Proximal's Research Infrastructure Bet

This points to a shift where the scarce input is no longer raw human labor, but systems that can generate and score tasks at or beyond expert level. Proximal is built around that bet. It uses automated rollouts, execution environments, and verifiers to turn code, traces, and workflow artifacts into harder training tasks, while Scale and Surge still lean more heavily on expert designed tasks and human judgment loops.1 2 3 4

• Proximal has made this concrete in coding. FrontierSWE gives agents up to 20 hours on tasks like compiler optimization, ML research, and system reimplementation, and its public benchmark uses automated verification, which reduces dependence on a human reviewer deciding whether a result is good enough.1 5 6
• Scale is moving the market in the same direction, but from a more human centered starting point. Its RL Environments product wraps simulated business workflows with expert designed tasks and verifiers, and Scale says nearly half of new training projects now involve these environments.1 3 7
• The competitive split is becoming expert supply versus research infrastructure. Surge highlights expert contributors, human written rubrics, and benchmark gains from its datasets, while Proximal explicitly positions itself as a research and engineering lab rather than a talent marketplace.1 4 8

If frontier models keep outrunning what even strong human evaluators can reliably judge, data vendors will look more like infrastructure companies than labor networks. That favors teams that can build durable environments, hard to game reward signals, and automated quality checks across coding, science, and other technical workflows.1 2 3 4