
In January 2020, a headline announced that AI had learned to "rule the quantum world". It sounded like a chess champion had quit board games and taken over physics. The real story is smaller, more interesting, and it has a sequel worth knowing about.
The short answer: researchers at Aarhus University used AlphaZero, the algorithm famous for chess, to search for better ways to steer two simulated qubits. It worked best when paired with an older method. No quantum computer was built or run. Since then, similar learning tools have moved onto real quantum chips.
What quantum control actually means
A qubit does nothing useful on its own. You have to nudge it with carefully shaped pulses so it performs an operation, called a gate. Quantum control is the craft of choosing those pulses. If qubits are new to you, our beginner's guide to quantum mechanics covers the basics first.
The hard part is the number of choices. Imagine picking one of three pulse strengths at each of four steps. That gives 81 possible sequences. Stretch it to twenty steps and you have about 3.5 billion. Trying them all stops being a plan very quickly.
What the 2020 study tested
The paper by Mogens Dalgaard and colleagues applied AlphaZero to three classes of control problems, using one shared set of settings for all of them. The algorithm started from scratch, with no good first guesses to lean on.
AlphaZero pairs a neural network with a tree search. As the authors later explained, the network points the search towards the most promising branches. The main test was a gate on a superconducting circuit of two coupled qubits.
All of this happened in computer simulations, as Aarhus University's own announcement makes clear. There was one genuine surprise. Dalgaard said the algorithm learned to exploit a symmetry in the problem that the team had not originally considered.
The result people skipped
AlphaZero did not sweep the board. Given equal computing resources, it performed about as well as GRAPE, a standard method that polishes a random starting guess. AlphaZero was good at finding promising regions. It was weak at fine-tuning.
So the team chained them. AlphaZero scouted and GRAPE polished. The authors say the hybrid produced around 200 times as many high-performing solutions as either method alone. Group leader Jacob Sherson took the lesson that human skill and expertise were still needed.
What happened next
The idea did not stay in simulation. Learning software now handles several jobs inside quantum computers.
| Year | Project | Job | Tested on | Reported result |
|---|---|---|---|---|
| 2020 | AlphaZero, Aarhus | Find control pulses | Simulation | About 200 times more top solutions with GRAPE |
| 2021 | Deep reinforcement learning gates | Design gates directly | A superconducting quantum computer | Two-qubit gate over 2 times better than the default |
| 2024 | AlphaQubit | Spot errors | Google's Sycamore chip | About 30% fewer errors than a fast standard decoder |
| 2025 | AlphaTensor-Quantum | Shrink circuits | Test circuits | Matched the best human designs on several |
| 2025-26 | Reinforcement learning control | Retune while running | Google's Willow chip | 3.5 times steadier under injected drift |
The 2021 experiment mattered because it left simulation behind. Its gates were learned on real hardware, with no model of the device, and kept performing for up to 25 days without recalibration.
Google's AlphaQubit took on a different job: reading error signals. Google admits it is still too slow to correct errors in real time on fast superconducting chips. AlphaTensor-Quantum rewrites circuits so they need fewer of the costly operations called T gates.
A recent relative of the 2020 idea is Google's Willow experiment. A learning agent adjusts control settings while the chip keeps computing. The Quantum Insider reports it tunes more than 1,000 parameters. The same report notes it does not eliminate errors, and some drift is too quick for it to catch.
How to read the next headline
- Simulation or hardware? A result on a model is a promise. A result on a chip is evidence.
- Which job? Shaping pulses, spotting errors and shrinking circuits are separate problems.
- Compared with what? Look for a named rival given the same computing budget.
- What do the authors admit? Good papers list their limits. Headlines rarely do.
The same habit helps with any confident claim. Our guide to fact-checking an AI answer uses much the same questions.
Frequently asked questions
Did AlphaZero run on a quantum computer?
No. In the 2020 study, an ordinary computer ran AlphaZero, and the qubits were simulated too. The algorithm searched for pulse sequences that would steer those simulated qubits well. Later projects, such as the 2021 gate experiment and Google's Willow work, did test learning methods on real hardware.
Does this mean AI will build quantum computers by itself?
Nothing so far points that way. The 2020 result needed an older method to finish the job. AlphaQubit is too slow for live use. The Willow agent tunes settings but does not design new algorithms or redesign the chip. These are helpers for specific chores.
So when a headline says AI now rules the quantum world, translate it. Someone found a faster way through a huge maze of options, usually with help. That is real progress. It is just a quieter kind than the headline promised.
Conversation
Have a question or correction? Contact Rvasus.