Beneath every model is data: who produces it, how it is produced, who it is sold to, and at what price. This series splits AI training data into four columns by domain: robotics and embodied AI, large language models, video generation, and AI for Science. All four are now live.
Robotics and embodied AI
A six-layer data pyramid: real-robot teleoperation, handheld UMI, human video, physics simulation, world-model generation and deployment feedback. Who produces it, who sells it, how models use it, and what an hour costs.
Large language models
A six-layer data chain: pretraining corpora, copyright licensing, post-training human data, RL environments, synthesis and distillation, and deployment feedback. Who supplies it, how models mix it, what it costs, and which legal line it runs into.
Video generation
Seven layers of data: open web video, proprietary platforms, licensed copyright libraries, synthetic rendering, cleaning and captioning, post-training preference data, and new categories. Who supplies it, how models use it, what it costs, and which legal line it runs into.
AI for Science
A nine-layer data chain: public databases, instrument capacity, wet-lab ground truth, simulation and synthesis, model recipes and validation loops, plus benchmarks, vendor policy and prices. Mainly biology.
Robotics and embodied AI
Robots have no internet-scale data source: every demonstration with action labels has to be produced by people, equipment and a site. The column organises robot data into six layers, ordered by distance from the deployment distribution and by unit cost: human video and motion capture, handheld UMI capture, physics-simulation synthesis, world-model generation, real-robot teleoperation, and deployment feedback with real-robot RL. Across 10 chapters and 188 entity cards it covers who produces each layer, who sells it, how models mix it, and what an hour costs.
Large language models
The LLM column follows a six-layer data chain: pretraining corpora, copyright licensing, post-training human data, RL environments, synthesis and distillation, and deployment feedback. It asks who supplies each layer, how models configure their mix, how much it costs, and which legal line each layer runs into. It runs to 10 chapters and 164 entity cards.
Video generation
Video models draw on seven layers of data: open web video, proprietary platforms, licensed copyright libraries, synthetic rendering, cleaning and captioning, post-training preference data, and new categories. As with LLMs, the column tracks who supplies each layer, how models use it, its price, and the legal line it runs into, across 10 chapters and 134 entity cards.
AI for Science
The AI-for-Science column follows a nine-layer chain from public databases and instrument capacity, through wet-lab ground truth and simulation and synthesis, to model recipes and validation loops, with cross-cutting chapters on benchmarks, vendors and policy, and prices. Its focus is biology. It runs to 9 chapters and 243 entity cards.
Read the AI for Science column →
How the numbers are sourced
This series is personal research. Every number is labelled in the text with its source and basis. Self-reported company data is marked as a company figure, and anything not independently verified is flagged. All judgements hold as of the date written and are not investment or R&D advice.
Adapted from the Spark Science research notes (originally published in Chinese at jimmyresearch.com). Last updated Oct 2026.