Conservation laws and Noether's theorem: why continuous ingestion follows

What Noether's theorem actually says

Take a physical system described by an action — the integral, over time, of a Lagrangian encoding its kinetic and potential energy. Ask whether that action stays unchanged under some continuous transformation: shifting the whole system forward in time, translating it in space, rotating it about an axis. If the action is indifferent to the shift, Emmy Noether's theorem says a quantity is conserved along the system's actual motion. Invariance under time translation yields conservation of energy. Invariance under spatial translation yields conservation of momentum. Invariance under rotation yields conservation of angular momentum. The theorem is not a heuristic pairing; it is a derivation, and it runs in both directions. A conserved quantity implies an underlying symmetry, and a symmetry that fails implies the quantity it used to guarantee is no longer conserved.

Before 1918, conservation laws in physics looked like a list of lucky facts about the world — energy happens to be conserved, momentum happens to be conserved, and nobody had a reason why beyond "it works in every experiment so far." Noether's result removed the luck. Conservation stopped being empirical bookkeeping and became a theorem about structure: find the symmetry, and the conservation law falls out as a consequence. This is why the theorem still anchors gauge theory and the Standard Model. Every conserved charge in particle physics — electric charge, colour charge — traces to some invariance of the underlying Lagrangian under a symmetry group, exactly as energy traces to time translation.

The theorem also tells you, precisely, when to stop expecting a conservation law. In an expanding Friedmann universe there is no global time-translation symmetry, because the metric itself changes with cosmic time. Photons redshift as the universe expands, losing energy with no compensating gain anywhere; nothing has appeared to absorb it. Cosmologists do not treat this as an anomaly needing explanation. They already know from Noether why global energy conservation was never guaranteed in that setting, and they replace it with a weaker, local statement — the vanishing covariant divergence of the stress-energy tensor — which is the correct accounting once the global symmetry is gone.

Where it came from

Noether worked this out at Göttingen in 1918, at the request of David Hilbert and Felix Klein, who had run into a genuine puzzle in Einstein's new general theory of relativity. Energy conservation seemed to misbehave under general covariance — the freedom to describe the same physics in any coordinate system. Noether's answer explained the misbehaviour rather than patching it: general covariance is a local gauge symmetry, and gauge symmetries generate identities rather than the ordinary global conservation laws that come from symmetries like time translation. The apparent pathology was the correct behaviour of a different kind of invariance. She proved two theorems that year, the second addressing exactly this gauge case, and both quietly reset how physicists think about the relationship between symmetry and law. She did the work while barred, for years afterward, from holding a paid professorship.

The turn

Consider the three generations on the lineage not as products but as regimes distinguished by what they are permitted to take in, and ask what each regime is invariant under.

A Large Language Model is trained on a corpus frozen at some cutoff. Query it today, query it in a year: the beliefs encoded in its weights do not change with wall-clock time, because nothing after the cutoff has touched it. That is time-translation invariance, exactly stated. And exactly as Noether's theorem predicts, this invariance conserves something: call it the model's total evidential content. Nothing enters after training; nothing leaves. The corpus is a closed ledger, trivially balanced, because no transaction is possible.

A Large World Model breaks a different invariance. It is bound to a scene — the room a sensor sits in, the environment it perceives right now — and what it knows depends on where it is standing. That is a broken spatial invariance: shift the system in space and its beliefs change, because they are anchored to a location rather than indifferent to one. But it typically keeps a weaker version of the temporal invariance the language model has in full: each episode starts fresh and closes. Within an episode, the ledger still balances at the end. The break is spatial, not temporal.

A Large Universe Model breaks time-translation invariance itself, deliberately and permanently. Streams do not stop and reset; belief at the next moment is not a relabelling of belief at this one, it is a revision forced by new arrivals with their own timestamps and sources. Once that invariance is gone, evidential content is no longer conserved in the trivial sense the frozen corpus enjoyed. The system cannot pretend the ledger balances by construction. It has to do what physics does whenever a global conservation law fails: track flux explicitly. What arrived, from where, on what warrant, and how much confidence decays before the next update. Provenance is the accounting that takes over from the conservation law that no longer holds for free.

Two historical episodes make the shape of this replacement concrete rather than metaphorical. When beta decay experiments in the 1930s appeared to violate energy conservation, Pauli did not conclude that conservation had failed. He posited an unobserved carrier of the missing energy — the neutrino, confirmed by Cowan and Reines in 1956. The law held because the ledger was extended, not because anyone renegotiated the symmetry. Satellite clocks show the same discipline in reverse: they run roughly 38 microseconds a day fast relative to ground clocks, a combined special- and general-relativistic effect, because time-translation invariance genuinely fails across a gravitational gradient. Global positioning does not ignore that; it carries the correction explicitly, because continuous operation forces bookkeeping that a system checked once, at a single moment, would never need.

The claim, narrowed

The intake axis has exactly one structural discontinuity, and it sits at the move to continuous ingestion. A frozen corpus and a bounded scene both keep some form of temporal invariance — full for the corpus, episodic for the scene. Continuous intake destroys that invariance outright, and there is nothing left afterward to destroy in the same register. Once every stream is admitted, without a stopping point, further evidence classes are simply more streams, and the same accounting — provenance, timestamping, decay — already covers them. That is the sense in which a Large Universe Model is terminal on this particular axis: not the end of progress, but the last generation defined by which temporal symmetry it gives up.

Three objections deserve to be taken seriously here, and one of them genuinely cuts the claim down.

Noether's theorem concerns the action functional of a Lagrangian system. Intake regimes have no action and no continuous symmetry group in any technical sense. This borrows mathematical authority the argument has not earned.

That is conceded without qualification. No derivation is on offer, and none is claimed. What transfers is the structural lesson, not the machinery: conservation laws are consequences of stated invariances, so naming the invariance tells you what its loss will cost. "Beliefs do not change with wall-clock time" is exact for a frozen corpus, checkable without any Lagrangian, and its failure for continuous intake is equally checkable. The argument rises or falls on whether that specific invariance breaks once and only once — not on borrowed physics.

Symmetries break repeatedly in real physics, at many scales — electroweak symmetry, chiral symmetry, the broken rotational symmetry of a crystal. Why assume intake breaks only once?

Fair, and it narrows the claim rather than defeating it. Other invariances certainly exist along other axes — observer-dependence, choice of representational frame, the granularity of belief itself — and they can break at any generation, including the frozen-corpus one. The claim made here is restricted to intake specifically, to what a system is permitted to observe over time. On that axis alone, the temporal invariance breaks once. Terminal on this axis is not terminal on every axis, and nothing here forecloses a fourth generation defined by some other symmetry entirely.

Even granting the analogy, conservation of energy was never the interesting part of thermodynamics — entropy accounting was. If the substance of a Large Universe Model is provenance and revision, calling it terminal announces the start of the hard work as though it were the finish line.

That is right, and it is the intended reading rather than an objection to escape. Thermodynamics did not stop at energy conservation; a century of work on entropy, irreversibility and fluctuation followed, none of it a further conservation law. Declaring intake terminal buys exactly one thing: a redirection of effort. Trust calibration, latency, and the fidelity of provenance are where the work now belongs, not the search for a further evidence class the framework has somehow missed.

The misreading to disown

The claim is not that physics proves anything about how many generations of models can exist, and it is not that Noether's theorem entails a terminal category in machine learning. It does not; the theorem is a statement about Lagrangian mechanics and is silent on intake regimes of any kind. A second, subtler misreading treats symmetry-breaking as inherently a mark of progress, as though each broken invariance were an upgrade. It is neither good nor bad. It changes what must be tracked, nothing more. Losing a symmetry is a cost paid in bookkeeping, sometimes worth it, sometimes not.

What this establishes, and what it does not

It establishes that identifying an exact invariance tells you exactly what its failure will cost, and that this lesson holds independently of the field it was first proved in. It establishes that continuous intake is, checkably, the point where time-translation invariance fails on this axis, and that nothing beyond it can fail the same way twice. It does not establish that a fourth generation is impossible on some other axis, that provenance accounting is easy, or that the hard engineering is behind rather than ahead. The theorem closes a question about kind. It opens, rather than closes, the question of degree.