If the environment is synthetic, the verifier is the product

Every model release ends with the same question, which is how many parameters. Z.ai’s Jie Tang posted this week that the question cannot be answered on its own, and that parameter count only means something alongside how much data you have, where you intend to spend your compute, and who will run the model under what conditions. The roundup I read offers GLM-5.3 as the evidence: reportedly the same core base model as 5.2, with the jump coming from about a month of additional reinforcement learning.

The part that deserves more attention than the headline is what they had to build to do that.

Their writeup, as relayed, says the difficulty in scaling post-training has moved from the model to the environment. So they synthesize the environments end to end. Research agents turn patterns from real work into runnable long horizon tasks with multi step dependencies and hidden state. A judge agent attempts each task first to confirm it is solvable at all. Verifiers are written without access to the reference solution, and a verifier only counts once it survives oracle, no op, and unsolved state checks.

Read that again as a claim about what is hard now. Not the reinforcement learning. The grading.

The vault’s tuning notes say the small team version of this. Reach for reinforcement learning only when the reward signal is worth its operational burden, meaning a sandbox, tests, a simulator, graders, or outcomes you can check automatically. And build the eval data before the training data, because a training run without a held out eval is an expensive prompt experiment.

The scale is not the same. The constraint is. Whoever can grade the work can improve it, and everyone else is running experiments they have no way to score.