Building High-Quality Medical Datasets in the Age of AI
A cross-domain talk written for the legal profession: from the paradigm shift and the changing role of data, to the 9 dimensions of data quality, dataset production for five classes of AI scenarios, the data flywheel, the Palantir ontology path, and finally the mapping to law and an action framework. 12 chapters.
- 00
Opening · Medicine Speaks, Law Listens — the Underlying Logic Is Shared
This is the opening of a cross-domain talk written for the legal profession. Medicine and law are the two professional industries most alike in the age of AI — both highly specialized, both heavily regulated, both reliant on expert judgment, both costly when AI gets it wrong.
4 min - 01
What the Age of AI Is (A 2026 View)
AI in 2026 is a different animal from two years ago. From classic deep learning before 2022, to the GPT moment of 2022, to reasoning + multimodal + agents in 2026 — each generation demands a completely different shape of data.
6 min - 02
Defining "Data" Clearly (the Concept Stack)
Before you build data, define the word \"data\" clearly — what metadata, datasets, data assets, knowledge graphs, and ontology actually are. Then three sets of easily confused concepts, so you don't govern the wrong target.
5 min - 03
The Role of Data Changed · From Fuel to Intelligence
The first generation needed labeled samples; the second needed massive corpora; the third needs reasoning traces, alignment data, and evaluation sets. Data's role has been upgraded from "training fuel" to "carrier of capability" — the core thesis of this whole talk.
5 min - 04
What High-Quality Data Is · the 4+5 Dimensions
The 4 dimensions of the national EMR grading standard (completeness, consistency, integration, timeliness) only solve “the business can run.” The age of AI needs five more — standardization & computability, representativeness, annotation quality, knowledge density, compliance & lineage.
8 min - 05
How High-Quality Data Forms · The Three-Piece Set + Three Hard Problems
Data governance, the data platform, and govern-as-you-use — the three-piece set is your entry ticket. What the age of AI adds: from governing business data to governing AI training assets; from BI platform to Data+AI platform; from passive discovery to an active closed loop. Then the three problems the three-piece set can't solve.
6 min - 06
Different Scenarios Need Different Data · The Five Classes of Fuel
Medical AI applications fall into five classes by their essential capability — knowledge apps run on knowledge bases, data apps on training sets, reasoning apps on reasoning chains, workflow apps on flywheels, embodied apps on multimodal trajectories. What each class demands of data is wildly different.
9 min - 07
How to Produce High-Quality Datasets
High-quality datasets aren't governed into existence — they're produced. Knowledge, data, reasoning, agentic, and embodied applications each demand a different production line.
7 min - 08
How Data Evolves · From Analytical Data to AI Brain
Data once helped us see the present clearly; now it trains intelligence; soon it will define the clinical and legal boundaries of AI. What matters isn't how much data you hold — it's whether you can spin a flywheel.
6 min - 09
Working Backward from Business to Data
Data governance isn’t the goal — winning the business is. In the age of AI, you build data by working backward from the business loop, the AI’s role, your data assets, and how deep to govern — not by launching a hospital-wide governance mega-project first.
8 min - 10
Mapping to Law · The Profession That Looks Most Like Medicine
Law and medicine are the two professional industries most alike in the age of AI. Eighty percent of legal AI starts by eating the knowledge base, but the real long-term goldmine is judicial reasoning, expert argumentation, and the agent-feedback flywheel.
7 min - 11
Action Framework · Start With What You Can Do Tomorrow
Building data in the age of AI isn't a technical question — it's a strategic one. The final chapter collapses the whole course into core one-liners, advice for three audiences, and an executable action checklist.
5 min