The article highlights the current state of reinforcement learning (RL) environments, drawing parallels with the semiconductor industry's early struggles before electronic design automation (EDA) revolutionized chip design. Just as EDA enabled scalable simulation and verification for chips, RL environments are becoming essential for the development of AI agents that can perform complex tasks. However, unlike EDA, which can prove correctness, RL environments face the challenge of defining what constitutes "success" in non-deterministic, evolving contexts. This makes verification the bottleneck in achieving reliable automation with AI. The shift in AI model requirements has also been significant. Initially, data volume was the limiting factor, but as models like GPT-3 emerged, the focus shifted to data complexity and the need for high-quality feedback over extended tasks. RL environments provide a way to simulate workflows, allowing for measurable actions and outcomes, thus transforming training into a continuous process rather than a one-time event. The article emphasizes that the competitive edge now lies in controlling the infrastructure that defines what tasks can be learned. In terms of market dynamics, the article notes that the first large-scale applications of RL environments are in coding and computer use. These areas are advantageous due to their high economic value, observable interactions, and relatively straightforward verification processes. Companies like Anthropic dominate the early market, spending tens of millions annually on RL environments, suggesting a trend similar to early adopters of autonomous vehicle data. Anthropic's contracts set the stage for significant growth in spending as RL environments transition from experimental phases to critical components of model training.