173: 对话姚颂：深鉴、东方空间、再出发，「天才少年」十年后

Summary

This LateTalk interview uses 姚颂’s three hard-tech startups, 深鉴科技, 东方空间, and Striding AI / 正行创新, to trace a decade of Chinese hard-tech entrepreneurship. The episode connects 清华大学 research culture, AI-chip fundraising, strategic acquisition by 赛灵思, commercial-space launch strategy, and physical AI route choice. Its main conceptual addition is that physical intelligence is a system stack, not a single model, robot body, or demo.

Key Claims

• Yao Song / 姚颂 frames his startup choices around two lines: important human problems such as survival space, health and happiness, efficiency and fairness; and personal achievement from turning a technical origin into products used by many people.
• DeePhi Tech / 深鉴科技 began from Tsinghua University / 清华大学 AI-chip and acceleration work by Wang Yu / 汪玉, Han Song / 韩松, Yao, and classmates, then had to translate technical promise into Investor Risk Narrative and market-product language before financing.
• The DeePhi Tech / 深鉴科技 sale to Xilinx / 赛灵思 was not presented as only a price decision; Yao says the strategic fit, IPO uncertainty, company risk, and team intensity all mattered, making it a case for Strategic Acquirer Fit and Post-Acquisition Founder Identity.
• Yao’s post-exit “gray” period and later role at 经纬中国 show Founder Motivation Evolution after a successful acquisition: the exit removed the first mission before the next direction was clear.
• Orienspace / 东方空间 was chosen from a China-specific commercial-space thesis: satellite internet and future communications demand would need launch capacity, and a rocket company should prioritize high success rate, enough payload capacity, and market timing before cost.
• 引力一号 made Orienspace / 东方空间 publicly visible through a sea launch, unusual solid-rocket configuration, and public launch livestream, extending Space Economy Infrastructure beyond SpaceX-centered cases.
• Striding AI / 正行创新 defines physical intelligence as data, compute, models, algorithms, software, hardware, scenarios, solutions, and remote systems working together, which makes Embodied AI Value Chain and vertical integration central rather than optional.
• Yao argues that VLA work was approaching a bottleneck by late 2025 and that some companies were shifting toward reinforcement learning, World Models, and World Action Models.
• The company starts from retail and 3C manufacturing scenes through CP Group / 正大集团 and an ODM partner, using scenario access as the seed for a physical-world data loop and Milestone Commercialization.
• Yao favors Wheel-Based Dual-Arm Robots for early delivery because bipedal humanoids and dexterous hands add cost, instability, and control difficulty before enough commercial value is proven.
• The source makes Robot Demo Authenticity explicit: Yao says all demo videos should state speed倍率 and whether the behavior is fully autonomous, because teleoperation presented as autonomy distorts the market.
• His risk posture is aggressive but bounded: he is willing to go all in on time, energy, and attention, but does not endorse selling homes or removing all fallback options.

Key Quotes

“boring 和 suffering” — Yao’s borrowed contrast for why a comfortable but slow path felt worse than startup suffering.

“好多快省” — his ordering for commercial rockets: reliability, payload capacity, speed to market, then cost.

“八分进攻、两分防守” — his description of continuous-founder risk posture.

“说出来的一定是真话” — his bottom line for demo and financing communication.

Connections

• Yao Song / 姚颂, Wang Yu / 汪玉, Han Song / 韩松, and Tsinghua University / 清华大学 — founder, mentor, collaborator, and institution behind the first startup path.
• DeePhi Tech / 深鉴科技, Xilinx / 赛灵思, Hard Tech Fundraising, Investor Risk Narrative, and Strategic Acquirer Fit — AI-chip company and acquisition path.
• Orienspace / 东方空间, Gravity-1 Rocket / 引力一号, Commercial Rocket Launch Priority, Space Economy Infrastructure, Satellite Connectivity Competition, and Space Mission Livestreaming — commercial-space branch.
• Striding AI / 正行创新, Physical Intelligence System Stack, Embodied AI, Physical AI, Embodied AI Value Chain, and Milestone Commercialization — physical-intelligence branch.
• Vision Language Action Models, World Models, World Action Models, Robot Reinforcement Learning, and Robot Form-Factor Pragmatism — model and robot-body route judgments.
• CP Group / 正大集团, Xie Guomin / 谢国民, Physical World Data Flywheel, and Production Robot Scenario Selection — scenario access and commercialization loop.
• Founder Motivation Evolution, Founder Risk Calibration, Post-Acquisition Founder Identity, and Hard-Tech Direction Selection — founder psychology and route-selection layer.

Contradictions

• No direct contradiction found.
• Productive tension to track: Poke Robotics and Physical AGI emphasize the household-general robot brain as the largest prize, while Striding AI / 正行创新 in this source starts from business scenarios, full-stack delivery, and milestone-level commercial value.
• Productive tension to track: Xinghaitu and Production Robot Scenario Selection already emphasize industrial-scene grounding; this source agrees on scene grounding but frames physical intelligence more explicitly as a vertically integrated system that must shorten the route from frontier research to paid deployment.