Technical Deep Dives

Quantum Computing and AI: Hype, Reality, and Where They Intersect

Machine learning already decodes errors inside quantum computers; quantum computers accelerating AI is still mostly theory. Google's Quantum Echoes is the first advantage claim classical researchers have not shrunk.

Gabriele Masetti · 2026-08-13

Two Hype Cycles, One Uneasy Overlap

Quantum computing and artificial intelligence are separately over-promised technologies that keep getting bolted together in press releases, as if proximity to "AI" makes a qubit more real or proximity to "quantum" makes a language model smarter. Most of the time the pairing is marketing. But there is a genuine, narrower intersection underneath the noise: machine learning is becoming one of the more useful tools for making quantum hardware work at all, and quantum computing's most credible near-term contributions to AI are still mostly theoretical.

Understanding where the boundary sits requires being precise about what quantum computers currently do, what "error correction" actually means, and which quantum-machine-learning claims have already been quietly disproven by classical computer scientists.

The Physics That Doesn't Get a Shortcut

A classical bit is 0 or 1. A qubit can be placed in a superposition of both, described by complex amplitudes, and two or more qubits can be entangled so that measuring one instantly constrains what you'll find when you measure the other, regardless of distance. That is not "trying every answer at once," despite the popular gloss — a quantum computer doesn't read out a superposition and hand you the best branch.

Algorithms have to be designed so that constructive interference amplifies the amplitude of correct answers and destructive interference cancels out wrong ones before measurement collapses the state to a single classical outcome. That constraint is why only a short list of problems (integer factoring via Shor's algorithm, certain simulation and search tasks) have known quantum algorithms with a proven asymptotic advantage over the best classical alternative — and why "quantum computer" is not a general-purpose accelerator for arbitrary AI workloads.

The other constraint is noise. Physical qubits — whether IBM and Google's superconducting circuits, QuEra's neutral atoms, or trapped-ion systems from companies like IonQ and Quantinuum — lose their quantum state (decohere) in microseconds to milliseconds, and every gate operation has some error probability. Physicist John Preskill named the current period the "NISQ" era — Noisy Intermediate-Scale Quantum — in a 2018 paper published in the journal Quantum, arguing that machines with dozens to a few hundred noisy qubits would be useful for exploratory physics but not yet for changing the world.

That framing has held up better than most 2018-era predictions in tech. It is still, in September 2026, a fair description of where the hardware sits — with one asterisk, added in October 2025, that the rest of this essay has to account for.

Quantum Supremacy, and Why IBM Immediately Disputed It

In October 2019, Google published a Nature paper claiming its 53-qubit Sycamore processor had achieved "quantum supremacy": it completed a specific, contrived sampling task in about 200 seconds that Google's team estimated would take the Summit supercomputer, then the world's fastest, roughly 10,000 years.

IBM disputed the comparison almost immediately, arguing that with better use of disk storage as extended memory, a classical simulation of the same task could run on Summit in about 2.5 days rather than ten millennia — a difference of many orders of magnitude, but not the difference between "possible" and "impossible" that "supremacy" implied.

The dispute was never resolved by either side conceding; it exposed something more useful than a verdict, which is that "quantum advantage" claims are benchmark-specific and classical algorithms keep improving in response, so any such claim has historically had a shelf life until someone found a smarter classical shortcut.

That pattern repeated in 2023. IBM and UC Berkeley researchers published evidence in Nature that error mitigation techniques let IBM's 127-qubit Eagle processor produce results for a quantum Ising model simulation that classical computers of the time couldn't match. Within months, multiple independent groups showed that classical methods — tensor networks and, notably, a technique called sparse Pauli dynamics — could reproduce or beat the quantum results, in some cases on a single laptop core.

IBM's own team acknowledged the classical catch-up was real. The lesson generalizes: treat any single "quantum beats classical" headline as a snapshot in an ongoing arms race, not a permanent verdict, and check whether the comparison task was contrived to favor the quantum device or whether it has independent practical value.

The pattern has one standing exception, and an essay built on skepticism has to say so. In October 2025 Google published a result in Nature it called Quantum Echoes: an out-of-time-order correlator measured on Willow, presented as the first verifiable quantum advantage. Verifiable means another quantum computer of comparable quality can repeat the measurement and get the same answer, rather than the claim resting on an estimate of how long a classical simulation would take.

Google put the speed advantage at roughly 13,000 times the best classical method on the fastest supercomputers, with the largest runs using 65 qubits for the OTOC-squared observable and 95 for the simpler correlator. Nearly a year later, no classical shortcut has been published that narrows it.

The nearest thing to a check ran the other way. An April 2026 preprint found that tensor networks with belief propagation — the cheapest plausible classical attack — cannot reproduce the measurement: the bond dimension required grows exponentially, and the method failed on systems as small as 23 qubits. The authors flag that their incompressibility assumption is empirical rather than proven, and exact classical verification has only been done at small qubit counts. Skepticism here means holding the claim open, not dismissing it.

Why Error Correction, Not Qubit Count, Is the Real Scoreboard

Adding physical qubits is comparatively easy and makes for a good spec sheet; making them reliable is the actual bottleneck. The standard approach is quantum error correction: encode one logical (error-protected) qubit across many physical qubits arranged, commonly, in a surface code, using continuous syndrome measurements to detect errors without collapsing the encoded information, then correcting them faster than they accumulate.

The theoretical promise of the surface code is that as you increase the "code distance" (roughly, how many physical qubits protect each logical qubit), the logical error rate should fall — but only once physical qubit quality clears a fault-tolerance threshold. Below that threshold, adding more physical qubits per logical qubit makes things worse, not better, because you're just giving noise more places to accumulate.

Crossing that threshold convincingly is what Google's Willow chip demonstrated in December 2024. Willow — a 105-qubit superconducting processor from Google Quantum AI (not to be confused with Google DeepMind, a separate part of Alphabet's AI research organization) — showed that scaling a surface code from a 3×3 to a 5×5 to a 7×7 arrangement of physical qubits cut the logical error rate roughly in half at each step, an exponential suppression rather than the flat or worsening curve NISQ-era devices had shown.

The headline number from Google's paper: a distance-7 logical qubit reached a per-cycle logical error rate of about 0.143%, and its effective lifetime exceeded that of the best individual physical qubit on the chip by a factor of roughly 2.4.

Google also ran a random-circuit-sampling benchmark it says would take Frontier, one of today's fastest supercomputers, longer than the age of the universe — a claim in the same genre as Sycamore's, to be read with the same expectation that classical algorithm designers will try to shrink it.

What's different and harder to dispute is the below-threshold error-correction result itself, since it was replicated by scaling the code distance in a controlled, incremental way rather than resting on a single hard-to-verify sampling task.

IBM has taken a different technical bet for its own path to fault tolerance: rather than scaling up surface codes, it is moving to quantum low-density parity-check (qLDPC) codes, which it says can cut the physical-qubit overhead per logical qubit by up to roughly 90% by using more efficient, non-local qubit connectivity.

The first named processor on that roadmap has arrived. IBM unveiled Loon on 12 November 2025, a full year ahead of its own schedule, alongside a 120-qubit processor called Nighthawk carrying 218 tunable couplers. Loon is the more consequential of the two: each of its qubits connects to six others, including vertically, and IBM used it to demonstrate real-time decoding of qLDPC codes on classical hardware in under 480 nanoseconds — the part of fault tolerance that has to keep pace with the machine or is worthless.

After Loon the roadmap is still a roadmap. Kookaburra remained the stated 2026 milestone in September 2026, undelivered; Cockatoo is set for 2027, building toward Starling, a large-scale fault-tolerant system targeted for 2029 and intended to run on the order of 100 million quantum gates across roughly 200 logical qubits at a new IBM Quantum Data Center in Poughkeepsie, New York.

Those later entries are company-stated targets, not independently verified achievements, and roadmaps in this industry have slipped before. On the first of them, at least, IBM was early rather than late.

Separately, in December 2023, researchers from Harvard, QuEra, MIT, and NIST published a Nature paper demonstrating a 48-logical-qubit processor built on reconfigurable neutral-atom arrays — individually trapped and moved atoms rather than superconducting circuits. Superconducting qubits are not the only credible hardware path; neutral-atom and trapped-ion approaches are contributing published error-correction results of their own.

Where AI Actually Helps Quantum Computers Today

The most concrete, unglamorous intersection of these two fields right now is machine learning models trained to do the error-correction decoding itself. Decoding — looking at a stream of syndrome measurements and figuring out, in real time, what error most likely occurred and how to fix it — is a hard combinatorial inference problem, and it has to run fast enough to keep up with a superconducting processor's microsecond-scale cycle time.

In November 2024, Google DeepMind, working with Google Quantum AI, published AlphaQubit in Nature: a transformer-based neural network decoder trained on hundreds of millions of simulated and real error samples from Google's Sycamore hardware. AlphaQubit outperformed the standard algorithmic decoders it was benchmarked against, with reported error-rate improvements on the order of 30% in faster, lower-accuracy decoding regimes and around 6% in slower, high-accuracy regimes, and it scaled to tests with up to 241 qubits.

It is worth being precise about attribution here: AlphaQubit is a DeepMind/Quantum AI research result tested against Sycamore data, distinct from the real-time decoder Google Quantum AI used in the separate below-threshold Willow demonstration. Both point in the same direction, though — that classical machine learning, running on ordinary hardware, is a genuinely useful tool for making noisy quantum processors more reliable, and this is arguably the most mature current use of AI in quantum computing, more mature than any quantum contribution back to AI.

Machine learning is also used more prosaically in quantum hardware calibration — tuning control pulses, characterizing crosstalk between qubits, and predicting drift — work that is real but rarely makes headlines because it looks like ordinary engineering optimization.

Quantum Machine Learning: The Pitch and Its Debunkings

The more speculative direction — quantum computers accelerating AI or machine learning workloads — runs into a problem that doesn't get enough attention: getting classical data into a quantum computer in the first place. Most proposed quantum machine learning algorithms (quantum support vector machines, quantum principal component analysis, and the linear-systems solver known as HHL, after Harrow, Hassidim, and Lloyd) achieve their theoretical exponential speedups only under an assumption that the input data can already be loaded into a quantum superposition efficiently, typically via a hypothetical quantum random-access memory (qRAM).

Building qRAM at the scale and fidelity these algorithms assume remains largely unsolved engineering, and several analyses have pointed out that if you had to pay the classical cost of loading arbitrary data into that structure, much of the claimed speedup evaporates before the "quantum" part of the algorithm even starts. The problem is known as the input or data-loading bottleneck, and it's one of the most honest reasons to be skeptical of QML speedup claims that don't specify how the data gets in.

The second, sharper reason for skepticism is dequantization. In 2018, as an undergraduate, Ewin Tang examined a quantum recommendation-system algorithm published two years earlier by Kerenidis and Prakash, which had been presented as an exponential speedup over the best known classical approach. Tang showed that, given the same kind of data structure the quantum algorithm assumed for input access (sample-and-query access with precomputed norms, essentially a classical stand-in for what qRAM would provide), a purely classical algorithm could solve the same problem with only a polynomial slowdown rather than an exponential one — eliminating the claimed exponential quantum advantage.

The technique, quantum-inspired classical sampling, has since dequantized several other proposed QML algorithms, including versions of principal component analysis and low-rank regression. Quantum algorithm researchers now check for it routinely: before claiming a speedup, does a classical algorithm exist that matches the same input-access assumptions and closes the gap? Several headline QML proposals from the mid-2010s did not survive that check.

None of this means quantum machine learning is a dead end — there are proposed variational quantum algorithms and quantum kernel methods being tested on NISQ hardware for narrow tasks, particularly in quantum chemistry and materials simulation, where the data is naturally quantum-mechanical rather than classical and so the loading bottleneck is less severe. Simulating molecules and materials is widely regarded, including by skeptics, as the most plausible domain for a genuine quantum advantage, precisely because the system being simulated is already quantum.

What "Where They Intersect" Actually Looks Like

Strip away the marketing pairing of two buzzwords and the honest map looks like this: AI, in the form of classical machine learning, is already doing real, measurable work inside quantum computers — decoding errors, calibrating hardware — because that's a well-posed pattern-recognition problem that plays to deep learning's strengths and doesn't require solving the data-loading problem, since the "data" is the quantum device's own noise, generated at the scale you need.

Quantum computers accelerating classical AI or machine learning workloads is the far more speculative direction, dependent on hardware (qRAM) that doesn't exist at useful scale, on caveats about input assumptions that have already dequantized several headline algorithms, and on quantum advantage claims with a track record of being narrowed or overturned by classical researchers within a year or two of publication — a track record that now has one live exception in Quantum Echoes.

The parts of the story that are solid and worth tracking are unglamorous by design: incremental, verifiable error-correction results tied to a specific code distance and a specific error rate (Willow's below-threshold demonstration, the neutral-atom logical-qubit results, Loon's sub-microsecond qLDPC decoding), and named, dated engineering roadmaps with specific claimed milestones (IBM's Starling target) that can be checked against reality as the years pass rather than taken on faith. Quantum Echoes provisionally belongs there too: a measurement another machine can repeat is different evidence from an estimate of what a supercomputer would have needed.

Anyone writing or reading about this space should apply the same test Ewin Tang and the sparse-Pauli-dynamics researchers applied to the claims that came before: ask what the classical alternative can do, under the same assumptions, before accepting that quantum made the difference.