In 2018, an eighteen-year-old undergraduate named Ewin Tang was given a research problem: prove that an ordinary computer could never match a certain famous quantum recommendation algorithm.
She could not prove it. Instead she found an ordinary algorithm that matched it.
Then the same technique dismantled several other headline quantum AI speed-ups, one after another.
| Weak claim | What it takes to be believed | |
|---|---|---|
| Baseline | beats a small neural net | beats the best known classical method on the task |
| Data loading | assumes the state is already prepared | counts the cost of getting data in |
| Scaling | shown at 4–8 qubits | the trend holds as qubits grow |
| Problem | constructed to suit the circuit | someone outside the field wanted it solved |
What went wrong
The quantum algorithms had assumed a very convenient kind of access to their data — the ability to reach in and sample it in a structured way, provided by hardware that does not exist.
Tang's insight was simple and devastating: give an ordinary computer that same convenient access, and it can usually do the job too.
So the enormous speed-up was never coming from quantum mechanics. It was coming from the assumption about how the data was stored. Nobody had cheated. Everybody had just forgotten to check.
Four questions for any quantum AI claim
What is it being compared to? Beating a small untuned neural network means nothing. The comparison has to be against the best ordinary method somebody would actually use.
Does it count the loading? If the paper assumes the data is already inside the machine, the cost has been moved outside the sums, not removed.
Does it scale? Results on four to eight qubits are simulations run on a laptop. The question is always whether the trend continues, and very often it does not.
Who wanted this problem solved? Many demonstrated advantages are on problems constructed to suit the method. That is legitimate research. It is not evidence of usefulness.
What survived
Being precise about this matters as much as the criticism:
- Simulating quantum things. Chemistry and materials. The original 1982 idea, and still the best one.
- Learning from quantum data. Real proven advantages, because an ordinary computer cannot even get at the input.
- Constructed kernel problems. Genuine theorems on artificial problems.
— Ewin TangThe lesson is not that quantum computing does not work. It is that we should be careful about what we are comparing against.
Where that leaves you
You now know how these methods work and how to check whether a claim about them is real. That combination is genuinely rare — most people who read about this field have one or the other.
Go and be the person in the meeting who asks the four questions.
Worth remembering
- Several quantum AI speed-ups were matched by ordinary computers given the same assumptions.
- The advantage came from how the data was assumed to be stored, not from quantum physics.
- Ask: compared to what, does it count loading, does it scale, and who wanted it solved?
- Simulating quantum systems and learning from quantum data remain genuinely solid.