The rapid advancement of large language models (LLMs) was built on a foundation of 'extraordinary luck.' For the first generation of models, the primary requirement was a vast corpus of text, which humanity had already been diligently digitizing for decades. Projects like Wikipedia, Stack Exchange, and the Internet Archive provided a massive, self-conscious 'data commons.' Even the raw, messy data of the web was captured by initiatives like Common Crawl, which archived petabytes of information. This windfall of static documents allowed models like GPT-3 and Llama to be pretrained on a scale that was previously unimaginable.
However, the frontier of AI research has shifted toward 'agentic' AI—models designed to perform complex tasks rather than just generate text. This shift introduces a fundamental change in data requirements: while pretraining consumes documents, posttraining consumes tasks. Human activities are typically situated, adaptive, and ephemeral; once a person performs a task in the real world, that specific moment is gone. To train an agent via reinforcement learning (RL), these fleeting actions must be 'frozen' into durable objects consisting of a world, an intent, a set of tools, and a grader to evaluate success. Unlike the web's archives of books and articles, there is no pre-existing public infrastructure for recording the ephemeral workflows of lawyers, doctors, or scientists.
Software engineering serves as the only significant exception to this scarcity. Because software is built on deterministic principles, it has naturally evolved into a 'task commons' through version control, continuous integration, and unit testing. Yet, even this 'easy case' reveals the difficulty of the transition. Efforts to convert open-source repositories into executable RL tasks have seen high failure rates; for instance, in some subsets of the TaskTrove dataset, over 90% of Java tasks were discarded because they were not executable. This suggests that turning human activity into machine-learning-ready data is a process of 'manufacturing' rather than simple curation. Without a deliberate, systematic effort to inscribe human work into reproducible environments, the ability to develop frontier AI may become a private asset held only by firms with the capital to build these environments from scratch, potentially limiting the influence of independent researchers and non-profits in shaping the technology's future.