Google Willow and "magic states"

On October 15, the Google Quantum AI team wrote on Google’s blog that a method called magic state cultivation has worked in practice for the first time, on the Google Willow processor. The results were published in the journal Science.

What exactly did Google Willow show?

According to the post by Emma Rosenfeld and Craig Gidney from Google Quantum AI, the method delivered:

1. around 40x fewer errors than magic states prepared without cultivation,
2. T states (one kind of magic state) with 99.99% logical fidelity, which the team calls a record,
3. first steps toward Google’s target architecture: a state was moved from a distance-3 color code to a distance-5 surface code.

According to Decrypt, the experiment used at most 54 of Willow’s 105 qubits.

So what are these “magic states”?

An error-corrected quantum computer handles some of the easy operations fairly cheaply, but universal computation also needs the “hard” ones. One way to get them is to prepare special helper states in advance, and those are the magic states. The name comes from the fact that they are exactly what lets a computation go beyond what an ordinary computer can simulate.

The catch is that they are expensive and single-use: each magic state can be consumed only once. According to estimates cited by Google, producing them could eat up as much as 95% of a quantum computer’s capacity. The team writes that this bottleneck was predicted back in 2004, and people have been looking for a way to widen it for about 20 years.

How does magic state cultivation work?

The theory was described in 2024 by Craig Gidney, Noah Shutty and Cody Jones in the paper “Magic state cultivation: growing T states as cheap as CNOT gates”. It works like a filter:

1. a noisy state is injected into a small patch of the chip,
2. a series of coordinated measurements checks it for errors and discards failed attempts,
3. the state that survives is “cultivated”, meaning grown into a very clean form that can be moved into a larger code.

Interestingly, the road from theory to hardware wasn’t smooth. According to Google, the first data from Emma Rosenfeld’s experiment “seemed to defy the laws of physics”. After months of work, it turned out the culprit was a wrong assumption in a measurement method the team routinely used. ScienceAlert reports that the problem was solved by a new technique called “kickback tomography”.

Is there a catch?

The filter has a cost. In the strictest runs, only about 8% of attempts survived and the rest were discarded. Moving the state into a larger code also added new errors, which the team still has to deal with.

Google itself notes that the method has to be shown at a larger scale before it becomes useful for practical computation. Decrypt points out that breaking RSA-2048 would require magic states with an error rate around 1 in 10 billion, roughly a million times cleaner than in this experiment. So “Q-day” is still a long way off.

What does this change for cryptography?

Not much, for now. Alex Pruden of Project Eleven told The Block that the result doesn’t necessarily move Q-day forecasts, because earlier resource estimates already assumed this technique. In his view, the experiment rather confirms those assumptions, so current forecasts become more credible. Nic Carter of Castle Island Ventures said there is “no immediate threat”.

This is another big result from the quantum world in recent days. I recently wrote about how Claude designed quantum circuits with up to 17x fewer gates. Better circuits and cheaper magic states are two sides of the same problem: how to fit useful computation onto limited hardware.

FAQ

What is Google Willow?

It’s Google Quantum AI’s quantum processor. The magic state cultivation experiment used at most 54 of its 105 qubits.

Can Google Willow break encryption already?

No. Breaking RSA-2048 would require magic states roughly a million times cleaner than in this experiment, and a much bigger machine.

Where can I read the paper on magic states?

The experimental results were published in Science, and the 2024 theory is on arXiv.