140. 对姚顺宇的4小时访谈:请允许我小疯一下!在Anthropic和Gemini训模型、技术预测、英雄主义已过去
140. 对姚顺宇的4小时访谈:请允许我小疯一下!在Anthropic和Gemini训模型、技术预测、英雄主义已过去
Podcast3 hr 50 min
Listen to Episode
Note: AI-generated summary based on third-party content. Not financial advice. Read more.
Quick Insights

Investors should prioritize Google (GOOGL) as it leverages its proprietary TPU hardware and vertical integration to secure a long-term infrastructure moat over competitors. Anthropic remains the "gold standard" for AI coding; while private, its dominance in this sector suggests Microsoft (MSFT) and GitHub are the primary public plays to capture the massive productivity gains in software engineering. The industry is shifting from simple chatbots to "agentic" models that manage entire projects, making companies focused on AI Coding Tools the most immediate source of high ROI. While NVIDIA (NVDA) maintains hardware leadership, watch for a transition toward "inference-time scaling," which favors companies that can optimize how models "think" rather than just those with the most raw compute. For exposure to consumer-facing AI innovation, monitor the Chinese AI landscape, as they are currently outperforming Western labs in user engagement and application design.

Detailed Analysis

Based on the interview with Yao Shunyu, a research scientist who has worked at both Anthropic and Google Gemini, here are the investment insights and industry trends extracted from the discussion.


Anthropic (Private)

Anthropic is highlighted as a highly efficient, "top-down" organization that has successfully challenged larger incumbents like Google and OpenAI.

  • Strategic Focus on Coding: The guest notes that Anthropic made a deliberate "bet" on coding and agentic capabilities (Claude 3.5 Sonnet/Opus), which has become their primary competitive advantage.
  • Top-Down Execution: Unlike Google’s "bottom-up" culture, Anthropic’s founders (like Jared Kaplan) are technical leaders who make centralized decisions, allowing the company to move faster and take bigger risks.
  • Product Innovation: The "Artifacts" and "Claude Code" features are cited as revolutionary UI/UX changes that have significantly increased user stickiness and developer productivity.

Takeaways

  • Market Leadership in Coding: Anthropic currently holds a "gold standard" status for software engineering AI. Investors should watch for how this translates into enterprise market share against Microsoft/GitHub.
  • Efficiency over Scale: The interview suggests that "know-how" and "clean" training methods are currently more important than raw compute power, giving smaller, focused labs a fighting chance against tech giants.

Google (NASDAQ: GOOGL)

The discussion reflects a "comeback" narrative for Google, specifically regarding the Gemini models.

  • Gemini 2.0/2.5 Progress: The guest notes that Google has "caught up" significantly. Gemini is now considered a top-tier model, particularly strong in daily use cases and reasoning.
  • Hardware Advantage (TPU): Google’s TPU (Tensor Processing Unit) is highlighted as a major structural advantage. Unlike GPUs, which are often limited by communication between individual cards, TPUs are designed as massive, integrated "pods" that offer better memory and scaling efficiency for large models.
  • Search vs. Chat: The guest suggests OpenAI actually "saved" Google by proving that chatbots are not a direct 1:1 replacement for search, but rather a complementary tool, allowing Google to integrate AI without destroying its core business.

Takeaways

  • Infrastructure Moat: Google’s vertical integration (owning the chips, the data, and the model) makes them a long-term powerhouse despite early stumbles.
  • Distribution Power: The integration of Gemini into the Android ecosystem and Chrome provides a distribution moat that startups like Anthropic or OpenAI struggle to match.

AI Coding Tools & Agents (Investment Theme)

The most immediate and "explosive" investment opportunity identified is in the AI coding sector.

  • Cursor (Private): Mentioned as a standout product that has redefined AI-assisted coding. It has moved from being a simple wrapper to a sophisticated tool that manages complex code logic.
  • Productivity Gains: The guest estimates a 20x to 50x acceleration in implementing research ideas and writing code compared to 18 months ago.
  • The "Agentic" Shift: The industry is moving from "chatbots that write snippets" to "agents that manage entire repositories."

Takeaways

  • Software Engineering Transformation: Companies that provide AI coding infrastructure (like GitHub/Microsoft, GitLab, or startups like Cursor) are at the forefront of the first real AI ROI (Return on Investment).
  • Risk to Junior Talent: The guest predicts that while AI won't replace programmers overnight, it will drastically reduce the number of people needed for a single project, concentrating power in the hands of a few "super-developers."

NVIDIA (NASDAQ: NVDA) & Hardware

While NVIDIA remains the leader, the transcript hints at the growing importance of specialized architectures.

  • GPU Limitations: The guest discusses the "bottleneck" of communication between GPUs (like the H100/H200).
  • Scaling Laws: There is a debate on whether "Scaling Laws" (just adding more chips) are hitting a wall. The guest suggests that while we haven't hit the wall yet, the focus is shifting toward "Inference-time scaling" (thinking longer) rather than just "Pre-training scaling" (bigger datasets).

Takeaways

  • Diversification of Compute: While NVIDIA is the current king, keep an eye on "custom silicon" efforts from Google (TPU) and Amazon (Trainium) as they seek to lower the massive costs of AI training.

Emerging Themes & Risks

  • Long-Horizon Tasks: A key future theme is "Long Horizon" AI—models that can plan and execute tasks over days or weeks rather than seconds. This is the next frontier for OpenAI (o1/Strawberry) and Anthropic.
  • China’s AI Landscape: The guest observes that while China may lag in "frontier" model research, they are exceptionally strong in C-end (Consumer) applications and product design, often finding "unnatural" but effective ways to engage users.
  • The "System" Risk: AI development is becoming a "system science." The risk for investors is that a single "bug" or a slight deviation in data strategy can cause a multi-billion dollar model to fail, making these companies high-risk/high-reward.

Disclaimer: This summary is for informational purposes based on the provided podcast transcript and does not constitute financial advice.

Ask about this postAnswers are grounded in this post's content.
Episode Description
姚顺宇戴了一副茶色眼镜走过来,这副眼镜会随光线变化而改变颜色。 硅谷AI业界有两位Yao Shunyu,他们曾是清华同一届毕业生,这让姚顺雨与姚顺宇时常成为话题人物。 前一位姚顺雨2025年从OpenAI跳槽到腾讯,他去年来过我们节目(第115集);后一位姚顺宇也于同年跳槽,从Anthropic来到Google DeepMind。 今天的嘉宾姚顺宇,毕业于清华和斯坦福大学,曾经的研究方向是理论物理——非厄米系统、量子物理与高能物理。他的人生奋斗姿态是——“总想挑战一些自己不太会的事”。 他人生最大的一次跨步是博士毕业,毅然决然离开深造9年的物理,来到崭新的AI行业。过去两年,他先后在Anthropic和Google DeepMind出任研究科学家,参与了Claude 3.7、4.5、Gemini 3等关键模型的开发过程。 姚顺宇一点也不nerd。有时,他会令你猝不及防,突然发表一点“小疯”言论。 第一次见面,他就对我说:“我在这个行业又没有什么导师,又没有什么旧友,我当然想喷谁喷谁。” 转行AI的两年,他变得越来越直接,越来越不害怕得罪人。访谈中,他也说了一些直白言论: “AI个人英雄主义时代已经过去了,所以也没有什么英雄,有时候甚至觉得旧时代英雄有点蠢。” “没有哪个老登是你的亲属,所以你觉得他傻,他就是傻,就可以直接说他傻。无所谓的啊。(笑)” “现在大家都是冲浪的人,本质上是那个浪,而不是你那个冲浪的人。” “AI这个事,本来也不太需要脑子——真的不太需要脑子。这个行业最重要的特质,就是靠谱,就是做事细,对自己做的事情负责任。” “你不用太担心因为自己的观点而惹到什么人。只要你的观点是自洽的,不是说随便喷人,你是有一套自己的理解。最终你在这个领域做的怎样,是有客观评价标准的——大家是会尊重你的。” 不过,在描述自己的研究时,姚顺宇又显得异常审慎。 他对于这些工作的描述是:“我自己对那个事没那么重要,更多的是,我很幸运,有机会在那个时候加入了一个重要的项目,做了一些事。” 他反复强调,AI个人英雄主义的时代已经过去了,现在都是集体主义的故事,要对神话个体的一切叙事充满警惕。 我们的节目录制于2026年3月,距离我们这次节目录制完,世界又发生了许多意想不到的变化:Meta对Manus的收购被撤销、Cursor可能被SpaceX收购、xAI将终止独立运营并入SpaceX,并更名为SpaceXAI等等。如果我们的谈话内容有一定滞后性,请大家多包涵——AI的世界实在变化太快、太出乎意料了。 可能还要特别说明的一点是,技术细节会涉及企业机密,有一些嘉宾是不方便分享的,也请大家能够包容。 在访谈中,我们尽最大可能和大家一起学习AI。你会收获姚顺宇在Anthroic和Google Gemini的技术探索历程与思考洞察。 而对于那些离经叛道的微小片刻,还请允许我们小疯一下。 接下来,就是我对顺宇的访谈。 OUTLINE: 00:02:41 两个Shunyu Yao 00:06:50 竞争与逃逸 00:26:37 “Pre-train没有到头” 00:36:23 Coding的爆发 00:51:25 字节和豆包 00:54:45 "硬蒸"和"聪明的蒸" 01:05:22 机器人 01:10:00 在Underdog之地赌一把 01:20:59 非厄米系统与量子物理 01:37:42 高能物理 01:44:24 物理与AI 01:53:47 在Anthropic训练Claude 3.7、4.5 02:36:18 "AI本质是简单的" 02:42:25 在Google DeepMind训练Gemini 3 03:02:43 "Pre-train也是一种RL" 03:08:04 技术预测 03:14:06 组织搭建 03:24:48 集体主义胜利 LINKS: 我们的播客在小宇宙、Apple Podcast、Spotify等全音频平台播出; 我们的视频播客在Bilibili、小红书、视频号、抖音等全视频平台播出; 如果你想服用文字版,请搜索我们工作室的公众号:语言即世界language is world。 本集的文字版:《独家对话姚顺宇:请允许我小疯一下》 DISCLAIMER: 本内容不作为投资建议。 CONTACT: xiaojunzhang@lisw.ai Jump into the new world-and explore with us!😉
About 张小珺Jùn|商业访谈录
张小珺Jùn|商业访谈录

张小珺Jùn|商业访谈录

By 张小珺

努力做中国最优质的科技、商业访谈。 张小珺:财经作者,写作中国商业深度报道,范围包括AI、科技巨头、风险投资和知名人物,也是播客《张小珺Jùn | 商业访谈录》制作人。 如果我的访谈能陪你走一段孤独的未知的路,也许有一天可以离目的地更近一点,我就很温暖:)