What, How, and Why
A model can recite the cause and still not possess it.
Here is the claim, plainly: today's AI understands What and How. It does not understand Why. And that is not a gap scale will close, because of where the three live.
Take the shadow picture from the first piece and put a machine in front of it. The ideal is the cause — what the world actually is. The recording is the shadow — the surface, the artifact. A language model is trained to predict the shadow: the next token on the surface. It models the artifact, not the thing that cast the artifact. It is a superb continuer of shadows on the wall, having never once modeled the object behind them.
Three rungs
Judea Pearl gives the cleanest map of my three words — a ladder of causation:
- What = association. "X tends to go with Y." Seeing. Rung one — and exactly what next-token prediction is: correlations across the face of the shadow.
- How = intervention. "If I do X, Y changes." A recipe you can run. Rung two.
- Why = counterfactual. "Had X been otherwise, Y would have been otherwise — and here is the cause that makes that true." Rung three.
My claim in this frame: the model is saturated at rung one, fakes rung two, and basically cannot reach rung three.
The faked rung
The faking is the subtle part. A model looks like it understands How — it writes correct code, explains a mechanism, lays out a procedure. But it is reproducing the recorded description of how, not running the mechanism. It learned how-explanations as more surface. That is why it can state a causal rule flawlessly and violate it two sentences later: it holds the text of the How without the machinery of the Why. How is genuine only when it sits downstream of a Why. Cut the Why away and How collapses back into a memorized What about procedures.
Why is the ideal. What and How are its shadows. A system trained on the surface gets What for free, gets How as a description, and structurally cannot get Why — because Why was never on the cave wall to begin with.
The cause never appears in the recording. Only its effects do. You cannot regress your way back to it from correlations, however many you gather. That is not a data-volume problem; it is a type problem. This, I think, is the strongest form of the "scaling alone won't get there" argument — and the useful thing about it is that it is mechanical, not mystical. It says exactly what is missing and exactly why more of the same cannot supply it.
Where Why could come from
If Why is not on the surface, it has to enter from somewhere else. Three candidates, and they are the real design fork:
- Intervention. The system has to act and watch what changes. You cannot infer the cause from the painting; you have to poke the world that cast it. (Agency, feedback, embodiment.)
- A latent world-model. Predict in cause-space rather than surface-space, so the representation is forced to carry the generator, not the generated.
- Counterfactual structure, built in. Architectures that represent variables and their dependencies, not just sequences of tokens.
All three share a single move: stop predicting the shadow; start modeling the thing that casts it. The art metaphor and the frontier of the field are pointing at the same door.
The honest doubt
I should not make it too clean. It is not obvious that humans get Why differently in kind. We may also be running very good What-and-How and then narrating it as Why — confabulating a cause after the fact. Split-brain patients do exactly this: the talking hemisphere invents reasons for actions it did not order, and believes them. If that is what we do too, then "real Why" might itself be a shadow we mistake for the object.
Which loops straight back to the fork from the first piece. Maybe nobody recovers the ideal. Maybe we all just generate convincing readings, and the only difference between us and the machine is how deeply we can invert the shadow — not whether we possess the Why at all.
So the question that decides everything: is Why a thing a system has — in which case giving it to a machine is an architecture problem — or is it a quality of how far you can run the inversion, a limit we are also under? The next piece takes the opposite side of this, and nearly talks me out of the whole shadow.