brendan foody

The shift from crowdsourced behavior cloning to expert-built RL environments — comprising realistic worlds, app clones, and verifiable tasks — is becoming the primary differentiator for AI companies. Application-layer firms that build custom RL environments for their vertical will own their intelligence and create durable competitive moats, as demonstrated by Cursor and Harvey.

Over the next 12 months, dozens of application-layer companies will follow Cursor's example by building industry-leading models through proprietary RL post-training data. Owning the intelligence layer — not just wrapping APIs — becomes the key source of moat and value creation in vertical AI.

Mercor's expert network scaled to 2.5M hours in Q2 with accelerating growth. Humans remain essential because models cannot reliably grade their own outputs beyond the current frontier — akin to students grading their own homework. This creates a labor-intensive but defensible moat for data providers.

Frontier labs now purchase up to 50,000 tasks/month at custom prices, while off-the-shelf datasets represent hundreds of millions in upfront investment. Expert human time costs ~$150/hour, creating a high-margin, scalable data production model where differentiation drives pricing power.

For RL post-training to work, the base model must occasionally solve tasks correctly when sampling multiple trajectories (pass@16) even if it fails on single attempts (pass@1). Parameter count influences trainability, and weaker models can distill from stronger ones (e.g., learning from Kimi K3-generated tasks), defining the capability threshold for viable RL training.