Jaggedness: what RL optimises, and what stalls

Karpathy's argument that models are advancing only where a reward can be computed, with a joke that hasn't changed in three or four years as the demonstration — and why that is the same boundary the harness cluster keeps running into from the other side.

From the NoPriors interview. The observation first: working with a current agent feels like talking to a brilliant systems programmer and a ten-year-old at the same time. His point is not that it is uneven — people are uneven — but that the combination is one you would never meet in a person. Human ability is far more correlated across domains than this.

The joke test

His demonstration is deliberately petty. Ask a frontier model for a joke and you get the one about scientists not trusting atoms, because atoms make everything up. It is the joke you would have got three or four years ago. Meanwhile the same model will work for hours on an agentic task and move mountains.

The joke has not improved because nothing is optimising it. Coding, maths and anything else with a checkable answer sit inside the reinforcement-learning loop; humour sits outside it, so it is frozen at whatever the pre-training distribution produced.

Asked directly whether competence at code generalises to everything else, he says no — maybe a little, "not a satisfying amount". This is worth recording plainly because the opposite is a common background assumption: that capability bought in verifiable domains arrives free everywhere else as a side effect. His answer is that it does not, and that the gains are clustered inside opaque models where you cannot see which parts got the attention.

Where the ridge falls

The formulation he uses: you are either on rails, inside the verifiable domains, where everything moves at the speed of light — or you are off them, and things meander.

What he names as off-rails and consequently weak:

Both interviewers volunteer the same complaint from the user side: the frustration is not that the agent fails, it is that it burns a lot of compute on something it should have recognised as an obvious dead end.

Why this is the same wall the harness cluster keeps hitting

Karpathy's constraint on Auto Research: taking the researcher out of the loop is that if you cannot evaluate it you cannot automate it. Jaggedness is that same line drawn from the training side rather than the harness side: the reason your loop only works on verifiable tasks is that a verifiable reward is also the only thing that made the model good at them.

That makes several scattered findings in this notebook one finding:

The practical reading: when you are designing a loop, the question "can this be checked mechanically?" is not a nice-to-have constraint you might engineer around later. It is a prediction of whether the model is any good at the task at all.

The same ridge, seen from the buyer's chair

Four months later, DoorDash describes the pattern without naming it, from the position of a company paying for the models rather than reasoning about them. On non-coding work — accounting, analytics — the sequence they report is: the team says "yeah, it works okay"; the data gets scrubbed and handed to a lab, an RL environment is built around the task, and the models crush it; the same thing pointed at real enterprise data does not hold.

Step two is the verifiability operation. Building an RL environment around a task is dragging it onto the rails, and the gap between step two and step three measures how much of the real task the rails did not cover. Their own open question — harness gap, or something missing from the models' data distribution — is Karpathy's unresolved part restated by someone with a procurement decision attached to the answer. Followed through in Benchmarking your own agent spend.

Someone who expects the transfer to hold

The first direct contradiction of this page in the notebook, and it is worth stating as a disagreement rather than smoothing into a synthesis. Greenblatt is asked more or less Karpathy's question — does capability bought inside the rails arrive outside them — and answers the opposite way: in practice the transfer "seems okay". His evidence is the absence of a counter-example. He says it is hard to point to a domain that is genuinely hard to verify where the improvement from GPT-4 to a current frontier model has not been large.

Both of them are describing the same models, so the disagreement has to be about what counts. Karpathy's instrument is a ceiling: the atoms joke is unchanged, which shows a capability that has not moved at all. Greenblatt's is a slope: hard-to-verify work is much better than it was, even where it is still worse than a competent professional — a distinction he makes explicitly and does not try to hide behind. Those are compatible descriptions of one situation, and they license opposite forecasts, which is why the crux matters more than the observation.

Two things sharpen it rather than resolve it:

The version of the disagreement that could actually be settled is on this page already, in the section above: DoorDash builds an RL environment around a task, the models crush it, and the real data does not hold. Greenblatt's answer to that specific pattern is that environments built from production data close the gap — which is a prediction, about a thing several companies are now doing, that will be visible either way.

The unresolved part

He does not claim to know whether the remaining roughness is a model limitation or an operator one, and says so. That hedge is discussed in the interview note; it applies here more than anywhere else in the transcript, because jaggedness is the one part of his account that is an argument about the models rather than about how you drive them.

There is also a precedent for the observation that is older than the vocabulary. Efros and Torralba said in 2011 that computer vision was optimising a definition of success it had chosen itself, and the field's best model at the time had been fitted to its benchmark's quirks one false positive at a time — see Optimising for the benchmark. What that episode adds is a reason not to read this page as a claim about a permanent shape: the ridge moved, and what moved it was a harder measure rather than a better model.

Two things follow that the interview does not address. If capability is concentrated where rewards exist, then a benchmark suite is a map of what was optimised, not of what the model can do — and the gaps are unmeasured by construction. And the same argument predicts that jaggedness gets worse as the verifiable domains advance, not better, which is the opposite of what "the models keep improving" implies. Neither is tested in the transcript. The first one has since acquired a worked example: the scrubbed task that passes and the real one that does not is a benchmark reporting on the version of the job it could grade.