AI's Next Frontier: Expertise is the New Bottleneck Resource
By Klaus Berger ·
As compute power commoditizes, high-skilled human judgment and structured data curation are defining the next wave of AI value.
The podcast “Sequoia Capital” episode on RL Environments Explained offered a dense, almost overwhelming look at the mechanics of advanced AI training. What struck me immediately was the explicit declaration that the data market has transitioned entirely away from the "low-skilled crowdsourcing era of behavior cloning" toward what Brendan called the "agentic era of data." This shift suggests that the value chain is rapidly maturing past mere quantity and into highly specialized, structured quality—a transition with profound implications for where capital will flow in the next decade.
Throughout the discussion, key claims emerged regarding this new industrial architecture. Brendan outlined an RL environment as having three core components: "Worlds" (real-world documents), "Apps" (high-fidelity clones of popular software like Salesforce), and "Tasks" (prompts with verifiers). He stressed that humans remain indispensable because models struggle to reliably identify their own mistakes in complex domains, making human-created rubrics essential. Furthermore, he argued that the primary barrier to scaling is covering the full distribution of these worlds, apps, and tasks across the entire economy, requiring massive scale-out efforts guided by expert input.
The Return of Human Capital as a Bottleneck Resource
The most telling aspect of this discussion was not the technology itself, but the economics underpinning its development. Brendan detailed three methods for curating high-quality datasets: paying per complex task, using off-the-shelf data, and "Expert Provisioning." While he noted that Expert Provisioning is becoming less central, the entire model still relies on finding and utilizing "the highest-skilled experts in the world that can work collaboratively in teams"—lawyers, doctors, bankers—to build these frontier evaluations.
This confirms a return to scarcity value. If compute power becomes commoditized, the truly differentiating factor remains data, but critically, this data requires expert scaffolding to achieve "Realism" and "Accuracy of Verifiers." The ability to architect an environment that reflects the real-world distribution of tasks is not merely technical; it demands deep domain knowledge—the kind of specialized human capital once thought to be obsolete in the age of automation.
Pricing Value from the End Result, Not the Effort
Brendan offered a surprisingly clear lesson on market pricing: the most natural way to price data is by working backward from the customer's goal—determining how much a specific outcome (e.g., achieving frontier status on a leaderboard) is worth to them. This fundamentally shifts the business model away from selling "data points" or "compute hours." Instead, it positions data providers as performance guarantors, selling access to an outcome.
This sophisticated pricing strategy suggests that the market recognizes the extreme difficulty of creating reliable, high-fidelity training environments. It is not enough to simply gather documents; one must prove that the environment accurately reflects complex human workflows and can be tested against rigorous rubrics. The focus on "trajectory analysis"—where both agentic quality control systems and human review validate scores—is a clear indication that systemic risk mitigation (i.e., ensuring accuracy) has become the premium service.
Systemic Bottlenecks and Structural Investment
What matters for those of us focused on macro finance is understanding the structural bottleneck. The current consensus, as presented, is that while AI tools can make task creation more efficient, the process remains "highly human-intensive." Models cannot reliably write detailed rubric criteria for mistakes; this noise level is unworkable for training purposes.
This reinforces a key insight: the primary investment is not in the algorithms (which are rapidly advancing), but in the infrastructure of knowledge—the architecture that maps complex, siloed customer processes into structured, measurable environments. The significant reliance on custom, siloed data owned by critical customers to maintain a competitive advantage confirms that the value accrues disproportionately to those who own and can structure proprietary institutional knowledge.
The era of general-purpose computational power is giving way to an age where specialized human expertise—the ability to map complexity into measurable environments—is the ultimate source of durable economic advantage, making high-skilled labor a more valuable asset class than previously imagined.