ทวีตที่ปักหมุด
Lot of new faces here this week, so a quick recap of who I am and what we actually do.
I'm Pratik. We build human data infrastructure for frontier AI labs at Anthromind, the training and eval data the best models in the world are made of.
The most surprising thing I've learned doing this is that the labs can't get enough good data, no matter what they're willing to pay.
@seanZCai (the data guy) puts it plainly - he has never once watched a lab turn down genuinely good data over budget.
The money is sitting in a pile, waiting. Budget was never the constraint.
So why can't the best funded companies on earth get the data they actually need?
Because great data was never a scale problem.
They have the capital, the researchers, the compute. None of it produces a great dataset on its own.
Quality comes from somewhere money can't reach. Iteration, feedback over months, and an almost unreasonable amount of rigor on every single item.
Especially on the hard sets, where there's no answer key and the model is already failing most of them.
And rigor is the first thing scale usually kills.
A frontier lab is a machine built to scale, so the one input that makes data great is the exact thing the machine can't produce.
The biggest data vendors hit the same wall from the other side. They grew into marketplaces of a million contractors, and the founder-level care behind their early work didn't survive the headcount.
We're building Anthromind to scale the thing everyone else gives up. The rigor itself.
Right now that still means us, the founders, are in the data, hands on every hard call. I can't tell you how many passes one item takes before it ships, because at some point we stopped counting.
The whole bet is that this obsession lives in the infrastructure we're building, so it gets sharper as we grow instead of breaking.
Everyone else is racing to spend more and label faster. We're building the thing that makes great data repeatable.
That gap in the market is wide open right now.
English
