Scaling RL task data is tedious, and each lab is making its own tradeoffs in terms of what to prioritize. Further, multi-objective training necessitates tradeoffs at the policy level. Optimizing for SOTA on SWE-Bench implies instruction following to the point of sycophancy, which is dangerous for domains like mental health where the opposite is required. Frontier labs are focused on scaling verifiable domains like coding and tool-use, at the expense of the very critical, non-verifiable domains like medicine, law, finance and long-form writing.