98: 李开复聊零一部分团队并入阿里：只有大厂能追逐超大模型
晚点聊 LateTalk
Deep Dive
Why did 01.AI decide to collaborate with Alibaba Cloud and establish a joint lab?
01.AI collaborated with Alibaba Cloud to establish the 'Industrial Large Model Joint Lab' to leverage Alibaba's resources for developing ultra-large models. This partnership allows 01.AI to focus on creating faster, cheaper, and smaller models for broader applications while relying on Alibaba's infrastructure for large-scale model training. The relationship between ultra-large models and smaller models is likened to 'teacher and student,' where the larger models enhance the performance of smaller ones through data annotation and synthetic data generation.
What is the significance of the 'teacher-student' model relationship in AI development?
The 'teacher-student' model relationship refers to the use of ultra-large models (teacher models) to improve the performance of smaller, more efficient models (student models). The teacher model generates synthetic data and provides annotations, which help the student models achieve better results. This approach is cost-effective and aligns with the trend of focusing on smaller, faster, and cheaper models for practical applications.
What challenges did 01.AI face in pursuing ultra-large models, and how did they address them?
01.AI faced challenges in scaling ultra-large models due to the high costs and diminishing returns of scaling laws. They decided to pivot towards smaller, faster, and cheaper models, as ultra-large models are more suited for tech giants with significant resources. By collaborating with Alibaba, 01.AI can now focus on practical applications while leveraging Alibaba's infrastructure for large-scale model training.
What is the role of synthetic data in improving smaller AI models?
Synthetic data, generated by ultra-large models, helps improve smaller AI models by providing high-quality training data. This data can be tailored to specific needs, enhancing the performance of smaller models without requiring massive real-world datasets. It also allows for more efficient training cycles, reducing costs and time while maintaining model accuracy.
What are the key factors driving 01.AI's shift towards commercialization?
01.AI's shift towards commercialization is driven by the need to balance technical advancements with financial sustainability. They aim to create models that are not only advanced but also cost-effective and fast enough for real-world applications. By focusing on industries like gaming, energy, automotive, and finance, 01.AI seeks to generate significant revenue while maintaining a competitive edge in the AI market.
What are the potential risks for AI startups in pursuing ultra-large models?
Pursuing ultra-large models poses significant risks for AI startups, including high costs, diminishing returns on scaling laws, and intense competition from tech giants. Startups may struggle to justify the massive investments required for ultra-large models, especially when smaller, more efficient models can achieve comparable results at a fraction of the cost. This makes ultra-large models a less viable option for startups compared to established tech companies.
What is the future outlook for AI-first applications in China?
AI-first applications in China are expected to grow significantly, with a focus on natural language interaction, universal reasoning, and understanding capabilities. These applications will redefine industries by integrating AI into core functionalities, creating new opportunities for startups and established companies alike. China's strengths in execution and commercialization are likely to drive the development of innovative AI-first solutions.
What lessons has Li Kaifu learned from his experience as a CEO in the AI industry?
Li Kaifu emphasizes the importance of adaptability and foresight in leading an AI startup. He highlights the need to balance technical innovation with commercial viability, making strategic adjustments based on industry trends and company performance. His experience underscores the value of leveraging partnerships, such as the collaboration with Alibaba, to achieve long-term success in a competitive and rapidly evolving field.
- 零一万物与阿里云成立产业大模型联合实验室
- 部分团队并入阿里，专注超大模型研发
- 零一万物专注于更快、更便宜、更小的模型及应用开发
- 超大模型作为Teacher Model，提升小型模型性能
Shownotes Transcript
「不再追求 AGI，他第一个讲了出来」
上周初开始，市场陆续出现有关零一万物的新调整传闻，关键词包括：“放弃预训练，资金链紧张、被阿里收购……”传闻出现的第二天，我们在零一万物办公地，中关村鼎好大厦，采访了李开复。
他解释了实际发生的变化：零一万物已经和阿里云成立了 “产业大模型联合实验室”，零一万物的大部分训练和 AI Infra 团队会加入这个实验室，成为阿里的员工，侧重超大模型研发。零一自己接下来会聚焦做更快、更便宜、更小，更能支持普惠应用的模型，同时自己做应用。
在李开复的描述里，超大模型和更快、更便宜、更小的模型的关系类似于“老师和学生”，超大模型可以通过标注结果和生成更多合成数据的方式帮更小的模型提升性能。
去年 5 月，我们也访谈过一次李开复，当时的话题是从中国“最年长的 AI 大模型创业者”开始的，在 2023 年创立零一万物时，李开复已经 62 岁。
和他行业地位还有人生阶段相似的人，更多会选择支持一个公司，而李开复这次是自己当 CEO，自己跳入了这场大模型的混战。
关于零一的新选择，有人认为是理性、务实，有人认为是收缩乃至“认输”。不管如何，在最新传闻后，李开复快速对外说明事实和阐释想法的姿态，展现了 CEO 的责任。当公司出现调整，CEO 是需要对内对外说明情况的人。
这次我又问了去年问过李开复的一个问题：功成名就时再来创业，会不会有心理包袱？
他的回答还是和上次相似。1983 年，李开复开始在卡耐基梅隆读计算机博士，当时他在研究计划里写：“AI 是人类认识并理解自己的最后一里路，我希望加入到这个全新绽放、充满前景的未来科学领域。”
李开复认为，这是他等了四十多年终于等到的 AI 时代，如果自己没有试一把，才是一个终身遗憾。
我们访谈李开复的两篇文字报道，我也贴在了 shownotes 的“相关链接”部分，感兴趣的听友可以阅读。
时间线跳转： 回应零一万物调整：“不会停止预训练，但不再追逐超大模型” 02:30 零一与阿里云成立“产业大模型联合实验室”，零一部分团队并入阿里 06:01 当开源追上闭源，任何公司没必要执着于自己预训练，但这件事现在还没发生 08:16 零一没有寻求过被收购
Scaling Law 在变慢；商业化灵魂拷问时刻已经到来 12:08 2024 年 5 月后，零一就做出抉择：聚焦更快、更便宜的模型；想做最大、最棒、最牛的模型，代价非常高，绝对不是一个初创公司可以做的事。 16:57 超大模型的作用是当 Teacher Model，提升较小模型的能力，Anthropic 和 OpenAI 都有类似实践 22:22 大模型时代，从技术竞争到商业落地拷问，一切在加快 25:13 不打打不赢的仗，不做看不到回报的大量投入 27:25 3 种 ToB 订单可以做：帮客户赚钱的；与行业客户紧密合作的；方案可复制性高的
“2025 年零一会有数亿收入”，怎么来？ 33:32 2024 年获得超 1 亿元人民币实际收入 35:03 2025 年会和适合大模型的行业公司建合资公司，结合行业数据、Know-how 和零一的技术
中国大模型创业公司会全军覆没吗？——“没有任何概率” 38:26 谈团队变化：有人禁不住诱惑，有人想追寻超大模型 40:19 中国大模型创业公司全军覆没有多大概率？——“没有任何概率”，因为 AI-first 的应用会足够颠覆 44:01 AI-first 应用的特性：自然语言交互；有通用推理、理解能力；无 AI 不成立 45:11 中国大模型创业公司在应用和落地上会有更多优势，这是互联网和移动互联网验证过的逻辑。
“等了 40 多年，不试才是遗憾” 47:28 不后悔自己当 CEO，一线创业 49:43 2025 年，应用会爆发，零一会在垂直细分行业找到有大价值的 PMF 51:19 工作是工业革命留下的魔咒，如果有了 Super Agent，我会花更多时间和爱的人在一起，这是 AI 取代不了的。 52:45 “每一位都是勇士，我们应该彼此鼓励” 53:32 2025 年的新年愿望
相关链接： 《晚点对话李开复丨他第一个讲了出来，不再追求 AGI》（25.01）) 《对话李开复：这次大模型创业，我十年都不会变现》(24.05))
登场人物： 嘉宾：李开复，零一万物创始人兼 CEO。 主播：程曼祺，《晚点 LatePost》科技组负责人。小红书 @曼祺_火柴Q) 即刻 @曼祺_火柴Q)
剪辑：甜食
☆《晚点聊 LateTalk》建立「 播客听友群」啦！☆
欢迎关注科技、商业大公司动态和创业创新的小伙伴进群交流，第一时间收听新节目。 这里有更多互动，更多话题讨论。欢迎贡献选题 & 推荐嘉宾。
请先添加「晚点」小助手的微信号，备注：“晚点聊”，我们邀请您入群。
关注公众号《晚点 LatePost》和《晚点对话》，阅读更多商业、科技文章
