Data Pricing In AI

从蒸馏到合成数据到 RSI，模型竞争的下一个焦点是什么？｜对谈 Evolvent AI 联创孟繁青 adds concrete pricing intuition from 孟繁青. He says long-horizon Synthetic Agent Data tasks can be expensive because the value is not the row itself; it comes from environment setup, trajectory generation, verification, and whether the resulting data measurably improves a model.

Data pricing in AI is the episode’s frame for why different kinds of data carry sharply different value. 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 argues that standardized pretraining-like data is cheaper, while feedback-rich, expert, customized, and embodied data can be much more expensive.

Key Claims

• Static or pretraining-style data tends to behave more like a commodity.
• Post-training and evaluation data is more customized because it depends on the model’s weaknesses, tasks, and desired behaviors.
• Embodied data can be priced by physical diversity, trajectory quality, labels, evaluation criteria, expert feedback, and whether failure-recovery sequences are included.
• Better pricing logic depends on Data Recipe Co-Creation: customers pay for data that measurably improves model capability, not just for hours or rows.
• RSI Data may price higher than ordinary agent traces because it includes training-loop expertise, long execution, and proof that one model or data process improved another.

Connections

• Synthetic Agent Data, RSI Data, and Environment-Based Agent Benchmarks — Evolvent AI source branch around long-horizon data value.
• Data As Education — why feedback and task quality can be more valuable than static examples.
• Embodied Data Pyramid — source of different embodied data layers and cost structures.
• Data Engine Learning Loop — system that can prove data value through repeated evaluation.
• AI Commercialization Pressure — broader pressure to turn expensive AI training and data into paid outcomes.