134. 【数据的综述】和谢晨聊，新时代的石油、历史、版图、数据金字塔、定价与Recipe

Summary

This 张小珺Jùn｜商业访谈录 episode interviews 谢晨, founder and CEO of 光轮智能, about data as the third pillar of AI beside compute and algorithms. The discussion reframes data from labeled files into Data As Education: task design, expert feedback, failure correction, evaluation, and environments that let models learn. For Embodied AI, Xie argues that real robot data is necessary but too costly to scale alone, so useful progress depends on an Embodied Data Pyramid centered on Robotics Simulation Evaluation, Data Engine Learning Loop, and Data Recipe Co-Creation.

Key Claims

• 谢晨’s path from physics and quantitative finance into dynamic pricing, Cruise, Nvidia, and autonomous-driving simulation shapes his view that simulation can be a real training and evaluation tool rather than only a demo environment.
• Data is increasingly like education: early datasets such as ImageNet looked like static textbooks, Scale AI industrialized annotation as a factory, and frontier post-training/evaluation now depends on experts who can set tasks, grade answers, and supply feedback.
• In the current large-model stage, pretraining text is less the only bottleneck; post-training, evaluation, and difficult feedback data become more important for large language models and digital agents.
• For Embodied AI, real teleoperation or robot-body data is accurate but expensive and hard to scale, so it should be the smallest and most valuable layer rather than the whole strategy.
• The episode’s Embodied Data Pyramid puts real robot data at the top, simulation data in the middle, and internet plus human first-person data at the bottom; Xie argues the layers should form a loop rather than remain isolated.
• Robotics Simulation Evaluation is framed as a prerequisite for general-purpose robots because real homes, factories, and tasks cannot be repeated at enough scale for systematic training and measurement.
• Good robotics data can include failure, correction, and diverse solution paths; “perfect” trajectories are not always the highest-value examples once models can learn from recovery.
• World Models and Vision Language Action Models are treated as complementary: world models can supply physical prediction and cloud-brain context, while VLA models translate perception and instructions into action.
• Xie contrasts robot data with Tesla’s Data Engine: robots do not yet have millions of deployed units producing cheap real-world feedback, so a simulation-centered loop has to substitute for the missing fleet-scale shadow mode.
• Data companies may move from Data Factory to Data Engine Learning Loop, selling environments, evaluation, feedback, and recipe discovery rather than standardized labeled files.
• Data Recipe Co-Creation matters because data companies and model teams have to discover which data actually improves training; recipe quality cannot be separated from model architecture, compute, and evaluation.
• Data Pricing In AI rises with customization and feedback value: pretraining-like data is more standardized, while post-training, evaluation, expert feedback, and embodied trajectories can command much higher prices.
• The embodied-AI industry is likely to split into model-brain companies, robot-body companies, data/simulation companies, and scenario owners rather than being monopolized by a single full-stack player.

Key Quotes

“数据约等于教育” — Xie’s core metaphor for moving beyond static labeled files.

“没有仿真这件事做不成” — Xie on why simulation is necessary for robotics.

“Data Engine” — Xie’s preferred frame for 光轮智能, contrasted with a Data Factory.

Connections

• 谢晨, 光轮智能, Cruise, Nvidia, and Scale AI — person, company, autonomous-driving/simulation context, infrastructure context, and data-industry comparator.
• ImageNet, Data As Education, Data Engine Learning Loop, Data Recipe Co-Creation, and Data Pricing In AI — source’s data-industry history and business-model frame.
• Embodied AI, Embodied Data Pyramid, Robotics Simulation Evaluation, Real Robot Data Strategy, and Physical World Data Flywheel — robotics data strategy and its tension with real-world fleet data.
• World Models, Vision Language Action Models, World Action Models, and Video Models — technical model routes around physical prediction, action, and simulation.
• OpenAI, Google DeepMind, Nvidia, ByteDance, Alibaba, and Qwen — model-company and platform-company players Xie expects to matter in physical AI.
• Unitree Robotics and Zhiyuan Robotics — robot-body commercialization examples used to separate body companies from brain/data companies.
• Xinghaitu and Aether AI — prior wiki robotics sources that make the same field’s route-level disagreements visible.

Contradictions

• No direct contradiction with prior wiki content.
• The source creates a productive tension with Real Robot Data Strategy and Physical World Data Flywheel from the Xinghaitu source. Gao Jiyang emphasizes the robot body as the real data carrier, while 谢晨 argues that real robot data is overestimated if it is treated as scalable by itself; the difference is a weighting dispute inside a shared embodied-AI data problem.
• The source also qualifies World Models and Video Models optimism: Xie sees world models as possibly becoming a class of simulation, but says ordinary video generation is not yet simulation unless it supports reproducible action, physical accuracy, and counterfactual control.