Will quantum computers make AI more powerful?

This is the loop in the name. Half of it ships today; the other half is real, slower, and wildly oversold. Here's how to tell them apart.

You'll be able to rank the four ways quantum might help AI, and say which survive the data-loading toll.

The loop, and which half is real today

Picture two brilliant specialists who can't function alone. AI is a genius that's confidently, fluently wrong just often enough to be dangerous. A quantum computer is a genius that knows nature's deepest secrets but forgets them the instant you stop babysitting it. Point them at each other and something useful happens, but the two directions are at very different stages.

Direction one, shipping now: AI is the repair crew. ML calibration and hardware tune-up are already routine in real labs; learned decoders (like AlphaQubit-class networks) match or beat hand-written methods at reading the error alarms (that's the error-correction story) and are moving toward real-time deployment. This half of the loop is the mature one.

Direction two, real but early: what does quantum give back to AI? Here is the version that survives scrutiny. Quantum computers are terrible at almost everything your laptop is good at: spreadsheets, video, browsing. They are naturally good at exactly one strange thing: being nature. Molecules, materials, and probability patterns that would take a classical computer longer than the age of the universe to imitate simply happen on quantum hardware, because it runs on the same underlying rules. AI has a bottomless appetite for rare data of exactly that kind. So the trade is real, but it's about growing data no one else can grow, not running today's neural networks faster.

And the catch that kills most of the hype: getting ordinary, everyday data into a quantum computer is so slow that for normal datasets the speedup usually dies before it starts. The loop runs on nature's data, not your spreadsheet. Hover the diagram below to see what flows in each direction.

The loop, both directions

Click either arrow. Solid violet = AI helping quantum (mature). Dashed teal = quantum helping AI (nascent; the dashes are the point).

Four ways quantum could help AI, ranked by how much to trust them

The data-loading toll. To touch N classical numbers a quantum computer must first load them into amplitudes. Without large-scale QRAM (which nobody has built) that costs ~N operations. So an algorithm that's exponentially faster after loading still sits behind an O(N) toll bridge, and the total time is no better than classical. The algorithm is fast; the door is slow.

Two more dragons. Dequantization: in 2018 an 18-year-old, Ewin Tang, found a classical algorithm matching a flagship quantum recommendation-system speedup under comparable assumptions; the exponential separation evaporated, and the trick generalized to several others. Barren plateaus: for big random quantum circuits the training gradient's variance vanishes exponentially with qubit count, so the landscape goes flat and gradient descent gets no signal: bigger trains worse, the exact opposite of deep learning. The rule of thumb that survives all of this: trust quantum-for-AI claims about quantum data; audit everything about classical data.

The reverse loop, with receipts

HHL solves Ax=b in Õ(log N · s²κ²/ε) time for an s-sparse, well-conditioned A with an efficiently preparable |b⟩. The "fine print" (Aaronson, Nat. Phys. 2015) is where the exponential speedup usually leaks out: preparing an arbitrary classical b costs Ω(N) without QRAM; the condition number κ enters polynomially; and the output is the quantum state |x⟩, so reading out classical components costs another ~Ω(N) sampling overhead. Three separate O(N) tolls on a "log N" algorithm.

HHL's own engine is quantum phase estimation: it reads out A's eigenvalues, inverts them coherently, then uncomputes. It is the same subroutine derived from scratch on The Machinery, where it is also named as phase estimation's other real use besides factoring.

Dequantization. Tang (arXiv 1807.04271) gave a classical recommendation-system algorithm polylogarithmic in dimension under ℓ²-norm sample-and-query access, collapsing the claimed exponential separation; the technique generalized to kill several QML speedups in low-rank regimes. The lesson isn't "quantum lost"; it's that a speedup is defined against the best classical algorithm, which may not exist yet.

Barren plateaus. For random parameterized circuits forming approximate 2-designs, Var[∂_θC] decays exponentially in qubit count n (McClean et al., arXiv 1803.11173); global cost functions make it worse (Cerezo et al.). Trainability then requires structured or symmetry-constrained ansätze: you must smuggle physics back in to make the "learning" learn.

Where the reverse loop pays rent. Problems whose only quantum edge is a modest polynomial speedup (Grover-style quadratic gaps) die under fault-tolerance overheads; the cited analysis finds even quartic speedups only marginally practical on early fault-tolerant hardware (the Grover audit; Babbush et al., arXiv 2011.04149). The survivors are exponential-gap, low-data problems: Hamiltonian dynamics, chemistry and materials ground states (with careful instance selection), and, in the constructive results, classical ML gains provable power from data, which shrinks the room left for a quantum ML advantage (Huang et al., arXiv 2011.01938), while a rigorous exponential advantage does appear when a quantum machine learns coherently from quantum experiments (Huang et al., arXiv 2112.00778). Real advantage clusters around tasks whose data is itself quantum. Synthesis: the defensible near-term reverse loop is quantum devices as data factories, sampling and Hamiltonian simulation feeding classical AI, not QML on classical data.

The classical side of this page's own loop, for further reading. Everything above assumes familiarity with how classical ML works: gradient descent, cost landscapes, what "training" means. This page doesn't re-teach that; the standard references are Bishop, Pattern Recognition and Machine Learning (Springer, 2006); Murphy, Probabilistic Machine Learning: An Introduction (MIT Press, 2022) and its companion Advanced Topics (MIT Press, 2023); Goodfellow, Bengio & Courville, Deep Learning (MIT Press, 2016); and, for the free, course-form version, Stanford's CS229 (Machine Learning) and CS25 (Transformers United).

So what's the one-line version?

Quantum won't make your chatbot faster. It may, within the decade, hand AI a firehose of data about molecules, materials, and quantum systems that no classical instrument can generate, and that is a real way to make AI more powerful at the science that matters. Meanwhile the other direction of the loop is already load-bearing. Both halves running at once, in one machine, is the whole thesis of this site.

Go deeper