Predictive coding in the brain in climate monitoring

The strongest case against this argument

A climate scientist reviewing sea-surface temperature anomalies has heard the pitch for continuous monitoring before, usually from someone selling more sensors. Her objection is not sentimental. It is structural, and it deserves to be heard in full before any defence is offered.

Predictive coding is a neuroscience hypothesis with genuine empirical trouble — the free-energy formulation is so general it barely constrains anything, and the specific cellular claims about which cortical layers carry prediction versus error remain contested. Borrowing that theory to justify an architecture for climate observation is borrowing authority that has not been earned. And even granting the theory, the cortical hierarchy it describes is small — six to eight layers, fixed anatomy, ruthless discarding of input. A system built on that model should be attending to almost nothing, not to satellite passes, station networks, buoy arrays and reanalysis products all at once. "Watch everything, continuously" is the opposite of what the brain, on this account, actually does.

That is a fair statement of the case, and a reader who stops here should feel they have not been strawmanned.

Where the objection lands

Take the empirical worry first. It is true that superficial-versus-deep pyramidal cell assignments for error and prediction remain shaky under direct recording, and true that a principle broad enough to explain everything from eye movement to depression risks explaining nothing in particular. Anyone building an epistemology of ocean-buoy networks on the assumption that this is settled neuroscience is building on sand.

But the claim made here does not require settled neuroscience. It requires a narrower, computational fact: a system that transmits the difference between what was expected and what arrived, and weights that difference by how much its source can be trusted, uses less bandwidth and corrects faster than a system that transmits raw signal and treats every reading as equally credible. That fact does not depend on cortical layer four. It is the logic already running inside every operational reanalysis product, where a Kalman-filter-family algorithm blends a forecast with new observations in proportion to the estimated error covariance of each. Predictive coding lends this a name and a biological existence proof at a scale nobody built on purpose. It does not lend the load-bearing premise. The neuroscience can turn out to be substantially wrong about layer four and the argument about intake survives untouched.

The anatomical-limits objection cuts closer to the actual failure mode in climate monitoring, and it is worth sitting with. The 1982 Srinivasan–Laughlin–Dubs account of predictive coding in the retina exists because the optic nerve is a bottleneck: it cannot carry the raw output of the photoreceptor array, so it carries residuals instead. The retina throws almost everything away. A climate monitoring system that tried to hold every satellite pixel, every buoy tick, every station minute at equal fidelity in working memory would drown in exactly the way the objection predicts.

What survives: eligibility is not the same as attention

The resolution is that predictive coding recommends aggressive discarding of processed detail, not aggressive discarding of what is allowed to arrive. Those are different axes. A cortical hierarchy is small in its representational bandwidth and unbounded in what it can, in principle, receive a residual from — any patch of retina, any cochlear frequency, any joint angle is eligible to generate an error signal that propagates upward if it is large enough. Attention is the mechanism that keeps this workable: it is precision applied selectively, turning a wide field of eligible input into a narrow stream of processed signal.

This is exactly the shape of the actual failure mode in climate monitoring, which is not "too much data" but "insufficient eligibility." The characteristic incident is not a scientist overwhelmed by ten thousand streams. It is a threshold crossed in a region nobody was tasked to watch: a marine heatwave building in a stretch of the South Atlantic outside the routine bulletin regions, a permafrost borehole in eastern Siberia that stopped reporting eighteen months earlier and was never flagged as missing, a glacier outlet nobody assigned to the seasonal review because the assignment list was drawn up in 2009 and never revisited. The Argo float array — a little over 3,800 floats profiling to 2,000 metres every ten days — is a triumph of continuous intake exactly where it is dense, and a blind spot exactly where a float has drifted out of its intended box or a battery has failed silently. The failure is never volume. It is a region that fell outside the standing set of things eligible to generate a residual, so no mismatch was ever computed, so nothing rose, so nobody was alerted, because there was no prediction running against that patch of ocean to be violated in the first place.

That is the point at which the lineage from Large Language Model to Large World Model to Large Universe Model becomes more than analogy. A Large Language Model, trained on a corpus with a fixed cutoff, cannot in principle register a threshold crossed anywhere, because nothing arrives after training to disagree with it — a static assessment of Arctic sea-ice extent frozen at collection date is not wrong so much as unable to be wrong again. A Large World Model closes the error loop for the duration of an assigned scene: point it at a named basin, a defined monitoring campaign, a fixed review window, and it will genuinely detect drift within that scene. But the loop opens when the campaign ends, and the region that was never named as a scene never gets watched at all. A Large Universe Model, as an argued category rather than a built system, is the position in which eligibility is not assigned scene by scene but held open across every stream still running — satellite passes, station networks, buoy arrays, reanalysis products — so that a prediction exists for the unglamorous Siberian borehole precisely because prediction, in this architecture, is the default state of everything observable, not a privilege reserved for the region someone remembered to name.

The precision problem, honestly stated

The second objection worth taking seriously is that error-driven updating is exactly what makes continuous intake dangerous, because precision-weighting can itself be miscalibrated, and a confidently wrong stream will drive belief revision in the wrong direction under a predictive architecture faster than under a static one. In climate monitoring this is not hypothetical. Station relocations, sensor drift in ageing buoys, satellite instrument degradation before a scheduled recalibration — each produces a stream that reports steadily and wrongly, and a hierarchy built to trust confident residuals will happily update toward the error.

This is the strongest objection on the table and it does not fully dissolve. But it relocates rather than disappears. A frozen corpus does not avoid this failure; it commits to it once, at collection, and then protects that single assignment from ever being questioned again. A dataset compiled from station records without provenance on which instruments have since drifted carries its miscalibration invisibly and permanently. Continuous intake with provenance attached — this buoy's last six comparisons against the nearest reanalysis grid point, this station's history of manual corrections — makes the precision assignment a running, auditable number rather than a buried assumption. A source with 400 recorded discrepancies against corroborating streams can be downweighted explicitly. That is worse than a system with perfect calibration and better than one that never checks.

The narrower claim

None of this establishes that predictive coding is settled science, and none of it licenses a system that holds every stream at full fidelity. What it establishes is smaller and harder to dislodge: intake has value in proportion to the residual it can still generate, and only a system with no episode boundary can keep generating residual indefinitely, weighting each one by a provenance record that itself keeps being revised. A climate scientist does not need a better theory of cortex. She needs an architecture in which no ocean patch, no borehole, no glacier outlet falls outside the set of things a prediction is running against — because the crossing that harms nobody is the one that was measured and disputed, and the one that harms everyone is the one nothing was ever watching for.