RLMs are right. I'm building for when you can't reach one.

I run a local AI squad, and I just spent an evening with Trampoline's predict-rlm and the Recursive Language Model paper behind it. I want to start by agreeing with them, because they are right about the thing they are arguing.

Their case is the Bitter Lesson. Don't hand-engineer how a model should think. Give it room, let it run, and capability rides the model. As models improve, the thing you built improves for free, and the scaffolding you hand-tuned just becomes a ceiling. I believe this. I have said it about my own squad: don't throttle the model, let it take its turns and its tokens, enable it instead of boxing it in. On that, the RLM people and I are on the same side.

So this is not a rebuttal. It is the case their bet does not cover.

predict-rlm assumes you can call a strong model. The examples hardcode the frontier. "Performance scales with the base model" only pays you back if you can keep reaching a better base model. For a company shipping to users with an API key and a network, that is a fair assumption.

It is not a safe assumption everywhere. The places I care about most are the ones where it breaks:

• A field unit with no datacenter in reach.
• A disaster zone where the network is down and staying down.
• A craft far enough out that a round trip to a server is minutes, or impossible.
• A person who decided, on purpose, not to depend on a model someone else can ration, price, or switch off.

There you don't get to ride the best model, because the best model is not there. You get the model that fits the hardware you brought and the watts you can generate. I have been measuring exactly that for weeks: small models making real decisions on an 8-watt board, where the most decisions per watt came from a model you can actually own.

In that regime the harness is not a ceiling. It is the floor. It is the difference between a modest local model doing work you can trust and doing nothing you should. And verification matters more there, not less. A smaller model needs a harder check on its work, and that check has to be something real, a compile that passes or fails, not the model grading itself. That is not a fight with the Bitter Lesson either. A verifiable check is exactly what lets a model learn and improve on its own runs.

So: two regimes. They build for the one where you can reach the frontier, and they built a good way to do it. I build for the one where the frontier is a luxury you might not have. Both are real. I would rather be ready for the second.

When the model is there, let it fly. When it is not, the harness had better be damn good.