Google's Willow chip cultivates magic states with errors cut roughly 40-fold
Google Quantum AI cut errors in the magic states made on its Willow chip roughly 40-fold, using a technique called cultivation. The states still fail about once in 10,000 attempts, Scientific American reported, against one in a million or better for simulating molecules.
The Scientist · Science desk
Willow's magic states now fail roughly once in 10,000 attempts. Simulating molecules would need one error in a million to one in 100 million. Breaking RSA-2048 would need states that fail only once in 10 billion tries.
| Figure | Value | Note | Claim |
|---|---|---|---|
| Willow's magic states now, roughly | 10,000 attempts/error | 14 | |
| Needed to simulate molecules, at least | 1,000,000 tries/error | up to one error in 100 million | 13 |
| Needed to break RSA-2048 | 10,000,000,000 tries/error | 12 |
What happened
- Instead of asking whether the state was 0 or 1, which would collapse it, the team asked whether it was the right magic state, then spread passing states across more qubits.
- The experiment used 54 of Willow's 105 physical qubits.
- Google researchers showed in 2024 that cultivation could in theory cut the qubits needed for consequential programs to one-tenth; New Scientist says this experiment shows it works in practice.
- Isaac Kim of UC Davis says only 8 percent of Willow's computations produced high-quality magic states, and Kenneth Brown of Duke puts the discard rate at 92 percent.
Why it matters
- What it limits constraint Simulating molecules, the use Rosenfeld calls more attainable, needs states between 100 and 10,000 times cleaner than the team has now.
- What it costs cost At an 8 percent yield a kept state takes about 12.5 attempts on average, and each discarded attempt costs time.
- What's now possible capability Dominik Hangleiter of ETH Zurich called the work a possible "gamechanger" for building intermediately sized machines with hundreds of logical qubits.
Magic states are expensive, and the standard way of making them, distillation, shows how much. Emma Rosenfeld, lead author of a paper that Scientific American says appeared in Science [20], said of distillation: "You can have up to 95 percent of your quantum computer being spent just generating these high-fidelity magic states." [3] She told New Scientist the same bottleneck is costly: "It eats up a lot of qubits and a lot of time." [4]
The two outlets' numbers fit together. Four faulty states in 1,000 attempts, cut 40-fold, is one in 10,000 [22], which is where Scientific American says the team now stands [14]. Breaking RSA-2048 encryption would need states that fail only once in 10 billion tries [12], a million times lower than that [25]. Rosenfeld described molecule simulation as the more attainable use [13].
The 8 percent yield is the main limitation, and Rosenfeld addressed it. She said the share ought to increase as the team refines the technique and improves Willow's hardware, and that it is also affected by a very careful benchmarking technique that may not be used in future applications [8]. She did not give a yield without that benchmarking.
Google's result is not the first of its kind. Experiments on ion-based quantum computers have also lowered error rates through new ways of making magic states and error-correcting non-Clifford gates [16]. Willow's error rates were similar to those, but Rosenfeld said its superconducting qubits work more quickly, potentially reducing overall runtime [16].
The sources support the claim that cultivation works. Kenneth Brown of Duke, who was not involved in the study, said: "The Google experiment shows that cultivation works in practice." [9] Isaac Kim of the University of California, Davis, said: "It's an approach that can make fault-tolerant quantum computing substantially cheaper." [19]
The one-tenth qubit figure comes from the 2024 theory [5]. The sources do not report a measured qubit saving from this run, which occupied about half the chip [23]. Zhenyu Cai of Imperial College London called the experiment a significant step forward but said Willow's errors must fall further before it can tackle truly useful problems [15].
In our view the evidence shows a validated component of a fault-tolerant machine, and a smaller machine doing useful work with that component is still a projection. Rosenfeld said: "This is the first demonstration that the concepts work. We're not done." [18]
What to watch
- Whether the 8 percent yield rises once the careful benchmarking technique is dropped from the experiment.
- Whether the failure rate moves from about one in 10,000 toward the one in a million that molecule simulation needs.
- A measured qubit count for a full cultivation-based non-Clifford gate, compared with distillation on the same hardware.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence72
- Adoption
- Insufficient
- Hype gap+25
- Incentives30
- Confidence70
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Emma Rosenfeld and colleagues at Google Quantum AI used the Willow quantum computer to show that error correction can be made more efficient; the new technique ultimately reduced Willow's errors 40 times.
- [2]
Without cultivation, about four attempts in 1,000 produced a faulty magic state; with it, errors fell by roughly 40-fold.
- [3]
Rosenfeld said of distillation, the standard recipe for magic states: "You can have up to 95 percent of your quantum computer being spent just generating these high-fidelity magic states."
- [4]
Rosenfeld said creating magic states has been a bottleneck: "It eats up a lot of qubits and a lot of time."
- [5]
In 2024 Google Quantum AI researchers showed that magic state cultivation could in theory enable quantum computers to run consequential programs with one-tenth the number of qubits previously expected; New Scientist says the new experiment proves cultivation works in practice.
- [6]
Only 8 per cent of Willow's computations produced high-quality magic states that could be put into further use.
- [7]
Cultivation relies on a quality assurance step that rejects low-quality states, discarding 92 percent of attempts.
- [8]
Rosenfeld says the 8 percent share ought to increase as the team refines the technique and improves Willow's hardware; it is also affected by a very careful benchmarking technique the team used, which may not be used in future applications.
- [9]
Kenneth Brown, director of the Duke Quantum Center, who was not involved in the study, said: "The Google experiment shows that cultivation works in practice."
- [10]
The team encoded a rough magic state in a few qubits and, instead of asking whether it was 0 or 1, which would collapse it, asked whether it was the right magic state; states that pass are spread across more qubits.
- [11]
The experiment used 54 physical qubits out of Willow's 105.
- [12]
Using a quantum computer to break RSA-2048 encryption would require magic states that fail only once in 10 billion tries.
- [13]
Simulating molecules, a use Rosenfeld says is more attainable, would need magic states with only one error in a million to one in 100 million.
- [14]
Right now the team stands at roughly one error in 10,000 attempts.
- [15]
Zhenyu Cai at Imperial College London says the experiment is a significant step forward but Willow's errors will need to be reduced further before it can tackle truly useful problems.
- [16]
Advances in creating magic states and error-correcting non-Clifford gates have also been achieved in experiments on ion-based quantum computers; Willow's superconducting qubits achieved similar error rates, but Rosenfeld says they work more quickly, potentially reducing overall runtime.
- [17]
Dominik Hangleiter at ETH Zurich says the work demonstrates one of the building blocks of universal quantum computing and could be a "gamechanger" for building intermediately sized quantum computers with hundreds of logical qubits.
- [18]
Rosenfeld said: "This is the first demonstration that the concepts work. We're not done."
- [19]
Isaac Kim at the University of California, Davis, said: "It's an approach that can make fault-tolerant quantum computing substantially cheaper."
- [20]
The study, with Emma Rosenfeld of Google Quantum AI as lead author, was published in Science.
- [21]
At an 8 percent yield, a kept magic state takes about 12.5 attempts on average.
- [22]
Four faulty states in 1,000 attempts cut 40-fold is one in 10,000, matching the team's current rate.
- [23]
The experiment used about 51 percent of Willow's qubits.
- [24]
Molecule simulation needs magic states between 100 and 10,000 times cleaner than the current one-in-10,000 rate.
- [25]
Breaking RSA-2048 needs magic states about a million times cleaner than the current rate.
Get this ledger inside ChatGPT. Ask about any of our stories and the answer brings its figure, each claim’s verdict and the sources. Free, no account.
Sources
2 independent publishers whose own reporting we read for this story.
- newscientist.comGoogle showcases way to fix quantum computer errors more efficiently
1 article · October 15, 2026
- scientificamerican.comGoogle’s quantum Willow chip may be able to grow ‘magic states’
1 article · October 15, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.