Quantum Advantage
Quantum advantage is not one milestone but a three-rung ladder, and conflating the rungs is how a narrow laboratory benchmark gets reported as if it means quantum computers already beat classical ones at something people care about. The bottom rung, contrived advantage, has been climbed; the middle rung, useful advantage on a problem anyone outside the field actually wants solved, is still contested; the top rung, large-scale quantum machine learning that outperforms classical ML at scale, is decades of hardware progress away and depends directly on clearing the NISQ-era limitations this term's companion entry lays out.
Contrived advantage means a narrow, artificially constructed benchmark with no practical use, built specifically to be hard for classical computers and easy for a quantum one. Google's 2019 Sycamore claim is the textbook example: a random circuit sampling task the company said would take a classical supercomputer 10,000 years, completed on their quantum chip in 200 seconds. The claim held for less than a year in the strict sense — IBM argued a classical machine could do it in 2.5 days with better simulation methods, and subsequent classical algorithms using tensor networks closed the gap further still. The scientific result survived; the "advantage" framing did not, because random circuit sampling answers no question anyone outside quantum computing research was asking. China's USTC followed with its own contrived-advantage claim on Gaussian boson sampling in 2020, same pattern: real physics, no application.
Useful advantage is the open, actively contested rung: a quantum computer beating the best classical approach on a real-world problem — chemistry simulation, optimization, or materials science — that someone outside the field actually cares about. As of now, no claim has conclusively cleared this bar in a way the field broadly accepts; every candidate result so far has either been matched by an improved classical algorithm within months, like Sycamore's sampling claim, or has involved a problem small enough that its practical value is itself debatable. The candidates getting the most serious attention are molecular simulation with the variational quantum eigensolver, where IBM and multiple pharma and chemical partners are running early trials, and Quantinuum's and IonQ's work on optimization problems using trapped-ion hardware's higher gate fidelity to run deeper circuits than superconducting chips currently tolerate. Nobody in the field claims this rung is climbed; the honest state of the art is "closer than five years ago, not there yet."
Large-scale quantum machine learning is the aspirational top rung: quantum hardware running ML workloads at a scale and reliability that meaningfully outperforms classical GPU-based training, not on a toy dataset but on problems where classical ML is currently the industry standard. Reaching it requires fault-tolerant hardware that does not exist yet, because today's NISQ devices cannot hold a large enough, deep enough, error-corrected quantum state to encode a useful ML model's feature space without the noise overwhelming the signal — exactly the circuit-depth and error-correction ceiling described in the NISQ-limitations entry. IBM's public roadmap targets its first fault-tolerant system, Starling, for 2029, and quantum-ML-specific advantage is explicitly downstream of that, not a separate, earlier milestone; Google DeepMind and IBM Quantum's own research groups both frame large-scale QML as a 2030s-horizon goal rather than a near-term product line.
Reading today's quantum computing news against this ladder is a useful discipline: ask whether a new result climbed a rung or just climbed higher on the rung it was already standing on. A faster random circuit sampling demo, however impressive the raw numbers, is still rung one. A chemistry simulation result that a pharmaceutical partner independently verifies as beating their existing classical workflow would be the first confirmed rung-two result in the field. Rung three will be recognizable because the comparison will not need qubit counts in the headline at all — it will just be a machine learning benchmark result, reported the way GPU-based results are now, because by then quantum hardware will have earned the right to compete on the same terms classical computing already does.