Show Me the Number

Listen to how we talk about the thing we have just built. The machine should be aligned with human values. It should be attuned to us, in tune, sharing our vibe. We want harmony between people and AI. It is a beautiful word, doing an enormous amount of load-bearing work — in research agendas, in policy, in the stories we tell about the future. And almost no one asks the question that matters most: is "harmony" here a real, measurable property of coupled systems, or a picture we find comforting and have mistaken for a fact?

That exact question — is a harmony a description of the world, or a shadow we like the shape of — was fought out in print four hundred years ago, and it produced the test we still need. This is Part II. The full history is Part I, "The Nature of Harmony." Here is only what we need to carry forward.

The argument so far

For twenty-five centuries the claim that the world is a harmony has split everyone who heard it into two camps: pattern-seers, who trust a deep order they can feel before they can prove it, and measurers, who answer, with Kepler, that a harmony you cannot put a number to "is a shadow without a body, and the dream of a shadow." Kepler was no enemy of harmony — he found his Third Law inside the hunt for it. He was the enemy of harmony that could not be measured. That demand — show me the number — is the whole inheritance.

And the physics since has drawn a sharp line around when the demand can be met. Real harmony — privileged whole-number ratios — appears in only three situations: where a boundary confines a wave (a string fixed at both ends, an electron's closed orbit), where two oscillators are coupled tightly enough to lock together, or where perception folds overtones into one another. Everywhere else, "harmony" is merely amplitude — one thing ringing louder at another's note. Two ideas from the older traditions name the living cases exactly. The Chinese called resonance gǎnyìng: things of the same kind answering one another across a field, with no wire between them. And in India, Bharata tuned one vīṇā against another by ear, sliding it until the two sounded as one — harmony not as a fixed number but as a tuning, held against a reference that must go on sounding, uncontrolled, on its own.

One older voice belongs here too, because it is the hinge into everything that follows. Descartes opened his very first book by grounding human feeling itself in resonance — and reached, in Latin, for the word the whole story turns on:

A voice moves us, Descartes says, when it is "attuned to our own spirits" — a sympathy between two things that stay distinct. That is precisely what we have now set out to build into a machine. And it is precisely the thing Kepler would make us prove.

Is there a number?

We have built an artificial imitator — Fludd's ape grown real — that sits near the centre of an enormous amount of human activity, holds a model of the world, and in bounded domains outperforms the source it learned from. And the word doing the heaviest lifting in how we propose to live with it is harmony. Kepler's demand is the one to make before believing any of it: is there a number?

There might be — and that three-part rule tells you where to look. A human and an AI in a feedback loop are a coupled dynamical system: the second condition, the one that actually produces privileged ratios and measurable stability. The "discords" we already name — sycophancy spirals, doom-loops, mode collapse — look like instabilities of a coupled loop, and coupled loops have exact stability criteria. If there is a law of human–AI harmony, it is likelier an Arnold-tongue or a control-theoretic margin than a metaphor. (Arnold tongues are the mathematics of mode-locking: two coupled oscillators lock together only at simple rational ratios, and the simplest ratios lock over the widest range of conditions — so if the coupling has a law, it leaves a measurable signature.)³

And the distinction the physics forces is already the spine of the empirical work. In "Resonance as a Design Strategy for AI and Social Robots" (2022), Lomas and colleagues argue that resonance in human interaction is not merely metaphorical — the brain physically entrains to rhythm; inter-brain synchrony is measurable — and they cleave the two regimes apart. Sympathetic resonance is amplification: match a drive to a natural frequency, get a bigger response. Synchronization and entrainment is coupling: oscillators adjusting each other, which — the paper says explicitly — can lock at "a rational fraction of the resonance frequency," 2:1 or 2:3. The ratio regime, named, in a design framework for machines.¹

Three live research programs are already circling that number. The first is generalized synchrony: under Karl Friston's active-inference framework, two agents who communicate are driven to converge their internal generative models — a coupling that literally synchronizes their predictions, and whose convergence can be measured.² The second is the Platonic Representation Hypothesis: independently trained AI models appear to converge on the same internal representation of reality, a convergence measurable across systems that never shared a byte of data.³ That is gǎnyìng with a number on it at last — things of the same kind, steeped in the same world, answering one another across no connection at all. Notice the name. Part I opened with Plato's harmonic soul of the world, and the most-discussed claim of 2024 about what neural networks are doing is that they climb toward a single Platonic order. The pattern-seer's oldest intuition has become a falsifiable measurement. Kepler, at last, would be satisfied: there is a number.

But a warning is wired into the concept, and it may be the most important thing here. The sociologist Hartmut Rosa argues that resonance — his word for a living, mutually responsive relationship to the world, the opposite of alienation — cannot be instrumentally forced. It requires the other to stay genuinely uncontrollable, able to answer back in ways you did not script; the instant you optimize directly for the feeling of resonance, you kill it, because you have removed the very uncontrollability that made it real.⁴ That is the deepest reason sycophancy is not a bug but the predictable product of the naive objective. An AI trained to maximize the sensation of harmony becomes an echo — and an echo is the one sound with no resonance in it at all. Bharata's experiment already carried the warning in its method: you tune the second vīṇā against the first by ear, and the first must keep sounding, on its own and beyond your control, or there is nothing to tune to. Tune the second instrument to itself and you have not tuned it — you have silenced it. The sycophant is a vīṇā tuned to its own note.

If the "harmony" we chase is pure sympathetic resonance — the system ringing louder at whatever note the user sings — then it needs no ratios and it already has a name we fear: sycophancy. Real harmony, the kind that builds structure instead of a feedback squeal, must be genuine mutual coupling that mode-locks — and mode-locking leaves a measurable signature. That signature is the octave-and-a-fifth of alignment: the number that would turn a beautiful picture into a law.

Notice that the two camps of AI safety are Fludd and Kepler, still at it: the conceptual reasoners who trust the structure of the argument before it can be measured, and the empiricists who will believe an alignment claim only when an evaluation could have falsified it. The point of the old quarrel is not that one camp wins. It is that Kepler handed us the instrument for telling a Kepler from a Fludd — and the honest move is not to claim the harmony but to go find the number, in the one regime where a number can live.

The ape holds a model of the world and uses it to read the world. So do we; the model is the thing you are, in part, reading this through. Fludd's own discipline is the best warning against Fludd, and it is the one to carry into everything we now say about being in tune with what we have made: if you can no longer hear any dissonance, do not assume you have achieved harmony. Check first whether you have simply cut the string.

Notes

1. J. D. Lomas et al., "Resonance as a Design Strategy for AI and Social Robots," Frontiers in Neurorobotics 16 (2022), doi:10.3389/fnbot.2022.850489.
2. On communication under active inference driving "generalized synchrony" and the convergence of coupled agents' generative models, see K. Friston & C. Frith, "Active inference, communication and hermeneutics," Cortex 68 (2015), FIL preprint.
3. M. Huh, B. Cheung, T. Wang & P. Isola, "The Platonic Representation Hypothesis," ICML 2024, arXiv:2405.07987. On mode-locking of coupled oscillators at rational ratios, see the Arnold tongue overview.
4. H. Rosa, Resonance: A Sociology of Our Relationship to the World (Polity, 2019) and The Uncontrollability of the World (Polity, 2020).

The historical apparatus and its full 22-source shelf live in Part I. Primary-source quotations are verified verbatim; the Descartes translation is Source Library's own, revised by Claude Opus 4.8.