Korva is a scroll of the internet’s best ideas.
Founders are doing the annotation themselves
Founders do the work
As models get better, the skill bar for annotation keeps rising, and a lot of small startups have answered by putting the founders on the work. Someone raises a bunch of money, goes to a lab, and asks for its business. The founders are smart people who used to be great technical employees, and they run the projects and do the annotation themselves.
VC-subsidized work comes out totally mispriced, and labs love it. Some of these startups even run as cash flow businesses, with the founders taking the profits home.
It points to where the field is heading: higher-skilled experts, with the best people in the world doing the annotation.
Where it stops scaling
The model stops working past a few data points, or past what one founder and some full-time employees can produce. Customers know it won't scale when they want ten times the throughput or ten times the number of projects, and Mercor competes against founder-led annotation directly.
How did annotation work change as models improved?
Early image datasets relied on crowds of workers with no special training. Stanford's ImageNet project, led by Fei-Fei Li, used Amazon Mechanical Turk workers in the late 2000s to label millions of photographs. Training chatbots to follow instructions took a smaller group of contractors who wrote and ranked answers, as OpenAI described in its 2022 InstructGPT paper. As models passed general tasks, labs began seeking doctors, lawyers, and engineers to judge answers.
Why can a services startup run on cash flow?
A services business pays workers per project and bills the customer for the same project, so it needs little money upfront. Staffing and consulting firms have long run this way, earning the gap between what clients pay and what workers are paid. Software companies, by contrast, spend heavily to build a product first and recover the cost across many customers.
Why doesn't founder-led work scale?
A founder's hours are fixed, so output rises only by adding people, and each added person adds coordination work. Fred Brooks made the point about software teams in his 1975 book The Mythical Man-Month, arguing that communication overhead grows as a team grows. Services firms face the same limit, because quality depends on the experienced people who set standards for everyone else. Professional services firms typically grow by building layers of managers beneath their founding partners.
More from 20VC with Harry Stebbings
