Build an advantage, then destroy it
Quantum computing spent a decade learning one lesson in public: a result appears, it is impressive, and then somebody finds an ordinary algorithm that reaches the same place on ordinary hardware. The technique has a name: dequantization. Reading about it produces agreement. Performing it produces the ability to do it again to the next claim, so this page makes you perform it.
Step one: build an advantage that looks real.
There is a table with a value in every row, and a quantity that needs all of them to compute exactly. Every row also carries a cheaply known weight: a rough, noisy proxy for how much that row matters. Now grant a machine one power the ordinary computer does not have: it can sample rows in proportion to that weight, and correct for the bias afterwards.
The weight is deliberately imperfect: within ±20% of what the row is really worth, because a weight that is exactly right would be the answer.
- rows read
- —
- error
- 0
Reads every row. This is the answer the other two are trying to reach.
- rows read
- —
- error
- —
- work saved
- —
Samples in proportion to the weight, then corrects for the bias.
- rows read
- —
- error
- —
- work saved
- —
Given the same sampling access, on ordinary hardware. Not run yet.
Run the machine and read the middle card. Then press attack.
What just happened
The advantage was never in the hardware. It was in the access model: the ability to sample in proportion to a known weight, and that access was granted by assumption rather than earned by the machine. Hand the same access to an ordinary routine and the gap closes exactly, because both are running the same estimator.
The general form: a complexity advantage is a claim about a gap between two algorithms, and gaps of that kind are closed by ideas. Ideas are cheap relative to fabrication. A field whose central claim can be attacked for the price of a graduate student's attention, while defending it costs a fabrication run, is on the wrong side of an asymmetry.
So what does survive?
A new machine that intends to exist in ten years has to answer two questions at once, and each is narrower than the version usually offered.
On work the incumbent already does, what does this deliver per joule?
Against a competently optimised classical implementation, not a naive one, and stating whether the figure was measured or modelled. This is where credibility is earned because it is the side a reader can check. Expect to lose ground here as classical methods improve: that is normal. Claiming Side A as permanent is the error.
What work does this make reachable that the incumbent does not?
The usual phrasing: impossible on traditional hardware: is false as literally stated. A classical computer can compute anything computable. Any claim of the form only our machine can compute X is either wrong or is smuggling in a resource bound without saying so. Say the bound out loud and it becomes checkable.
The three ways to state Side B so it has teeth
The budget form
A task must complete inside an energy and latency envelope fixed by the physical situation: a device on a battery, a loop that must close in milliseconds, a sensor that must run a year on a coin cell. The incumbent cannot fit; this machine can. Most checkable, because the envelope is imposed from outside and neither party can negotiate it.
The native-operation form
The machine performs in one physical step what the incumbent emulates with many. Drawing a sample is the clearest case: a thermodynamic device can be arranged so its equilibrium fluctuations are the sample. The claim is not that the incumbent cannot, but that it pays a conversion cost. State that cost at the boundary: it is where such claims usually fail.
The scale form
Reachable in principle, but the incumbent's version needs more energy than the operator can obtain or more time than the answer stays useful for. Weakest of the three because both quantities move. A claim of this kind should be dated and revisited rather than treated as settled.
Three conventions that make a claim checkable
| Convention | The failure it exists to prevent |
|---|---|
| Record how a figure was obtained, beside what it was | A measured joule, a modelled joule and a joule projected from a datasheet are three different statements. Stored as bare numbers they become interchangeable, and that is the mechanism by which a simulation becomes a headline. |
| When receipts compose, sum the energy and take the weakest provenance | Averaging provenance lets a simulated component inherit a measured component's credibility. The single most consequential error available in this area. |
| Name the baseline and state its optimisation level | A hundredfold gain against an unoptimised reference is a statement about the reference. Where the baseline is not named, the claim is not yet a claim. |
This exercise is the one described in Building the Energy Compute Future (TR-2026-26) §4, with the report's own defaults: 250 samples from 5,000 rows. The estimator is ordinary importance sampling with a Horvitz–Thompson correction, computed in this page: no figures are pre-baked. Constraint vocabulary: the eight.