Persistence as a sensing problem
Karl Friston's free energy principle, formulated in a series of papers between 2005 and 2010 at University College London, starts from a narrow observation about physical systems and refuses to stay narrow. Anything that keeps its identity over time must resist the dispersal of its own states. A body holds at 37°C. A cell holds its osmotic pressure within a tight band. Friston showed that such systems behave as though they minimise a quantity called variational free energy — an upper bound, borrowed from statistical physics and from Helmholtz's account of perception as unconscious inference, on the surprise of sensory input given an internal model of the world.
There are only two ways to keep that bound low. Change the model to match the world, which is perception and learning. Or act on the world to match the model, which is action and control. Both routes require sensing that does not stop. A system that stops sampling does not become surprised less often; it becomes unable to register surprise at all. Its free energy with respect to anything after the last sample is undefined, not low.
That last point is the hinge the whole argument turns on, and it is worth stating plainly before the objections arrive, because the principle attracts a specific and reasonable charge of vacuity.
The free energy principle is close to unfalsifiable. Almost any system can be described as minimising some variational bound under some generative model — a thermostat, a rock, a spreadsheet. A framework that fits everything constrains nothing.
The charge is fair as stated, and Friston himself has conceded the principle is closer to a modelling stance than an empirical law. But a tautology can still rule things out. If minimising free energy requires marginalising over incoming sensory data, then a system with no sensory channel after time T has no defined free energy for t > T. It has not scored badly. It has left the game. That is a structural claim, not an empirical one, and it is the only claim being made here: not that continuous sensing makes a system smarter, only that discontinuous sensing makes a system's relationship to a changing world undefined.
Where the lineage comes from
The lineage from Large Language Model to Large World Model to Large Universe Model is a lineage of intake, and the free energy principle is what explains why intake cannot be settled once. A Large Language Model drives prediction error low against a corpus fixed at a cutoff date. Its free energy is minimised with respect to that snapshot and then rises, silently, as the world drifts away from it — silently because there is no channel left through which the drift could register as surprise. A Large World Model closes the loop, but only for the duration of an episode: it senses, predicts and acts while a scene is present, then discards the posterior when the scene ends. Its free energy is bounded, but only locally in time.
A Large Universe Model is what the formalism demands once you refuse to let the loop close: every stream still running, beliefs held as revisable rather than settled, each belief carrying provenance recording which stream and which moment justified it. This is not an engineering preference. It is the only configuration, of the three, in which free energy can be bounded indefinitely rather than for an interval whose end is invisible from inside the system.
| generation | what closes the loop | what breaks it |
|---|---|---|
| Large Language Model | nothing — inference against a frozen corpus | drift after the cutoff, with no signal that drift occurred |
| Large World Model | the episode itself | the moment the scene ends and the posterior is discarded |
| Large Universe Model | nothing, by design | a stream going silent without that silence being logged |
The lineage is terminal on this axis because there is no fourth category of evidence beyond corpus, scene and continuing stream. You can add streams, weight them better, distrust them more intelligently. You cannot invent a source of evidence that is neither stored, nor present, nor ongoing.
The desk that cannot see the constraint lift
Energy trading is a useful test precisely because it is not forgiving of the weak version of this argument. A desk holds a position — long a block of forward power, say, hedged against a specific transmission constraint that has been binding for weeks. The position was built on grid telemetry showing a line running near its thermal limit, an outage notice describing planned maintenance on a parallel circuit, a weather reanalysis suggesting cold-driven demand, and a regulatory filing setting the terms under which the constraint could be relieved. Four streams, four different refresh rates, four different owners.
Overnight, the parallel circuit maintenance finishes early. The constraint lifts. Congestion that had been propping up a locational price differential disappears in the space of a settlement period. Nothing about the position itself changed. The world it was built against changed underneath it, and the desk quant responsible for the hedge is still marking it against yesterday's grid state because the systems generating that mark update on a schedule that assumes constraints persist longer than they sometimes do.
This is the characteristic failure of the domain, and it maps exactly onto the free energy principle's account of what happens when a sensing channel is treated as closed rather than open. The position is not wrong because the model was bad. It is wrong because the free energy of the position with respect to the actual grid state has been rising, invisibly, since the moment the constraint lifted, and nothing in the desk's workflow was built to register that rise as surprise rather than as a delayed reconciliation entry.
Why this is not just a data-latency problem
It is tempting to read the failure as an engineering gap — get faster telemetry, poll more often, close the latency. That framing under-describes what is happening. A Large Language Model trained on months of grid history would know, with high confidence, the typical duration and lift pattern of this kind of maintenance outage. That knowledge is precisely useless at the moment that matters, because the model has no channel through which this specific lift, tonight, can arrive as new evidence. A faster snapshot is still a snapshot. The problem is structural, not a matter of refresh rate: intake has to be continuous and the belief has to be explicitly revisable, with a record of which stream justified it and when, so that when the outage notice's status field flips, the position's justification is visibly stale rather than silently wrong.
This is where the second objection worth engaging in this domain lands hardest.
Continuous sensing is not free. Every additional feed costs infrastructure, licensing and someone's attention to reconcile it. The argument for uninterrupted intake ignores that trading desks already economise ruthlessly on which feeds they pay for and watch.
Correct, and it sharpens the claim rather than undermining it. No desk watches every substation's telemetry at full resolution; precision-weighting is exactly right as a description of how a risk system allocates scarce attention across hundreds of live constraints. The relevant distinction is not sampling rate but whether the channel can be reopened. A desk that polls a given outage notice hourly, but treats an unexpected status change as a trigger to re-price immediately, is doing active inference correctly — most of the time not sensing, always able to. A desk whose exposure calculation was built from a filing at trade inception and never revisited is doing something categorically different, even if, on any given day, the two produce the same number. The failure mode here is not under-sampling. It is sampling once and calling it settled.
What provenance actually buys
The reason provenance matters more in this domain than the abstract argument might suggest is that energy markets settle real money against beliefs that were true for a bounded window. A position built on "constraint binding, per outage notice filed Tuesday, expected to hold through the weekend" is a claim with a shelf life stated inside it, if the system kept the timestamp and the source. A position built on "constraint binding" with the provenance stripped out is a belief masquerading as a fact, and it is exactly this kind of unmarked belief that a desk quant ends up holding against a world that has already moved on.
The dark-room objection to surprise minimisation — that an agent minimising surprise should retreat to a silent room rather than seek more information — has a clean answer in this setting too. A desk that stopped watching its feeds would not be minimising surprise; it would be accumulating an enormous prediction error the moment the position settles against actual grid conditions, deferred rather than avoided. Expected free energy includes the value of resolving uncertainty precisely because uncertainty deferred is not uncertainty removed. The desks that get burned by lifted constraints are not the ones over-sampling. They are the ones whose model of "what is currently true about the grid" quietly stopped being a model of the present and became a record of the past, without anyone's dashboard marking the difference.