Search Ideas

Drawing samples from random quantum circuits, Willow did in under five minutes what would take one of today’s fastest supercomputers 10 septillion years. That suggests quantum computation happens in many parallel universes, as David Deutsch first predicted.

Most of Kalai’s post re-analyzes Google’s 2019 experiment, which was long ago superseded anyway. Since then, several other experiments, some by other groups, such as USTC’s and Quantinuum’s, challenge his position, so it no longer rests on Google’s claims alone.

We haven’t studied these particular claims yet, and our specific points of contention bear on Google’s newer supremacy experiments, but not directly on the quantum error-correction experiment.

My general conclusion applies to these claims too: Google Quantum AI’s claims, published ones included, should be approached with caution, particularly extraordinary ones. Its 2019 claim of quantum supremacy overestimated the classical running time by ten orders of magnitude, and its fidelities matched its predictions too well to be plausible. Such claims may stem from significant methodological errors, and reflect the researchers’ expectations more than objective scientific reality.

On our Willow chip, logical qubits spread over squares of 3 by 3, 5 by 5, and 7 by 7 qubits failed less often the bigger they were: each step up in size cut the failures by more than half. The errors are far enough below the threshold for error correction to work, and the biggest logical qubit outlived the best physical qubit in it.

The threshold theorem assumes errors strike independently, but in real devices they don’t: noise is correlated with the very states being computed. So no quantum device will get far enough below the threshold for error correction to work.

Originating ideas is something brains do, and brains are physical objects. Whatever the laws of physics let a physical object do, a universal computer can be programmed to do. So with the right program, a computer could originate ideas just as people do. What’s missing is the program, not the machine.

Quantum errors can be corrected faster than they arise, as long as each step errs less often than a certain threshold. Nest codes inside codes, and a quantum computation can run as long as needed, as accurately as wanted.

Error correction can’t save a quantum computer in practice. It needs many more qubits and steps, each of which errs too, and the large entangled states it relies on are extremely fragile, so it would add errors faster than it removes them.

Quantum errors can be corrected without looking at the data. Spread one qubit over nine, tied together so that checking how they relate to each other shows whether one of them went wrong, and how to undo it, without revealing what the qubit holds.

A quantum computer could never run long enough. Its quantities vary continuously, like those of an old analog computer, so small errors pile up and can’t be corrected the way a classical computer corrects a slipped bit: you can’t even look at a quantum state to check it without disturbing it.

A quantum computer could factor numbers hundreds of digits long in hours, by finding how often the remainders of a number’s powers repeat. That would break the codes that protect most of the internet’s secrets, which rely on factoring being too hard.

Every physical system that can be built can be perfectly simulated by a universal computer working by finite means – but only by a quantum one. A universal quantum computer can do whatever any physical machine can do, and some things far faster than any classical computer could.

A classical computer can’t simulate quantum physics efficiently: the numbers it needs double with every particle added, so a few hundred particles would take more numbers than there are atoms in the universe. To simulate nature, build a computer of quantum parts that behaves as nature does.

A computer can run by the laws of quantum mechanics alone. A Turing machine can be built of quantum parts that carry it through its steps one after another, as an isolated system, without losing energy.

Then the test needs tightening: put the competitors in a telepathy-proof room.

The argument from extra-sensory perception: the statistical evidence for telepathy is overwhelming. In the imitation game, a telepathic person could name the card in the interrogator’s hand far more often than a machine guessing at random, and so be told apart from it.

That confuses rules of conduct, which we can know and act on, with laws of behavior, the laws of nature that govern our bodies. We can be sure there are no complete rules of conduct, but never that there are no complete laws of behavior: we can never say we’ve searched enough. Even a small program of mine on the Manchester computer can’t be predicted from its replies.

The argument from informality of behavior: no set of rules can say what a person should do in every conceivable circumstance. If people followed a definite set of rules of conduct, they’d be no better than machines. There are no such rules, so people can’t be machines.

A discrete-state machine is different from a continuous one, but in the imitation game, the interrogator can’t take advantage of the difference. A digital computer couldn’t predict exactly what a continuous machine, like a differential analyzer, would answer, but it could give the right sort of answer.

The argument from continuity in the nervous system: the nervous system isn’t a discrete-state machine. A tiny change in an impulse reaching a neuron can make a large difference to the impulse going out. So a discrete-state machine can’t be expected to mimic it.

No support is usually offered for these claims. They generalize from the machines people have seen, which are ugly, each built for a narrow purpose, and able to do little else. Many of those limits come from their tiny memories, and computers’ memories are growing.

Arguments from various disabilities: machines may do all you say, but you’ll never make one be kind, have a sense of humor, tell right from wrong, make mistakes, fall in love, enjoy strawberries and cream, learn from experience, or do something really new.

Then the only way to be sure that a machine thinks is to be the machine, and the only way to know that a man thinks is to be that man. That’s solipsism. Instead of arguing over it, we have the polite convention that everyone thinks. And if a machine could defend its sonnet’s wording under questioning as well as a person can, I don’t think anyone would call that an easy contrivance.

The argument from consciousness: a machine would have to feel, not just act. As the English neurosurgeon Geoffrey Jefferson put it in 1949: “Not until a machine can write a sonnet or compose a concerto because of thoughts and emotions felt, and not by the chance fall of symbols, could we agree that machine equals brain.”