132. 对星海图创始人高继扬的3小时访谈：鲶鱼、曾国藩、Waymo与Momenta的两面、一只狼与许华哲的离开

Summary

This 张小珺Jùn｜商业访谈录 episode interviews Gao Jiyang, founder of Xinghaitu, and uses his path through SenseTime, Waymo, and Momenta to explain a pragmatic route into Embodied AI. The discussion frames robotics as a long-chain business where whole-machine work, supply chain, real-world data, models, post-training tools, distribution, and customer value have to fit together. Gao’s core argument is that Xinghaitu must build the robot body as well as the “brain” because the robot is both the commercial product and the carrier for a Physical World Data Flywheel.

Key Claims

• Gao Jiyang describes his own method as diligence, induction, decomposition, measurement, and pragmatic iteration rather than pure technical romanticism.
• He chose autonomous driving after his PhD because it was one of the few industries where AI was a bottom-level variable rather than an add-on to an already functioning business.
• Waymo is presented as an engineering training ground with strong systems, but also as an organization that lacked a sufficiently strong founder-like top-down force during rapid expansion.
• Momenta is presented as a harder, more result-oriented environment whose mass-production pressure, customer delivery, and “battle merit” culture accelerated Gao’s operator training.
• Gao says Xinghaitu had to start from whole machines and supply chain, even though the founding team came from AI and autonomous driving, because a physical robot is the data carrier and the saleable product.
• The episode treats Real Robot Data Strategy as the core moat: simulation, videos, UMI-style data, third-person data, and human-centric data all matter, but the useful mix has to be found through experiments and cost accounting.
• Xinghaitu chose Wheel-Based Dual-Arm Robots rather than humanoids because manipulation intelligence was already difficult, while wheels were sufficient for many near-term production scenes.
• Gao says Xinghaitu uses a dual-system architecture: a vision-language model layer decomposes ambiguous instructions, while Vision Language Action Models execute physical actions.
• Production Robot Scenario Selection depends on supply-side capability and demand-side tolerance: good early scenes should not require extreme speed, should have tolerable failure costs, and should scale globally.
• Gao confirms Xu Huazhe’s planned departure and says Xinghaitu supports his move toward 2C family-application entrepreneurship, while denying that the departure means algorithm innovation is unimportant.
• The episode frames Embodied AI Value Chain as the real innovation system: algorithms matter, but so do machines, supply chain, data, AI infrastructure, models, distribution, and customer value.
• Gao says the company had closed a recent round, with valuation growing from roughly 300 million RMB in January 2024 to roughly 10 billion RMB, but he frames shipment volume and real productivity scenes as the more important feedback.

Key Quotes

“以客户为中心” — Gao on the deepest operating lesson he took from Momenta.

“先做整机” — the source’s shorthand for why Xinghaitu began with the robot body and supply chain.

“链条长、周期长” — Gao’s reason that robotics does not easily allow pure technical romanticism.

“一只狼” — Gao’s chosen metaphor for Xinghaitu’s state in a hard, long-cycle robotics race.

Connections

• Gao Jiyang, Xinghaitu, Zhao Hang, Yang Zeyi, and Xu Huazhe — founder, company, founding/partner context, hardware-system partner, and departing scientist.
• Waymo, Momenta, Cao Xudong, and Tesla — autonomous-driving comparison cases behind Gao’s later robotics strategy.
• Embodied AI, Physical World Data Flywheel, Real Robot Data Strategy, Embodied AI Value Chain, Wheel-Based Dual-Arm Robots, and Production Robot Scenario Selection — core robotics strategy from the episode.
• Vision Language Action Models, World Models, and AI Plus Terminals — model and terminal context linking this source to prior AI-robotics synthesis.
• AI Commercialization Pressure, Product Led Willingness To Pay, Customer Pull, and AI Organization Design — business and organization themes surfaced by the interview.

Contradictions

• No direct contradiction with prior wiki content. The source reinforces Yin Qi’s terminal-led view of physical AI and Zhang Yi’s real-deployment view of robotics, while adding a more industrial/productivity-oriented route through Xinghaitu.
• The episode creates a productive tension with Aether AI’s causal-world-model framing: Xinghaitu emphasizes real robot data and a practical VLM plus VLA architecture, while Huang Biwei argues that pure VLA routes may be limited without causal grounding. This is a difference in route emphasis rather than an explicit contradiction.