Google shows magic state cultivation works on Willow, cutting errors about 40-fold

In plain words

Google says it has shown, on a real quantum chip, a way to make a scarce ingredient that future quantum computers need. The ingredient, called a magic state, is hard to produce without errors, and making it could eat up most of a machine's capacity. Google reports about 40 times fewer errors with its method than without it. The authors say the work still has to be repeated at much larger scale. Nothing here changes what ordinary people should do today.

What happened

Google Quantum AI says it has demonstrated "magic state cultivation" on its Willow processor, and that the work was published in Science. ¹ The company reports about a 40-fold reduction in errors compared with states prepared without cultivation. ¹ It also reports T-states with a logical fidelity of 99.99%. ¹ Fidelity here measures how closely a protected quantum state matches its intended form. ²

The problem dates back to 2004, according to Google. A universal quantum computer needs a particular set of operations, and magic states help supply them. ¹ They are consumed when used, so a long calculation must keep producing replacements. ² Google says some estimates put the share of a machine's capacity devoted to making them as high as 95%. ¹ The Quantum Insider points out that this figure comes from resource estimates, "rather than a measured allocation on Willow." ²

Google researchers Craig Gidney, Noah Shutty and Cody Jones developed the cultivation idea in 2024, building on earlier literature. ¹ It works like a filter: a noisy state is injected into a small region of the chip, coordinated measurements check for errors, and flawed attempts are removed. ¹ The experiments used a Willow processor with 105 superconducting qubits in a square lattice, according to the paper as described by The Quantum Insider. ²

The team also used artificial intelligence in the calibration. The researchers used reinforcement learning, a form of machine learning that adjusts controls based on feedback, as a final calibration step. ² For the experiment with an error of roughly one part in 10,000, training ran for about four hours just before data collection. ² The researchers report that this improved both the logical error probability and the share of retained attempts by about an order of magnitude. ²

Measuring the result was hard. Google says Emma Rosenfeld saw data that seemed to defy the laws of physics, and traced it to an incorrect assumption in the standard measurement technique, with noise from equipment drift and temperature variation interfering. ¹ The team added extra control and invented a new benchmark that uses cultivation to measure itself. ¹ The paper calls this kickback tomography and describes it as measuring a lower bound on T-state fidelity. ²

Google also moved cultivated states from a distance-3 color code to a distance-5 surface code, which it calls grafting. ¹ The supplement says the distance-5 experiment involved 54 qubits. ² After grafting, the logical error rate was about seven times higher than simulations predicted, and the researchers name leakage as the likely main cause. ²

What is confirmed and what is not

Confirmed by the sources, in the sense that Google and the outlet report the same figures from the paper: the Science publication, the roughly 40-fold error reduction against uncultivated states, the 99.99% fidelity, the use of reinforcement learning in calibration, and the grafting experiment. ¹²

Only claimed, or qualified by the sources: the 40-fold figure is specific to this state-preparation task, not an improvement in the processor overall. ² The reported quality applies to accepted attempts after filtering, so the share of attempts kept matters for eventual cost. ² The 95% capacity figure is an estimate. ²

Unknown: whether cultivation holds up at larger scale, what it would cost in total hardware and time for a useful computation, and how states can be stored and used while they wait. Google says it still has to demonstrate the method at larger scale. ¹ The Quantum Insider says the work does not necessarily establish that the resource problem is solved at a practical scale. ² This article rests on the maker's own blog post and one trade outlet's reading of the paper. We have not seen the paper itself.

Analysis

Technology and industry

The mechanism is about overhead. If a large share of a future machine must be given over to producing magic states, the machine needs many more physical qubits to do useful work. A cheaper way to make them lowers that requirement. The sources support this only as a possibility: Google says the idea could substantially reduce the size of the computer needed for practical tasks, and The Quantum Insider calls the result an incremental advance with potentially lower hardware demands. ¹² we are fairly confident the result is a real engineering step, because it is reported on hardware and in a peer-reviewed journal. We are much less sure about the size of the saving in a full machine, since no source gives one.

The grafting result is a useful warning. The step that connects cultivation to the wider architecture performed about seven times worse than simulation predicted. ² If the gap comes from leakage, as the researchers suspect, then better calibration and leakage removal could close it. If other error sources are involved, the full pipeline may be harder to build than the isolated cultivation result suggests. The strongest reason this view could be wrong is that a single sub-step is being weighed against the whole system. Evidence that would change it: grafting results at larger code distances that match simulation.

AI in science

The reinforcement learning detail is modest but real. The AI was not the discovery. It was a calibration tool that tuned controls to lower the rate at which the circuit detected errors, and it improved results by about an order of magnitude on two measures. ² The step took about four hours before data collection. ² This suggests that machine-learned control may become a routine part of operating quantum hardware, though one experiment cannot show that. A sceptic could reasonably say that hand-tuned calibration might have reached similar numbers with more effort, and the sources do not test that.

Security

Quantum computers matter to security because a large, error-corrected machine could threaten some current encryption. The sources do not discuss this, so any link is general knowledge rather than a sourced claim. If cultivation lowers the cost of building such machines, then timelines for that threat could shorten at the margin. But the sources describe a small-scale demonstration, and they say larger tests and better storage methods are still needed. ² Readers following how blockchain groups are defining quantum readiness can see that debate in our earlier signal. Nothing in this result gives a date, and we would not infer one.

Measurement and trust in results

One under-noticed point is the measurement problem. The team found that its standard technique could produce impossible estimates, and that the same flaw could overstate fidelity without an obvious sign. ² That matters for anyone comparing quantum results across labs. If other groups use similar tomography with filtered data, some published fidelities may be less reliable than they look. The sources do not say this has happened elsewhere, so it is a question rather than a finding. The researchers also note that possible error mechanisms at much lower noise levels remain to be investigated. ²

Financial markets and daily life

The sources make no claims about markets, products or consumers, and we draw none. Practical, everyday effects would depend on machines that do not yet exist at useful scale.

What can be done

For ordinary readers, no action follows. The result is a laboratory step, and the sources give no deployment timeline.

For specialists and organisations following quantum computing, sensible options are to read the paper and supplement for the post-selection rate alongside the fidelity, since the cost depends on both. ² Teams doing similar tomography could check whether their own measurements combine differently filtered data. Security planners who already track quantum risk can note this as one input among many, without changing plans on this evidence alone. For Google and other labs, the sources point to better calibration, leakage removal, simpler circuits and larger codes as next steps. ²

Questions readers ask

What is magic state cultivation?

It is a filtering method for making high-quality magic states, which quantum computers need for certain operations. A noisy state is placed on a small part of the chip, checked by coordinated measurements, and discarded if errors are found. ¹ Google says it works on its Willow processor. ¹

Does this mean useful quantum computers are here?

No. Google says it still needs to demonstrate the method at larger scale, and that technical challenges remain. ¹ The Quantum Insider calls it an important but incremental advance. ²

Did AI play a part in the result?

Yes, in calibration. The researchers used reinforcement learning as a final calibration step to reduce the rate at which the circuit detected errors. ² They report it improved the logical error probability and the retained share of attempts by about an order of magnitude. ²

Is the 40-fold improvement for the whole chip?

No. It compares cultivated states with states prepared without cultivation, and applies to this state-preparation task rather than the processor as a whole. ²