Doing ML to Build Quantum Computers to do ML With

The title is borrowed, with affection, from Spacemen 3's 1990 album Taking Drugs to Make Music to Take Drugs to.² I first described my own motivations in those terms in 2019, and the phrase has been rattling around my head ever since. It says something I had trouble putting more respectably: machine learning and quantum computing form a loop, each field feeding the other. (A strange loop, if you want to be Hofstadterian about it.) For most of the years since, only one half of that loop had much evidence behind it. A paper posted this month gives the other half a rigorous foothold, and it suggests the loop may be tighter than I'd imagined.

How I got here

I came to quantum computing from the machine learning side. My training is in physics and mathematics, but my career was built on applied learning systems, including robotics, from computer vision for autonomous underwater vehicles to reinforcement learning for robot control. When I started looking seriously at quantum computing's problems around 2018, what struck me was how familiar the engineering felt. I'd spent years trying to get machines to do useful things in noisy, partially observed environments, where the dynamics were only roughly known and the feedback came late and incomplete. At the systems level, a quantum computer poses that same problem.

Superconducting qubits drift, neutral atoms shift around in imperfect traps, and photons get lost in waveguides. These are control problems, and what makes them quantum is that the state space is exponentially large, measurement collapses the state you're measuring, and the noise is rich and entangled (literally) with the computation itself. What you do about them has the same structure as every other control problem I'd worked on: sense, decide, act, and learn from the outcome so you do better next time.

That parallel is what pulled me in. The systems being built to control quantum hardware were going to need the kind of intelligence roboticists had been developing for decades, and few people seemed to see it. Calibration, decoding for error correction, allocating resources across a processor, compilation that adapts to the device's current condition: I expected learned, adaptive policies to beat hand-crafted heuristics at all of these eventually, for the same reasons they win in every other domain once it gets complicated enough.

I've spent the seven years since building teams and infrastructure to make that case concrete. At Q-CTRL I worked across several qubit modalities, including neutral atoms. At the AWS Center for Quantum Computing I led the distributed runtime and control stack for superconducting hardware, and at PsiQuantum I work on AI and machine learning across fault-tolerant quantum computing, including neural decoders for error correction, reinforcement learning for adaptivity, and multi-agent systems for calibration, control, and operations. Since 2025 I've also been doing independent research on the agent infrastructure and scientific reasoning tools I think this field will need. All of that is the first half of the loop, doing ML to build quantum computers.

The skeptical case

The second half, quantum computers doing ML, has a messier history. Quantum machine learning in the 2010s was dominated by variational methods and by algorithms that assumed QRAM, and the field built up a reputation for overpromising. Ewin Tang's dequantization results, and the work that followed them, showed that many claimed speedups disappeared once someone thought carefully about the classical baseline. Results on barren plateaus cast doubt on whether variational circuits could be trained at scale. And QRAM, the assumption underneath most quantum linear-algebra algorithms, turned out to hide an enormous cost. Maintaining coherent random access to classical data took so much classical machinery that the machinery could often just solve the problem directly.

Plenty of reasonable people looked at that record and concluded that quantum advantage for machine learning on classical data probably wasn't real, or at least wasn't practical. I understood the skepticism, and I shared some of it. I also thought the verdict was premature, because the field had mostly asked whether existing ML algorithms could be sped up on a quantum computer. I was more interested in whether there are information-processing tasks on classical data where quantum mechanics gives a structural advantage that no classical approach can replicate.

A paper posted to the arXiv on April 8, 2026, answers that question with a yes, and I find the argument convincing.¹

Quantum oracle sketching

The paper is "Exponential quantum advantage in processing massive classical data," by Haimeng Zhao, Alexander Zlokapa, Hartmut Neven, Ryan Babbush, John Preskill, Jarrod McClean, and Hsin-Yuan Huang, from Google Quantum AI, Caltech, MIT, and Oratomic. If you follow the field, that author list will get your attention.

Its central idea is a technique the authors call quantum oracle sketching, aimed at the problem that sank the older approaches: getting classical data into a quantum computer. Quantum algorithms want to query the data in superposition, but the data is classical, and storing all of it in QRAM defeats the purpose. Oracle sketching builds an approximate quantum oracle a piece at a time from a stream of classical samples. Each sample is processed once and thrown away, so the full dataset is never stored and no QRAM is needed. The approximate oracle comes with a rigorous error bound, and it's accurate enough to run any quantum query algorithm against.

The clever part is how each sample updates the oracle. If every sample just drives a small random rotation, the randomness eats away at coherence and the error grows like N²/M, where N is the dimension of the data and M is the number of samples. That's far too slow to be useful. The authors instead design the rotations so that different data points act on orthogonal subspaces, which brings the error down to N/M. Holding the error fixed, that's the difference between needing on the order of N² samples per oracle query and needing on the order of N, a factor of a million for a million-dimensional dataset. With M = O(N) samples per query, you can run any quantum algorithm that only needs oracle access.

Pair that with classical shadow tomography for reading results out, and you get end-to-end quantum advantage for classification (through a least-squares support vector machine), dimensionality reduction (through PCA), and solving linear systems. A quantum computer with poly(log N) qubits can do these tasks, while any classical machine needs exponentially more memory to reach the same accuracy. The authors check this numerically on real datasets, including sentiment analysis of movie reviews and single-cell RNA sequencing of blood cells, and find reductions in machine size of four to six orders of magnitude using fewer than 60 logical qubits.

The advantage is also unconditional, which is unusual for claims of quantum advantage. It depends only on quantum mechanics being correct, with no appeal to conjectures from complexity theory, and it holds even if BPP = BQP and even if the classical machine gets unlimited time. It's an advantage in space rather than time, and it's information-theoretic, which is a different kind of claim from the ones quantum computing people usually argue about.

The authors also argue that verifying these results experimentally would test quantum mechanics at the complexity frontier, the way Bell-inequality experiments test nonlocality. I think the comparison holds up. If a small quantum device can compress a huge classical dataset into an accurate model that no classical machine of comparable size can produce, that tells us something about the physical reality of exponentially large Hilbert spaces that no earlier experiment has.

Where the loop closes

Put the two halves together and you have the loop in the title. We need ML to build quantum computers. Running fault-tolerant hardware means decoding, calibration, adaptive control, and resource management at a scale and speed far beyond what people can do by hand, so AI-based control has become a prerequisite. I've made that argument at length elsewhere, and the evidence keeps piling up. Google's AlphaQubit showed a neural decoder beating the best existing decoders on real data from Google's Sycamore processor. And the designs for the next generation of machines, from Google, Oratomic, PsiQuantum, IBM, and the neutral-atom community, are converging on architectures where the classical control plane has to be much smarter than lookup tables and threshold checks.

Now there's a strong theoretical case, backed by numerical results on real data, that the quantum computers we're building with ML will be able to do ML that classical computers can't. Better ML for control makes better quantum hardware, and that hardware can in turn do things for ML, like the compression of classical data that Zhao et al. describe.

The part I find most interesting is where the two halves meet. The quantum computers of the next decade will be run by ML systems, and those systems will generate enormous amounts of operational data: syndrome measurements, calibration trajectories, and performance telemetry from thousands of qubits over months of continuous operation. Turning that data into better control policies for the next generation of machines is itself a massive classical data problem, and if these results hold up in practice, the quantum hardware that produced the data might process it more effectively than a classical machine could.

I want to be careful about scope. Zhao et al. prove an advantage in a specific regime: very large datasets where the size of the machine is the bottleneck. Classical ML will remain the workhorse for nearly every application for a long time. My claim is only that ML and quantum computing form a technical loop in which progress on each side speeds up the other, and that the recent results give reason to believe the quantum side has more to offer than the skepticism of the last decade suggested.

What I'd build next

On the ML-for-quantum side, getting from lab demonstrations to utility-scale machines will take control systems that are adaptive, robust, and fast enough: neural decoders with microsecond latency, RL agents that keep calibration tuned as the hardware drifts, and compilers and resource managers that respond to the machine's condition in real time. All of it can reuse the training infrastructure, model architectures, and inference optimizations the wider ML community has built for other problems. Quantum hardware has tight latency limits and unforgiving physics, and mistakes show up immediately, which makes it a good place to push real-time learned control forward.

On the quantum-for-ML side, the paper opens a concrete research program. Oracle sketching needs to be extended to more kinds of learning problems. The authors point out that the algorithm is mostly commuting operations, which they suggest leaves a lot of room for parallelization, and exploring that will take serious hardware-software co-design. Someone needs to build hybrid pipelines that combine oracle sketching with classical pre- and post-processing, scale up the numerical experiments, and eventually run them on quantum hardware. And we should map out which other fields are bottlenecked on processing massive data, with the physical sciences, genomics, and earth observation as obvious places to start.

Seven years on

The years since I first used this phrase have been long on engineering and shorter on vindication than I'd have liked. The skeptics of quantum ML were right to doubt the specific claims being made back then. I think they were too quick to close the book on the broader question.

I hope the Zhao et al. paper gets more people to treat ML and quantum computing as one coupled system. For now they're mostly two fields that occasionally collaborate. I keep coming back to the robotics idea that brought me here: the body and the brain co-evolve, and the interesting behavior comes out of their interaction. I think something similar is true of quantum computing and ML, and that the capability neither field can reach alone will come out of the loop. We're early, and nobody knows yet what that looks like, but both halves now have evidence behind them, engineering results for the first and an unconditional proof for the second.

Doing ML to build quantum computers to do ML with. I've been saying it for seven years, and it's nice to finally have a paper to point to.