Data Recipe Co-Creation

Data recipe co-creation is the process of data companies and model teams jointly discovering which data improves model behavior. In 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, 谢晨 compares the current robotics and post-training stage to earlier frontier-model periods where data vendors and model labs had to iterate together.

Key Claims

• A data company cannot prove value by delivering files alone if the model team cannot show that the data improves training or evaluation.
• Model teams and data teams can blame each other when a run fails, so recipe discovery requires shared metrics, experiments, and trust.
• For Embodied AI, the recipe may mix real robot trajectories, simulation, human first-person data, internet data, expert feedback, and challenge evaluations.
• Finding the right proportions may require large compute budgets, so data strategy is coupled to Frontier Model Scaling rather than separate from it.
• The process makes Data Pricing In AI less commodity-like: higher-value data is judged by learning impact, not only collection cost.

Connections

• Embodied Data Pyramid — data layers whose proportions need to be discovered.
• Data Engine Learning Loop — operating system for repeated recipe testing.
• 光轮智能 and Scale AI — source company and historical comparator.
• OpenAI, Google DeepMind, Nvidia, ByteDance, Alibaba, and Qwen — model and platform companies the source expects to participate in physical-AI data demand.