Proximal's automated coding data

The real split here is whether AI training data is made like outsourced labor, expert services, or software. Proximal is pushing furthest toward software, using automated environments and verifiers to generate long horizon coding tasks that look more like real engineering work. Scale grew from high volume human labeling workflows, while Surge sits in the middle with expert built datasets and environments that still rely heavily on managed human production.1 2 3 4 5

• Scale’s roots are in large labeling operations, first for autonomous vehicle images and sensor data, then broader enterprise and frontier model work. That model is strong when customers need audit trails and lots of human review, but it is more labor intensive than a system that can auto run tests, score outputs, and generate new coding episodes programmatically.2 3
• Surge’s coding product looks more like a premium research vendor. It sells off the shelf frontier data, evaluations, and RL environments built by expert contributors, and it runs custom projects for labs. That gives customers faster starting material than pure manual labeling, but the production engine is still centered on paid experts rather than mostly automated task generation.1 4 5
• Proximal’s advantage is that coding can be checked by machines. If a model edits code, runs tests, debugs, or completes a multi step repo task, the environment can often tell whether it worked. That makes it possible to create more data with less human touch, and to target messy domains like legacy codebases where long context and safe rewrites matter more than one shot labels.1 6

This market is heading toward data factories where humans design the rubric and software does more of the production loop. Providers closest to real developer workflows, with environments that can automatically create, run, and grade tasks, should capture more of the coding agent budget as labs move from buying static datasets to buying continuous improvement systems.1 4 6