← back

State of Data — Sean Cai, Independent / State of Data

6.0K views · Jul 26, 2026 · 18:22 min · Watch on YouTube ↗
Takeaway

Authentic workflow data and credible verification are central bottlenecks in extending AI into expert work.

Summary

  • Sean Cai argues that AI data supply chains are unbundling into specialists for expert sourcing, environments, reward design, and evaluation.
  • He distinguishes static outputs from process data capturing professional trajectories and decisions.
  • His type-one/type-two distinction separates real workflow capture from contrived expert examples, favoring ongoing partnerships with operating businesses.
  • Task verifiability depends on decomposability, agreement about correctness, and availability of fresh verified examples; these properties help explain coding’s early progress.
training-datadata-marketsverification
Original description
GPT 5.5 and Opus 4.8 landed within three points on the same finance benchmark while failing in opposite directions. GPT got the arithmetic right. Opus got the methodology right. One leaderboard number flattened both failures into a single noisy sample, exactly what happens when benchmarks reward one scaffold and vendors sell the data used to climb the tests they designed.

The scarce asset is no longer another isolated answer. It is process data: the reasoning trace, sequence of decisions, state changes, failures, recoveries, and verified outcomes that turn general competence into real expertise. Static datasets depreciate as models improve, so the durable moat is a live pipeline into real work plus the infrastructure to retrain when the base model changes.

Speaker info:
https://x.com/SeanZCai
  / sean-z-cai  
https://www.seancai.com/philosophy/st...

Timestamps:
0:00 - The data market nobody sees
1:15 - Data as industrial fuel
2:31 - Type one and type two data
3:23 - Compute, data, and talent
4:24 - State data versus process data
5:54 - The three axes of verifiability
8:14 - When benchmarks become snake oil
10:08 - Three finance benchmark tests
12:04 - Predicting the next AI domain
13:21 - The robotics counterexample
14:22 - Where the economic value lives
16:02 - Why data companies move enterprise
17:09 - The durable moat