing in promising developers. Li’s World Labs was acquired by Advanced Micro Devices Inc. in a multibillion-dollar deal last month, which AMD Chief Executive Officer Lisa Su said was for “access to world-class talent.”The developer behind Jev, a new, cheaper alternative to large language models, has also attracted investor interest at a valuation of more than $10 billion, according to the Financial Times.

田可宇是一篇论文的第一作者,该论文在顶级人工智能研究会议NeurIPS 2024上被评为最佳论文之一,提出了一种更高效的图像生成方法,其运行方式类似于大语言模型,通过预测即将出现的像素集群来生成图像。

Tian was the lead author of a paper honored as one of the best at NeurIPS 2024, a premier AI research conference, proposing a more efficient image-generation approach that operates like a large language model by predicting upcoming pixel clusters.

在获得资金支持和发表论文之前,田可宇曾在字节跳动实习。字节跳动在2024年将其辞退,此前他在该公司工作了三年。字节跳动当时在网络声明中表示,已因“严重违规”解雇了一名实习生。该公司没有透露更多细节,但表示指责田可宇破坏基础模型训练的中国社交媒体传言具有误导性。

The funding and the paper follow Tian’s stint at ByteDance, which dismissed him in 2024 after three years. ByteDance said in on online statement at the time that it had terminated a former intern for “a serious violation.”It gave no further details, but said Chinese social media rumors accusing Tian of sabotaging foundation model training were misleading.

田在一次采访中表示,他曾终止另一名实习生的项目,以便为其他研究人员释放计算资源。他承认自己行为不当,但同时称自己的举动是一种“正义执行”。根据彭博社看到的裁决,在字节跳动提起的民事诉讼中,法院裁定田因违反合同需支付50万元(7.5万美元),并没收其全部工资。

Tian said in an interview that he shut down another intern’s program to free up computing resources for fellow researchers. He acknowledged wrongdoing while characterizing his actions as a vigilante act. In a civil suit filed by ByteDance, a court ordered Tian to pay 500,000 yuan ($75,000) for breaching his contract, forfeiting his entire salary, according to a ruling viewed by Bloomberg.

“我当时是在以火攻火,”田说,“如果我能带着今天所知道的一切回到过去,我肯定会采取更明智的做法。”字节跳动的发言人未回应有关田被解雇的置评请求。5Y和IDG的代表确认他们投资了田的初创公司,但未作进一步评论。

“It was me fighting fire with fire,” Tian said. “If I could go back with what I know today, I would definitely take a wiser approach.”A ByteDance spokesperson didn’t respond to requests for comment on Tian’s dismissal. Representatives for 5Y and IDG confirmed their investment in his startup, without offering further comment.

通过他的新事业,田希望利用自己获奖的论文,为人工智能创造一种新的代币化语言,用于处理视觉媒体。他表示,人类语言无法高效地捕捉这类信息。在他的初创公司中,除了他本人外,还有其他前字节跳动员工。田称,他们为人工智能构建了一个类似词典的系统,其中包含人类无法识别的20万个符号,以帮助系统处理视频。

With his new venture, Tian seeks to leverage his award-winning paper to create a new tokenized language for AI to process visual media, information he says human words cannot efficiently capture. At his 10-person startup, which includes other former ByteDance workers, Tian said they have built something akin to a dictionary for AI – with 200,000 symbols unrecognizable to humans – to help systems process video.

“你永远无法用人类语言完美地描述一段视频,”田说,“我们首先为人工智能创造一种语言,然后向其输入1亿小时的视频。”田表示,使用这种方法可以将生成一秒视频的成本至少降低一个数量级。他计划在2027年完整模型发布之前,通过现场活动展示这些功能。目前的首要任务是交付这项技术,他表示自己并未将时间花在开发消费者产品或构建商业模式上。

“You can’t ever perfectly describe a video using human language,” Tian said. “We are first creating a language for AI and then feeding it 100 million hours of video.”Using this technique would compress the cost of generating a second of video by at least an order of magnitude, Tian said. He plans to showcase these capabilities at live events ahead of a full model release in 2027. The priority is to deliver the technology and he said he’s not spending time on a consumer product or developing a business model.

田的日程安排正值全球世界模型竞赛加速推进之际。Meta Platforms Inc. 前首席人工智能科学家勒昆为其初创公司 Advanced Machine Intelligence Labs 筹集了 10 亿美元种子资金。在中国,视频生成工具 PixVerse 和 3D AI 开发商 Vast 等公司已在开发类似 Google DeepMind 的 Genie 3 的交互式世界模型,该模型允许用户实时在虚拟 3D 环境中导航。

Tian’s timetable comes as a global race in world models gathers momentum. LeCun, the former chief AI scientist at Meta Platforms Inc., raised $1 billion in seed funding for his Advanced Machine Intelligence Labs startup. In China, companies such as video generator PixVerse and 3D AI developer Vast are already working on interactive world models similar to Google DeepMind’s Genie 3, which allows users to navigate virtual 3D environments in real time.