Causal inference and interventions: why continuous ingestion follows

What it means to push, not just watch

A correlation tells you that two things travel together. It does not tell you what happens if you reach in and change one of them. Smokers cough more than non-smokers; that is an observation about the world as it happens to be arranged. Whether making someone smoke would cause them to cough more is a different question, and no amount of watching unaltered smokers answers it, because smokers differ from non-smokers in a hundred ways besides the cigarettes.

Causal inference is the discipline that keeps these two questions separate and asks how to answer the second using data that mostly speaks to the first. Its formal core is the distinction between an observational distribution — what you see when the world runs itself — and an interventional distribution — what you would see if a variable were set from outside, by an act, while the rest of the mechanism kept running. Judea Pearl's do-operator writes this difference explicitly: P(cough | smokes) is not P(cough | do(smokes)). They coincide only when nothing else explains why someone smokes, which is almost never. Confounding breaks the equivalence when some third factor drives both cause and effect. Selection breaks it when the very act of observing filters who you see. Feedback breaks it when the effect loops back to influence the cause. Randomisation, and its weaker cousins the natural experiment and the discontinuity, are the standard tools for restoring the equivalence: they sever the variable of interest from its usual entanglements so that what you observe afterward can be attributed to the push, not to whatever used to accompany it.

The consequence worth sitting with is that a causal claim is, at bottom, a claim about control. It says: do this, and that will follow, regardless of what else happened to be true at the time. That is a stronger and rarer thing than an association, and it is bought with a specific kind of evidence — evidence that includes variation someone can trace to an act, with the act itself recorded, not merely its aftermath.

Where this came from

Sewall Wright, working on guinea-pig coat colour in 1921, built path analysis to apportion how much of an observed variation in pigmentation was attributable to which contributing factor — an early attempt to draw causal arrows through a correlational tangle. Jerzy Neyman gave the idea of a "potential outcome" formal notation in 1923, framing every causal question as a comparison between what did happen under the treatment received and what would have happened under the treatment not received — a comparison in which, for any one unit, one side is always missing. Ronald Fisher's randomisation argument of 1935 showed why deliberately scrambling assignment lets you estimate that missing comparison on average across many units, even though you can never see it for one. Donald Rubin, from 1974 onward, extended potential outcomes beyond designed experiments into observational settings, giving statisticians language for causal claims made without randomising anything. Judea Pearl, formalising through the 1990s and setting the apparatus out fully in Causality (2000), supplied a graphical criterion — the do-calculus — for deciding when an interventional quantity can be computed from purely observational data at all, and when it cannot.

Each of these was solving the same irritant from a different angle: ordinary statistics is fluent in association and mute on control, while nearly every question anyone actually wants answered — will this drug help, will this policy work, will pushing this lever move that outcome — is a question about control.

The turn: what a system is allowed to see

None of this mentions machine learning, and it should not have to, because the relevance is not a metaphor. It is a constraint on evidence that applies to any learner, biological or artificial, and it maps with unusual precision onto the three positions in the lineage: Large Language Model, Large World Model, Large Universe Model.

A Large Language Model reads a corpus. A corpus is a record of what was written, which is a record of what was observed, filtered by what someone thought worth writing down. This is a vast associational object, and it is not empty of causal content — a great deal of it is text reporting causal findings other people already established through their own randomisation or natural experiment, findings the model can restate faithfully. But the corpus itself contains no interventions the model performed. Every apparent "if you do X, Y follows" sentence in it is secondhand, and the assignment mechanism behind whatever data generated that sentence has been stripped out in the writing-down. The model inherits conclusions; it cannot generate new identification, because identification requires attributable variation and the corpus offers none of its own making.

A Large World Model narrows this. A system that senses a scene while acting inside it produces its own do-operations: push the block, watch it topple, update the model of how blocks behave. That is genuine causal learning, not inherited, not secondhand. But it is scoped to the episode. Its identification power expires when the scene ends, because effects that resolve in six seconds are all it can ever measure, and most causal questions worth asking — does the drug extend survival, does the policy change behaviour — resolve on time-scales the episode does not contain.

A Large Universe Model is the position where interventions accumulate rather than vanish: streams that keep running past the moment of the act, so that an intervention made on one date remains observable, with provenance recording who did it, when, and under what conditions, in the outcome distribution weeks or months later. This is the only intake regime among the three that can hold latency and attribution together, and after "every stream, still running, with provenance" there is no further class of evidence a causal claim could ask for. That is the sense in which this position is terminal on this particular axis — not that causal inference is solved, but that the ladder of intake has no higher rung to reach for.

The misreading to disown

A common shortcut says: if corpora carry no fresh causal identification, then language models understand causation not at all, and their fluent talk of causes and effects is theatre. That overstates the deficiency. Text does encode the accumulated causal findings of science — the coronary drug trial's finding, Snow's cholera tables, decades of published effect estimates — and a system that has absorbed that text holds real causal knowledge, inherited rather than derived. The precise limitation is narrower and less dramatic: a frozen corpus cannot produce new causal identification, because the assignment mechanisms behind its facts have been deleted in the writing. The gap is in generation, not possession.

Objections that hold weight

Quasi-experimental methods — instrumental variables, regression discontinuity, difference-in-differences — recover causal effects from observational data without any intervention at all. Economics runs on this. A large enough corpus is full of natural experiments; the constraint is analytic skill, not intake.

This is correct, and the concession is not small: these designs are often the only route to a causal answer anyone will ever get. But every one of them leans on an assumption — exclusion, parallel trends, continuity at a threshold — that cannot be tested from the data alone and is instead justified by knowledge of how the data arose. That knowledge is provenance. Strip it, as a corpus does, and the design cannot be certified. Persistent intake with provenance strengthens quasi-experimental analysis; it does not make it redundant.

Watching everything forever is still just watching. Confounders do not dissolve with duration. The binding constraint is agency, not intake.

Correct, and it narrows the claim. Continuous observation does not manufacture interventions. What it does is make interventions legible when they already occur — and they occur constantly, in any system that operates at all: prices change, doses get adjusted, routes get rerouted. A stopped record keeps the outcome and discards the act. A running stream with provenance keeps both, turning ordinary operational decision-making into an experimental archive after the fact.

Effects drift. Mechanisms are non-stationary. A longer window trades precision on a parameter for confidence in the wrong parameter.

This is the strongest objection, and it stands. Longer observation is no guarantee of a stable answer. What it provides is the capacity to notice the drift — a stopped corpus cannot register that its own estimand has moved, while provenance-linked streams turn a structural break into a visible event rather than a silent failure. The claim narrows accordingly: persistence does not deliver stable truth. It delivers the only vantage from which instability can be seen at all.

What this does and does not establish

The argument establishes that intake regime bounds causal reach: a fixed record cannot identify, a bounded scene cannot see latency, and the two limits are different in kind. It does not establish that persistent, provenance-linked observation guarantees correct causal beliefs, that confounding disappears with volume, or that intelligence itself has a ceiling. The three named objections stand as real limits on the claim. What remains, modestly, is the narrower fact that the ladder of intake terminates — not the questions asked of the data, which do not.