Coding agents are getting most meaningful improvements in coding ability from RLVR now, with priors formed by ingesting open source code and manuals directly, not SO, as the basis. The former doesn't rely on resources external to the AI companies at all, and can be scaled up as much as they like, while the latter will likely continue to expand, and they don't really need more of it if it doesn't. Not to mention that curated synthetic data has been shown to be very effective at training models, so they could generate their own textbooks based on open codebases or new languages or whatever and use that. Model collapse only happens when it's exclusively, and fully un-curated, model output that's being trained on.