Loading lesson…
Quantum AI / Quantum meets ML
A clear-eyed status report on quantum AI.
Reviewed by Yuvaraj
Strip away the headlines and one question decides whether a quantum result matters: did a machine do something useful that classical computers cannot, or did it do something contrived that merely happens to be hard to simulate? Almost every genuine milestone so far is the second kind. This lesson is a hype filter. It explains the noisy machines we actually have, why noise caps what they can compute, what error correction would take to fix that, and what the famous "supremacy" experiments really demonstrated. The goal is that you can read any quantum announcement and place it correctly as established science, emerging engineering, or speculation.
John Preskill coined the term NISQ, Noisy Intermediate-Scale Quantum, in 2018 to name exactly where the field sits. "Intermediate-scale" means tens to a few hundred physical qubits: enough that classical simulation is strained, far too few for the textbook algorithms. "Noisy" means there is no full error correction, so errors accumulate as a circuit runs. Every quantum processor available today, across superconducting, trapped-ion, and neutral-atom platforms, is a NISQ device. This is why claims that current machines can break encryption are simply wrong: Shor's algorithm against RSA-2048 needs fault-tolerant hardware with millions of physical qubits, and no NISQ device is close.
Two physical facts limit these machines. First, decoherence: a qubit slowly leaks its quantum information to the environment, characterized by a relaxation time (energy decay, often written ) and a dephasing time (loss of phase coherence, ). Beyond these timescales the state is scrambled. Second, : every operation is imperfect, with two-qubit gate error rates on the best hardware today around a few tenths of a percent to about one percent, plus readout errors on top.
Ask about this lesson, or about anything in AI. Answers cite the lessons they draw on.
Finished this lesson?
Mark it complete to earn XP, keep your streak, and schedule a review.
The consequence is a hard ceiling on circuit depth. Errors compound multiplicatively: a circuit with gates each succeeding with probability finishes intact with probability roughly . Put in numbers, a two-qubit error rate of across gates gives , under one percent of runs are error-free. You run out of coherence and accuracy long before a deep algorithm finishes, which is the central practical constraint of the NISQ era.
The escape route is quantum error correction (QEC): encode one logical qubit across many physical qubits so that errors can be detected and undone without ever directly measuring, and thus collapsing, the protected information. You measure only syndromes, extra checks that reveal what error occurred while leaving the logical state intact.
The leading scheme is the surface code: physical qubits on a 2D grid with only nearest-neighbor interactions, a relatively forgiving error threshold near one percent, and a code distance that corrects up to errors. The threshold theorem is the key result: if the physical error rate sits below the threshold, increasing suppresses the logical error rate exponentially, so arbitrarily reliable computation is possible in principle.
The price is overhead. A single useful logical qubit may need hundreds to thousands of physical qubits, and full estimates are stark: factoring a 2048-bit RSA key has been estimated to require on the order of 20 million noisy physical qubits running for hours (Gidney and Ekerå). Recent experiments are real but early, in 2024 Google reported a surface-code memory operating below threshold, with the logical error rate dropping as code distance grew from 3 to 5 to 7. That is a genuine milestone for the principle, still at the scale of a single logical qubit.
For a curious beginner
Protect information by spreading it over many copies and constantly checking for damage, like a document backed up across many drives with error-checking. The subtlety unique to quantum: you must detect and repair errors without reading the data itself, since reading would destroy the superposition. It only works if your hardware is already clean enough, below a noise threshold, so that adding redundancy helps rather than adds more error.
How it is actually used
Encode one logical qubit in a grid of physical qubits (the surface code). Repeatedly measure syndrome qubits, feed the results to a fast classical decoder, and apply corrections. A distance- code corrects up to errors, and you buy reliability by enlarging , paying in qubit count and decoding speed.
The underlying mechanism
The threshold theorem states that if the physical error rate is below a threshold , the logical error rate falls as . When the base is less than one, so decays exponentially in the code distance , arbitrarily reliable computation at polynomial overhead. When , more redundancy makes things worse.
In 2019 Google's 53-qubit Sycamore processor performed random circuit sampling: run a random quantum circuit and sample its output distribution. Google reported the task took about 200 seconds versus an estimated 10,000 years on the leading classical supercomputer, and called it "quantum supremacy", the first computation beyond practical classical reach.
Read it carefully. The task was deliberately contrived to be hard for classical computers and easy for the quantum device; sampling random-circuit outputs has no known use. It demonstrated that a quantum processor can do something a classical one cannot readily reproduce, a real statement about computational complexity, but not that quantum computers now solve useful problems faster. And the classical bar moved: better simulation algorithms and larger clusters cut the classical estimate dramatically (IBM argued days, not millennia; later tensor-network methods narrowed it further). "Supremacy" or "advantage" demonstrations are scientific milestones about hardness, not evidence of practical usefulness, keep the two ideas apart.
| Claim | Status | Notes |
|---|---|---|
| Superposition, entanglement, and unitary gates are real physics | Established | Confirmed by decades of experiment |
| NISQ processors with tens to hundreds of physical qubits exist and run circuits | Established | Available across several hardware platforms |
| Random circuit sampling can be classically hard (a complexity milestone) | Established, with caveats | Contrived task; classical estimates have shrunk over time |
| Quantum error correction works in principle; small below-threshold demos exist | Established principle, emerging engineering | Still around a single logical qubit |
| Practical quantum advantage for ML, optimization, or chemistry of real value | Emerging, not yet demonstrated | No useful task beats strong classical methods today |
| Fault-tolerant machines breaking RSA-2048 | Speculative (future) | Needs on the order of millions of physical qubits |
| Near-term NISQ machines solving valuable real-world problems better than classical | Speculative, unproven | Active research, no clear win yet |
| A qubit "tries all answers at once," giving free exponential parallelism | Myth (false) | Measurement returns one outcome; advantage comes from interference |
Useful fault-tolerant quantum computing is plausibly years to decades away, and honest roadmaps say so. Near-term value is mostly scientific: studying quantum systems, advancing error-correction engineering, and developing algorithms for the machines to come. One practical exception is worth noting: migrating to post-quantum cryptography is prudent now, even though the threat is in the future, because encrypted data captured today could be decrypted later once a capable machine exists, the "harvest now, decrypt later" risk, which is why standards bodies have already finalized post-quantum schemes. That is a rare case where a still-speculative capability justifies concrete action today.
Common mistakes