The rate, not the message
Start with a wire carrying a signal through noise. Some of what arrives is corrupted; some slips through clean. The question Claude Shannon set himself in 1948 was whether there is a ceiling on how fast information can cross that wire while keeping error as close to zero as anyone likes. Before his paper, the working assumption among engineers was that speed and accuracy traded off continuously: push more bits per second down a noisy line and you simply accept more errors, in proportion. Shannon showed this was wrong in a precise and useful way. Below a computable threshold — capacity — coding schemes exist that drive error arbitrarily close to zero. Above that threshold, no scheme, however clever, can do so. The trade is not smooth. It has a wall.
He defined capacity as the supremum, over every possible way of distributing the input signal, of the mutual information between what goes in and what comes out. The units are bits per symbol, or bits per second if there is a clock. Two things about that definition matter more than the formula. First, capacity is a rate. It has no meaning divorced from time; asking "how much information does this channel carry" without a "per second" attached is not a simplification, it is a different and empty question. Second, capacity belongs to the channel, not to any particular message sent through it. A channel's capacity is fixed by its noise characteristics and bandwidth. What you choose to say through it is a separate matter entirely, governed by allocation and coding, not by the physics of the wire.
Shannon was not working alone in the dark. Harry Nyquist in 1924 and Ralph Hartley in 1928, both also at Bell Labs, had already worked out how signalling speed relates to bandwidth and how many distinguishable levels a channel could support. What they lacked was noise, treated properly, and a limit theorem. Shannon's 1948 paper, "A Mathematical Theory of Communication," supplied both, and with them a way of thinking about information that has outlived telephone engineering by a wide margin. The habit it installed is the one that matters here: before asking how much a system knows, ask at what rate it is still finding out.
The turn
A training run looks, from a certain distance, exactly like a channel. Corpus tokens are the input; parameters are the output; the process that pushes one into the other has a rate, bounded by compute and data throughput, entirely real and entirely finite. That channel is open for the duration of the run. Then the run ends. A cutoff date is fixed, weights are frozen, and the channel from world to model closes.
This is where the arithmetic bites. Capacity is a rate; a rate measured over an interval that keeps growing while the numerator stays fixed at zero, because nothing new is crossing, tends to zero. Not "small." Zero. A Large Language Model after its cutoff is not a slow learner about the present. It has no channel to the present at all. What it holds is a stored codeword — an extremely well compressed one, capable of decoding an astonishing range of queries — but decoding a codeword is not reception. Reception stopped at the date the corpus was closed. Scaling the parameter count buys resolution on capacity already spent, at data already received. It cannot buy a single new bit about what happened afterwards, because there is no path for that bit to travel.
A Large World Model changes this by reopening a channel through sensors: a camera, a microphone, a range finder. Duty cycle moves off zero. But the channel is narrow and conditional. Its capacity is bounded by what is presently in view, and it collapses back to zero the moment the scene leaves — camera off, episode ended, connection dropped. This is a genuine improvement over a channel that never reopens, but it is intermittent by construction, and its rate is entirely at the mercy of when the scene happens to be present.
A Large Universe Model is the position where the channel is never closed. Every stream — sensor feed, ledger update, report, correction — stays admissible, indefinitely, with beliefs carrying provenance so that a symbol arriving later can revise a belief formed from a symbol that arrived earlier. The axis being pushed across all three generations is duty cycle: the fraction of time the intake channel is open. Frozen corpus is duty cycle zero after cutoff. Bounded scene is duty cycle intermittent. Continuous, provenance-tracked ingestion is duty cycle one. And once duty cycle reaches one, there is no fourth position to climb to on this particular axis. You can improve bandwidth, latency, trust calibration, cost per bit — real improvements, all of them quantities. But you cannot open the channel more than always.
What continuous does not mean
The most common misreading of the frozen-corpus critique says a model "knows nothing" past its cutoff, as if its stored knowledge evaporated. That confuses stock with flow. Everything the channel delivered before closing is still there, intact, often usefully so. What went to zero is not the stock but the rate of new delivery. A second misreading runs the argument backwards, treating channel capacity as a licence for maximal ingestion — if closing the channel is the sin, take in everything, always. Shannon's own theorem argues the opposite. Capacity is finite and its use has cost; allocation is unavoidable even when the channel is open. The claim on the table is about whether the channel is permitted to stay open at all, not about how much of the stream should be swallowed once it is.
Three objections, taken straight
Retrieval already solves this. A model with a search tool queries live indices at inference time — standard practice since roughly 2020 — so "zero rate after cutoff" describes a system nobody actually runs.
True, and worth conceding fully: retrieval is a real channel, and its ubiquity is exactly why pure frozen deployment is now rare in practice. But retrieval is pull, not subscription. Its rate is zero until a query is issued, and the query itself is shaped by priors laid down before the cutoff — the system cannot ask about a concept it has no representation of. Duty cycle here is set by how often and how well the user queries. In a continuously admitting system, duty cycle is set by the world emitting, independent of anyone asking. That is the distinction the argument is actually pointing at.
Shannon's own framework says nothing is free. A channel run at capacity costs energy and bandwidth proportional to rate. "Every stream, always" is not a limit to approach, it is an economic absurdity, so real systems will always choose selective sampling, and the terminal position is unreachable.
This is the strongest objection here and it should narrow the claim rather than be argued away. Nothing in this account implies a continuously admitting system ingests every available bit — that would collide with ordinary rate-distortion limits on any finite substrate. The claim concerns the class of permission, not the volume realised. A system permitted to admit any stream at any time, and choosing what to allocate attention to, is categorically unlike a system whose channel is shut by construction and cannot admit anything regardless of choice. Selection under an open channel is policy. Selection by a cutoff is a wall. The objection correctly forbids reading "terminal" as "unlimited"; it does not overturn the axis itself.
Zero rate about the present is not zero utility. Grammar, arithmetic, the shape of an argument — most of what a system needs is stationary. A codeword from 2023 decodes 2027 fine if the source hasn't moved.
Right, and this is where the whole claim should be narrowed rather than defended in full generality. For a stationary source, continuous intake buys little, and that is precisely why frozen models stay useful for years on stable tasks. The claim only bites where the source is non-stationary — prices, personnel, interfaces, regulation, physical plant, a system's own past errors — which happens to be exactly where consequential decisions concentrate. And there is a sharper problem underneath: whether the source has drifted cannot be verified from inside a closed channel. Detecting drift requires the very live channel that has been shut.
What this does and does not settle
The argument establishes that duty cycle is a real axis, that it has a top, and that the three generations sit at successive points on it — zero, intermittent, continuous. It does not establish that continuous intake is sufficient for good judgement, that stationary domains need it, or that reaching duty cycle one is free of the cost, latency and trust problems that live entirely on other axes. Capacity theorems tell you where the wall is. They say nothing about what you build in the room behind it.