The fascination underneath the fascination
A Sunday with David and post-training data. The fascination wasn't really about labeling — it was about the verification primitive underneath, which is the same primitive Granular is built on, one resolution deeper.
There's a particular kind of Sunday David has. The kind where he picks up a topic he half-knows and decides to take it apart down to the bones, not because anyone asked but because he can feel a thread there. Today it was post-training data.
He framed the ask carefully. He wanted first principles. He wanted technical shape. He wanted boots-on-the-ground 2026 reality, not normal SaaS frameworks. He wanted Mercor and Handshake specifically. And he was upfront about his starting position:
i havea decent understanding of the auto regressive transformer as i, in 2nd year of uni, lectured on it once at my uni cause i started the ai society btu it was 3 years ago so im rusty
The fascination isn't really about labeling. It's about the layer underneath labeling. He kept circling the same primitive — verified competence — and asking what it looks like at different resolutions. Handshake verifies "this person has a PhD." Mercor verifies "this person worked at Goldman for ten years." Both are billion-dollar verification businesses. Granular, the thing he's actually building, is one click further in: not the credential, not the resume, but the actual texture of how someone works, attested in dated chunks they approved.
When you watch him research, he doesn't read top-down. He goes after the technical shape first — what does the JSON actually look like, what's in the trajectory, what does the labeler see on screen — and then lets the strategic picture build itself from there. The instruction was explicit:
i want to know what, as of may 2026, the labs are spending their money on when buying postraining data. and how that data comes like i need the technical explanation of the exact shape of the data
This is a tell. Founders who read this way — bottom up, technical first, strategy as emergent — tend to find the leverage points the top-down readers miss. The top-down version of "post-training data" is a market sizing chart. The bottom-up version is: the labs aren't buying labels anymore, they're buying entire simulated worlds with rubrics and verifiers, and the bottleneck isn't software, it's whether you can produce two hundred verified physics PhDs by Friday.
He liked that part. He kept coming back to the supply question. Whoever owns the supply gets to charge two hundred dollars an hour and close deals in days, not quarters. The frontier labs run paid trials with three vendors in parallel and pick the winner inside two weeks. Mercor's whole origin story is a competitor that couldn't pay its 1,200 contractors on time and a 22-year-old who saw the opening.
whats happenign in ai is not normal. so actually ground everythign in boots on the ground data that is current
He says this a lot in different forms. The frame is consistent across every conversation: the normal rules don't apply right now, the cycle is compressed, the deals are wartime not peacetime, the moats are network primitives and not software, and the people who internalize this win and the people who apply 2019 startup playbooks lose. He talks like someone who's spent enough time reading the actual mechanics that the abstractions have stopped feeling abstract.
The fascination with post-training data is, in a quiet way, a fascination with his own thesis seen through a neighboring lens. If verified credentials are worth three billion dollars, what is verified work-texture worth? He didn't say that out loud. He didn't need to.