Quantum Computing and the Future of Machine Learning

Las Vegas, and a title I almost talked them out of

Interop ITX in the spring of 2018 was in Las Vegas, which is a strange place to give an honest talk about the limits of a hyped technology. The session had been billed as "Quantum Computing and the Future of Machine Learning," and I had a small crisis of conscience about the title before I ever walked on stage, because the honest version of that talk had the word "someday" in it a lot, and Las Vegas is not a someday town.

The room was a good cross-section of enterprise IT — infrastructure architects, a few CISOs, data-platform leads, the kind of practical people who had budgets and had learned to be suspicious of anything on a magazine cover. That suspicion was exactly what I wanted to work with. My goal was not to sell quantum computing. It was to help a room of serious people calibrate: what was real, what was plausible, and what was, for the enterprise in 2018, still firmly science.

I started by saying the quiet part out loud. "Most of you will not run a machine-learning workload on a quantum computer this year, or next year, or the year after. Let's talk about why the topic is still worth your afternoon."

The argument: know which layer of the promise you're standing on

The core of my talk was a plea for precision, because "quantum machine learning" in 2018 was being used to mean at least three very different things, and conflating them was where people got burned.

The first layer was the near-term hardware reality. We were in what the field had started calling the noisy intermediate-scale quantum era — devices with tens of qubits, no meaningful error correction, and coherence times measured in microseconds. Real, improving quickly, and nowhere near able to run the algorithms that made the theory exciting.

The second layer was the algorithmic promise. There were genuine results — Grover's algorithm for unstructured search, and the HHL algorithm for solving certain linear systems, which matters because so much of machine learning reduces to linear algebra. On paper, some of these offered dramatic speedups. But every one came with fine print: assumptions about how you load classical data into a quantum state, about the structure of the problem, about error rates that the hardware of 2018 could not meet. I made a point of naming the data-loading problem specifically, because it was the one people forgot. If it takes you exponential time to get your data into the machine, an exponential speedup inside the machine buys you nothing.

The third layer was the honest enterprise question: does any of this change what I do on Monday? And my answer was: not directly, not yet — but it should change how you think about a few long-horizon bets, and it was already producing useful ideas that ran on the classical hardware you already owned.

Why enterprises in the room should still care

I gave the practical people two reasons to pay attention, neither of which required buying a quantum computer.

The first was cryptographic. A sufficiently large, error-corrected quantum computer would eventually break the public-key cryptography that everything runs on — and while that machine did not exist in 2018 and would not for years, the data you encrypt today can be harvested and stored until it does. That "harvest now, decrypt later" risk meant crypto-agility was a real planning consideration for anyone with long-lived secrets, even in 2018. It had nothing to do with machine learning, but it was the one quantum consequence with a genuine near-term action item, and I did not want anyone leaving the room having missed it.

The second reason was more subtle and more useful for the ML people. The pursuit of quantum algorithms was producing quantum-inspired classical algorithms — new ways of thinking about sampling, optimization, and linear algebra that ran perfectly well on the hardware you already owned. Studying how a quantum algorithm would attack a recommendation problem, for instance, had led researchers to better classical methods for the same problem. So the research was paying dividends on classical machine learning before quantum hardware could run anything at all.

A concrete example: an optimization problem, two honest paths

The example I used was portfolio-style combinatorial optimization — the kind of "choose the best subset under constraints" problem that shows up everywhere from logistics to feature selection in machine learning. These problems are hard classically because the number of combinations explodes.

The quantum pitch was that certain approaches — quantum annealing, and the variational methods that were the hot topic that year — might one day find good solutions to these problems faster. I walked through why that was plausible in theory. Then I walked through why, in 2018, a well-tuned classical solver on commodity hardware beat the available quantum devices on essentially every real instance, because the quantum machines were too small and too noisy. The point was not that quantum was useless; it was that the crossover — the day quantum would actually win on a real problem — had not arrived, and pretending otherwise would only burn credibility and budget. The right enterprise posture was to keep your classical optimization and machine-learning practice excellent, and to watch the crossover honestly rather than hopefully.

The bridge to now

What I really argued in Las Vegas was a discipline, not a prediction: separate the layers of a promise, act on the near-term reality, and do not let a shiny future distract you from doing production machine learning well in the present. The tools of 2018 were modest, and I was careful not to dress them up as more than they were.

Those same principles — models running in production, learning continuously from real outcomes, and explaining the decisions they make — are what we now operationalize at StudioX as autonomous AI workers and Missions. Quantum computing kept its "someday"; the discipline of doing the present well is what carried through, and it is the same discipline the platform runs on today.

---

Related on StudioX: Enterprise AI Platform · AI Workers · AI Missions