From what I understand the world is not shaped in a way that makes "scaling latent reasoning unto infinity while having gotten rid of visible CoTs entirely" appealing or likely.

Which is to say, spending your marginal flops on recurrent depth seems to be Pareto-dominated by all of: more pretraining tokens, more RL, and more params—unless you find yourself faced with some rather extreme constraints. Like "I'm out of VRAM, but my MFU is low, I can't get my hands on any more pretraining data, I've already run several epochs on the data I do have, I'm hit severely diminishing returns on the RLVR envs available to me, and making useful new envs is becoming slow and expensive."

Increasing recurrent depth is a way of increasing flops-per-param. That makes the model more data-efficient and more expensive/slow without requiring more memory. The increase in data efficiency, combined with the diminishing returns of scaling along this axis (as with every other), means that once I've added C flops of recurrent depth, spending the next C flops on RL for CoT becomes more, not less, attractive.

And I was never going to spend the next C flops on more recurrent depth anyway, because my MFU isn't low any more.