The Data Goldmine

​On the brink of AGI, a strange new industry is printing money. Twelve-person teams are closing $30M quarters selling training data to frontier labs, and the product depreciates every time a model gets smarter.

​We're spending an evening pulling that world apart:

• ​Pipelines: how frontier training data actually gets manufactured: expert task generation, RL environments with rubrics and verifiers, QA through consensus grading and camera-verified provenance, dedup and decontamination before anything ships to a lab

• ​What labs buy right now: post-training expert judgment (SFT and preference data from doctors, physicists, lawyers), simulated enterprise environments for agent RL, physical capture from sensor rigs, and why plain human labeling is already dead

• ​Economics: purchase orders instead of ARR, tasks discarded once models pass them 70% of the time, 4-5x exclusivity multiples, gross versus net once the experts get paid, and who's left standing when the gold learns to mine itself

​Talks, a panel, then food and arguing.

​Hosted at AGI House.