Mercor Head of Product on Revenue Concentration from Frontier Labs

20 insights saved from this video by @business

@business· Business Models
As models improve, annotation demands higher-skilled experts, increasing the value of technical talent producing training data.
@business· Business Models
Apex benchmarks put top models near 50% on long-horizon workflows, far below the frequently cited 90% automation figure.
@business· Business Models
Some startups turn founder-led annotation into venture-capital-subsidized labor, giving labs skilled work below its true cost.
@business· Business Models
A maximally flexible annotation platform became chaotic after supporting hundreds of heterogeneous workflows and every customer request.
@business· Business Models
Evals do double duty: they specify desired model behavior and become the optimization objective for improving it.
@business· Business Models
A training set can increase revenue when it improves performance against an evaluation set modeling customer demands.
@business· Business Models
Reinforcement-learning environments are the fastest-growing data type because they simulate the tools, files, and tasks agents face.
@business· Business Models
Agent training data is moving toward deployment-like environments, not abstract examples stripped of the tools and tasks agents use.
@business· Business Models
Specialized models will emerge for individual companies because growth, margin, and work-life balance are different objectives.
@business· Business Models
As better models create more economic value, demand persists for evaluation sets, training sets, and human-generated data.
@business· Business Models
AI-generated code is accelerating software output and expanding cyber risk as offensive and defensive capabilities continuously adapt.
@business· Business Models
Healthy cash flow beats outsider-selected metrics: ending every week with more money in the bank signals business health.
@business· Business Models
A previously difficult-to-scale environment market became viable through persistent experimentation once strong demand justified continuing.
@business· Business Models
To forecast AI demand, stay close to leaders across diverse labs and keep validating hypotheses. Leadership judgment remains partly guesswork.
@business· Business Models
AI compresses implementation details and removes skill bottlenecks, making prioritization and business-impact judgment central PM capabilities.
@business· Business Models
The annotation and evaluation platform lets experts support training and evaluation datasets while customers access data and run evaluations.
@business· Business Models
Every company-specific model needs enterprise-specific evaluation and training data to prove how it performs in context.
@business· Business Models
Companies outsource generic HR and procurement sooner than workflows central to their legal, advisory, or operational advantage.
@business· Business Models
AI spending makes sense when unit economics work, not when customer-service token costs exceed generated revenue.
@business· Business Models
Open-source models raise customer expectations without necessarily cannibalizing demand for frontier evaluation and training data.

More from 20VC with Harry Stebbings

See everything from 20VC with Harry Stebbings

Post on X

Keep what you found in Mercor Head of Product on Revenue Concentration from Frontie…

A free account saves any of these cards to your Collection, and Korva resurfaces them so you actually remember.