Context scaling: A new frontier of AI scaling
Abstract
A recurring failure mode in deployed AI lies at the boundary between the state a task depends on and the model-visible context available when the system acts. In such cases, what looks like weak reasoning can instead reflect failed access: relevant user state, repository state, tool traces, sensor readings, or environment changes are absent, distorted, or poorly selected at decision time. Viewing a model as a conditional Pθ (d|c) embedded in a system with three budgets (training budget Bθ, decision-time budget Bd, and context-infrastructure budget Bc), we read recent progress as a sequence of bottleneck shifts. Pretraining scaling spends Bθ to improve the base model; decision-time scaling spends Bd to search, refine, or verify candidate decisions under a supplied context; context scaling spends Bc to improve what the system can make available before a decision. The three programs interact. We treat context infrastructure as a distinct investment and measurement target, while leaving the form of any general scaling law open. Reporting only the model and per-query inference compute obscures the contribution of context infrastructure. The relevant unit of analysis is the model--infrastructure pair. Stronger models can exploit richer decision-time information states; stronger context infrastructure can acquire, represent, deliver, evaluate, and persist those states; infrastructure-mediated deployment traces can improve both. Effective-context value is a task- and intervention-dependent attribution measured after observing the utility of model-visible context. Availability, fidelity, selection, and utilization failures separate potentially relevant deployment state from information that improves a decision. The paper closes with three problem clusters: grounded evaluation and reporting, interaction-data supply and full-cost infrastructure, and safety and governance for persistent and adaptive systems.