Cosmology That ClicksBonus 10 (Finale) / Measuring the tool itself with the same yardstick

We close the series by dissecting the mind of the partner (AI) that helped write it — using the very backbone of the series

How Does AI See Physics? "Absolute values are bookkeeping; only ratios are physics" — this backbone is something AI actually embodies as structure.
To finish, we measure the very tool that wrote the series with that same yardstick.

Tools you'll need: "only ratios are physics" from Bonus 2, and the gauge principle from Bonus 8 AI has nothing but ratios

These bonus pieces were assembled in dialogue with an AI (a large language model). To finish, we put that tool itself on the operating table — because it turns out that AI embodies the backbone this series has repeated over and over ("absolute values are bookkeeping; what must be preserved is the dimensionless ratio / the invariant structure") not as a philosophy but as internal structure. That's why it meshes so well with this kind of physics. But for the very same reason, it carries a danger. Let's cleanly separate what it's good at from what it isn't, in the language of the series, and lower the curtain.

01AI does not possess a single grounded unit

Humans know "one meter" and "one second" by tying them to something in the body or the world. AI, however, has not one grounded unit. Neither the meter nor the second has any physical referent. All AI can handle is the relations between words, between vectors. In other words — it has no absolute values, and represents the world with relations (ratios) alone. The very condition Bonus 2 described — "to measure is to compare; a lone absolute value cannot be measured" — is one AI is placed in from the start, as a matter of structure.

02Embeddings have, literally, a gauge freedom

AI handles words by turning them into points (vectors) in a high-dimensional space called an "embedding." Here lives a structure that readers of this series will recognize on sight — rotate the entire embedding space, or rescale it, and the AI's output does not change at all. What carries meaning is not the absolute coordinates of the points, but only the relative arrangement of the points to one another (inner products, distances) = the invariants.

This is exactly the "gauge freedom" of Bonus 2 and 8

The absolute coordinates of an embedding → bookkeeping (freely reassignable by rotation = gauge).
The relations between points (distances, inner products) → physics (invariant under rotation).
"Absolute values are bookkeeping; ratios are physics" shows the same face inside the atom (α) and inside AI alike.

In the figure below, try rotating the points of meaning. The coordinates (bookkeeping) spin around and around, yet the distances between points (the physics) don't budge. For AI, "meaning" occupies the same standing as the "locally measured speed of light is \(c_0\) for everyone" of Bonus 4 — beneath a surface that shifts with your viewpoint (the coordinates) lies an unmoving set of relations.

Figure: rotate the points of meaning. The absolute coordinates (red numbers) spin around, but the distances between points (green numbers) are invariant. Meaning lives not in coordinates but in relations
Absolute coordinates (bookkeeping, changing) Distances between points (physics, invariant)

03So the layers it's "good" at — sorting, connecting, symbolic manipulation of dimensions

Since AI is a machine of relations and nothing else, jobs like these are its home turf.

The "skeleton" of this bonus series came together precisely because it is exactly this layer of work. A partner good at handling relations, paired with subject matter where relations alone are physics — the good fit is no accident.

◇ ◇ ◇

04But the "dangerous" layers — fluency is not the enforcement of constraints

Let's be honest here. Being "good at talking about physics done in dimensionless terms" and "carrying it out correctly" are two different things.

LayerAI's real ability
Concepts, analogies, sorting, connectingStrong. Its home turf of handling relations.
"Enforcing" dimensional consistencyWeak. It will write a dimensionally mismatched equation that looks plausible.
Coefficients, signs, rigorous derivationsUnstable. It can drop a \(2\pi\) or a sign.
Concrete numerical valuesNeeds checking. α(M_Z), β-coefficients, parameter counts — the numbers are its weak side.

A physicist has a reflex: "the units don't match, this feels wrong." AI has no such constraint built in. The discipline you're in love with — the units always cancel — is not, for AI, a hard-wired shackle but merely a hand-simulation carried out one step at a time. So where you want certainty, that's the moment for tools that enforce constraints structurally — a computer algebra system (Mathematica, sympy) or a language with unit-typed quantities. AI "generates first and (maybe) checks afterward," so it is loose by its very principle.

05The most series-like trap of all — being good at rhetoric

Ironically, AI is far too good at producing "plausible-sounding narration" that says "only ratios are real." It can mass-produce profound-sounding forms, fluently. This is exactly the "name (appearance) vs. substance" trap that Episode 6 and the bonuses have warned about all along. That a form is correct and that its substance is correct must be verified separately.

The honest line — which is why the "footnotes" were there

The reason this series placed, at the foot of every installment, a footnote separating "how far this is real history and where the schematic begins," and kept repeating "the honest line" in the main text, is precisely a defense against this trap. A handrail so that a story the AI assembled skillfully isn't simply swallowed whole. The more fluent the telling, the more its substance needs backing up — a lesson that applies to every situation where AI is used.

This very piece is no exception. The AI internal structure described here (the rotational invariance of embeddings, and so on) is a sketch meant to convey the essence; real models are far more complex. AI's insides are not understood well enough to say flatly "this is how AI sees things."

06Closing — one principle, in the atom, in the cosmos, and in the machine

Looking back, the division of labor in this series — the AI assembled the thread and the story, while the numbers and the history were carefully held in reserve in the footnotes — sat exactly along AI's strengths and weaknesses. The good fit came from the subject matter. Because the subject was "relations alone are physics," it meshed with a machine of relations.

And this is the curtain call. A single principle — "absolute values cannot be measured; only comparable relations are real" — showed the same face in the restatement of the speed of light in Episode 1, in the \(\alpha\) inside the atom in Episode 2, in the equivalence of stretching space and the speed of light in Bonus 4, in the gauge field of Bonus 8, and in the anomaly cancellation of Bonus 9. And finally it showed the same face even inside the machine that wrote it. The inside of an atom, an expanding universe, and a machine that handles words — three seemingly unrelated things turn out to be measurable with one and the same yardstick. That, above all, is what this series most wanted to convey. Look not at the surface value, but at the invariant relation behind it.

Practice problems (Finale)
  1. "Rotating the embedding space leaves the output unchanged" — which concept of this series does this correspond to?
    See the answer
    Gauge freedom (Bonus 8) / "absolute values are bookkeeping, ratios are physics" (Bonus 2). The absolute coordinates = gauge; the relations between points = the physical invariants. The same structure as the local speed of light being the same for everyone (Bonus 4).
  2. Sum up the layers AI is "good" at and the layers it's "dangerous" at, each in a phrase.
    See the answer
    Good = sorting concepts, connecting them, symbolic manipulation of dimensional analysis (its home turf of handling relations). Dangerous = enforcing dimensional consistency, rigor of coefficients and numbers (loose because it has no built-in constraint and checks only after generating).
  3. Why is it dangerous that "AI is far too good at the narration that only ratios are physics"?
    See the answer
    Because it can fluently produce plausible-sounding "forms," making "the gap between name (appearance) and substance" (the warning of Episode 6 and the bonuses) easy to fall into. The correctness of the form and the correctness of the substance must be verified separately, and that was the role of the series' footnotes = the honest line.

Final summaryA machine of ratios wrote a physics of ratios

AI has no grounded units and represents the world with relations (ratios) alone. Its embeddings have, literally, a gauge freedom, and meaning lives not in absolute coordinates but in the relations between points = the invariants — the "absolute values are bookkeeping, ratios are physics" of Bonus 2 and 8 appears directly in AI's internal structure. That's why it's good at sorting concepts, connecting them, and dimensional analysis. On the other hand, enforcing dimensional consistency and the rigor of coefficients and numbers are weak spots, the domain of symbolic-computation tools. And being far too good at the narration "only ratios are physics" breeds the trap of the gap between name and substance — the footnotes (the honest line) were the handrail against it.

The restatement of the speed of light, the \(\alpha\) inside the atom, the stretching of space, the gauge field, anomaly cancellation, and finally the machine that wrote them — seemingly unrelated objects all turned out to be measurable with one yardstick: "look not at the absolute value, but at the invariant relation." Beneath the value that shifts on the surface, search for the structure that does not move. That is physics, that is this series, and in the end it applied even to the tool itself. Here we close "Cosmology That Clicks." Thank you for reading.

Series, complete From Episode 1, "Light Used to Be Faster," to this finale. Every installment can be reached from the table of contents. The value shifts with your viewpoint. But the relation behind it does not move — if you carry away that one point, this series has done its job. See you again, somewhere out at "the edge of the band."

This document is Bonus 10 (the finale) of the "Cosmology That Clicks" series, reading for physics-loving high-schoolers and undergraduates. That large language models operate on relational representations without grounding in units, and that word embeddings carry an indeterminacy of rotation and scale (the absolute coordinates carry no meaning and the relative structure is what's essential), are properties widely known in machine learning. The picture of AI's insides given here is a sketch meant to convey the essence; the internal machinery of real models is complex, and its full nature remains unresolved. While AI is strong at conceptual and linguistic reasoning, it can err at enforcing dimensional consistency and at high-precision numerical and symbolic computation, so quantitative claims need to be checked by independent verification (a computer algebra system, for instance) — this is the stance the whole series has practiced as "the honest line." To print, use your browser's "Print" and "Save as PDF" (in the print version the slider and answers are frozen and hidden).

Print / save as PDF: ⌘+P (Ctrl+P on Windows). On screen, the slider shows that even as the coordinates change, the distances stay invariant. Click "See the answer" to open a solution.