The Cheap Half of the Loop

Coding agents pulled away from everything else in 2025. The easy read is that code is simpler for a model. The more useful read is that the field found a reward it could check for free, and built an industry on it.

The half that grades itself

The training move that defined last year was reinforcement learning on tasks a machine can check. It pulled in compute that had been earmarked for pretraining. Code and math went first, because the feedback is cheap. A compiler returns yes or no. A test passes or it does not. There is a clean principle underneath this: the easier a task is to verify, the easier it is to train a model to do it. Capability spikes wherever an automatic grader exists, and stays jagged everywhere else.

So the boom is not really about code. It is about checkability. Code just happens to be the most checkable thing we do, which is why a signal you can grade for free is the one that pulled in the year's compute. Labs are now racing to manufacture more of these checkable environments, an effort reported to run into the billion-dollar range, because the scarce input is not compute or data. It is a task with a clean answer.

One caution, because the causal story runs ahead of the evidence. The verifiable reward looks like it sharpens what the base model already had more than it builds something new. Push the sampling far enough and the untrained model often reaches the same answers, and on some models the measured gain even survives when the reward is handed out at random. So the honest claim is narrower than "checking makes models smart." It is that checkable feedback is the cheapest place to push, so the field pushed there. Call it the cheap half.

The other half, and what training does to it

Everything so far is one move in a larger loop. Call the loop ground and diverge. Grounding is checkable contact with something that can say yes or no, and the cheap half is exactly that move, mechanized and poured full of compute. Divergence is the other move: opening the space of candidate answers wide enough that something non-obvious is in it. I wrote about the loop in Past the Obvious: get past the first answer, then let reality correct you. This is not that piece. This one is about what the year's training economics did to each move.

They moved in opposite directions. The cheap half got industrialized. The scarce half got trained out by the same process.

The methods that produced the gains narrow the model's range as a side effect. Preference tuning makes outputs more uniform. In reasoning training the policy entropy collapses, and the model's sampled answers converge toward one path. Left to a plain prompt, it commits early. So this is not two pillars rising together. It is one half getting louder while the other gets quietly suppressed by the method paying for the first.

Two kinds of divergence, and why one lasts

It matters to split divergence in two, because they are not the same problem and they are not on the same clock.

The first kind is the one you can score by the result: generate, deliberate, sample more, keep what works. It does not matter whether the model gets there by trying many options or by thinking longer down one, because the answer at the end is checkable either way. So this kind is getting solved. It is most of what more compute at inference time buys you, and the labs are closing it fast. If your edge is out-sampling the model, that edge is renting time.

The second kind is frame origination: not more answers inside a frame, but seeing the frame is wrong and holding a different one. The Frame Trap covered the mechanism, that a model works within whatever frame you set and not on the frame itself. What that piece did not say is why the limit lasts.

Two routes can teach a model a skill: imitation, and a reward signal. Imitation has already had its turn here. A model has read more reframing than any person ever will, every pivot and reversal in the written record, and it still waits for you to supply the new frame. So the route that is left is a reward, and a reward needs a grader. A frame's correctness is the hardest thing there is to grade. You only learn whether a frame was right long afterward, from what it led to, with nothing clean to score it against in the moment. By the same logic that solved the most checkable work first, the least checkable work gets solved last. Framing is not off the roadmap. It is at the end of it.

The handle, with an expiry date

This is why a build-your-own-divergence loop is worth it now, and how you will know when it stops being. Today an external loop that forces real divergence beats the default, because the default converges too early. Part of that loop is the checkable kind of divergence, and that part is turning into a feature you get for free. The other part is choosing the frame the model then works inside, and that part is last in line precisely because no one can yet grade it.

So the edge is real and it is dated. Out-sampling is rented and expiring soon. Framing is rented too, but the lease runs longest, because it is the hardest thing to score.

Close

This is not a theory of AGI, and the two-half framing is a working read rather than a settled result. What it is, concretely, is two bottlenecks moving in opposite directions on one clock, and the clock is checkability. The more checkable a task, the sooner it gets cheap. Grounding got cheap first. The scoreable kind of divergence is getting cheap now. Choosing the frame gets cheap last, because there is nothing to grade it against until the outcome is already in. So measure your edge against the next model, not this one. Whatever a model twice as good would simply do for you is rented, and the lease is short. What still needs you when the model doubles is the framing. That is the edge worth building on: being the source of the frames a fast, well-grounded model is, for a while yet, the last to originate for itself.