◄ UD0   two cubi   transmon   error correction
Perceptron theory · the deepest quantum · the scaling story

The perceptron in N cubi

Two cubi gave four amplitudes. Ten give a thousand; fifty give a quadrillion; three hundred hold more numbers than there are atoms in the universe — inside a chip you could palm. That exponential space is the entire promise of quantum machine learning, and, in the same breath, its entire problem. You cannot pour exponentially many inputs in, and when you measure you get back only N bits. The hard truth of this body is that the promise and the curse are the same fact, looked at from two sides.
N cubi2ᴺ amplitudes inside (the promise)  ·  N bits out per measurement (the curse)  ·  advantage = data-dependent, unproven
✓ STRONG

The exponential space & the hardware. dim = 2ᴺ is just linear algebra, and coupled transmons run real two-cubi gates at ~99.x% fidelity. The space is real; the gates are real.

◐ MIDDLING

NISQ quantum nets. Tens of noisy cubi, T₁≈100 µs, no error correction yet — real machines, real demos, but small and decohering before deep circuits finish.

◔ FRONTIER

A real learning advantage. Whether exponential space helps real data — past the loading and readout bottlenecks — is open, and for most tasks looks unlikely. Honesty over hope.

I · The climb — every cubi doubles the world

Add one cubi and the number of joint amplitudes doesn't grow by one — it doubles. The state of N cubi is a vector in 2ᴺ complex dimensions, and an entangling circuit can put a meaningful number in every one. This is the only place a quantum advantage can live: a feature space so large no classical computer can even hold it.

each square is one amplitude of the joint state · the count is 2ᴺ (drawn up to a cap, then stated)
64
2ᴺ amplitudes (the state)
classically, that's…

II · The bottleneck — a quadrillion in, N bits out

Here's the catch nobody puts on the brochure. That vast state is real, but it's sealed. You can't write 2ᴺ numbers into it — you only get to set N knobs. And you can't read it out — a measurement collapses the whole thing to a single string of N bits. The exponential richness exists only between loading and reading, and most of it is unreachable. A quantum model has to be cleverly built so the few bits you can read carry the answer.

2ᴺ amplitudes held inside · the funnel = measurement · only N bits emerge per shot

III · The honest verdict

What N cubi do and don't buy

Do: hold a feature space exponentially larger than anything classical, and run real entangling gates that explore it. For a few special problems with the right structure, that's a genuine, proven speedup. Don't: magically classify your spreadsheet faster. The cost of encoding ordinary data into amplitudes, and of reading answers back out a few bits at a time, usually eats the exponential — and most proposed quantum-ML advantages, on inspection, dissolve into classically-simulable maps. The space is a promise written in a language most data doesn't speak.

So we end the road where honesty demands: a body with the most spectacular resource of all, and the least settled payoff. We built the real thing — exponential state, real gates, real bottleneck — and we won't pretend the universe owes machine learning a free lunch.

the promise= 2ᴺ-dim feature space the curse= load N knobs, read N bits the hardware= transmons, ~99.x%, T₁~100µs, mK the verdict= real for some, illusory for most

It began in a branch predictor guessing a single bit, and arrives here: a register whose state no computer in the universe could write down, that still answers you one bit at a time. Eight bodies, twelve substrates, one neuron — and the same gate kept open the whole way: say what's true, mark what's hoped.

silicon · cortical · dot · light · crossbar · spin · josephson · cubi · two-cubi · sound · DNA · N cubi