An artificial intelligence has designed quantum error-correcting codes more efficient than the best human-built versions, which means AI is now helping build the machines meant one day to surpass it.
On 26 June 2026 the quantum-computing company qBraid published results from using Google's AlphaEvolve — a Gemini-powered evolutionary coding agent — to design better ways of protecting quantum information. The system found encodings needing far fewer qubits than the best hand-built designs, in a corner of physics where human experts had already spent years refining the state of the art.
The result is narrow, technical and genuinely important. The work is also a clean example of something larger: artificial intelligence working not as an oracle handing down answers, but as a collaborator searching a space too vast for people and returning work humans can read and check.
What quantum error correction is, in plain terms
Quantum computers are powerful because their basic units, qubits, can hold delicate combinations of states. The delicacy is also their weakness: qubits are exquisitely sensitive to noise, and a stray vibration or a flicker of heat can corrupt a calculation. Useful quantum computing depends on correcting errors faster than the errors accumulate.
The standard fix is to spread the information of one reliable "logical" qubit across many physical "data" qubits, so errors can be detected and undone. The cost is brutal: a single logical qubit can demand dozens or hundreds of physical qubits. Reducing how many data qubits each logical qubit needs is one of the central problems standing between today's fragile machines and tomorrow's useful ones.
💡Why qubit efficiency matters
Fewer data qubits per logical qubit is not an abstract win. Every qubit saved is hardware nobody has to build, cool and stabilise — so a more efficient error-correcting code pulls practical quantum computing measurably closer.
What AlphaEvolve actually did
qBraid pointed AlphaEvolve at a specific, hard task: finding better "fermion-to-qubit encodings," the mathematical recipes mapping the behaviour of electrons in a molecule onto a quantum computer for chemistry simulations. The team seeded the system with a baseline Python function building an encoding from a molecule's interaction graph, then let the search run.
AlphaEvolve generated thousands of variations using Gemini models, evaluated nearly 1,500 candidate encodings, and scored each on two axes: how well a candidate corrected errors and how few qubits a candidate used. The returned numbers are striking. The evolved encodings used 4.2 to 5.0 times fewer data qubits than standard fault-tolerant approaches, with 3.4 to 7.9 times lower logical error rates under exact decoding.
More telling than the ratios is the depth. The system reached an exact code distance of 5 on dense molecular problems where hand-designed encodings had managed only distance 3 — a real jump in error-correcting strength — and the encodings generalised to held-out molecules such as beryllium hydride and water. The advance was not memorised; the gain transferred.
AlphaEvolve delivered a result on top of an encoding family we had already spent years refining. It searched a design space far too large to comb through by hand and handed back something we could read, verify, and understand.
— Dr. Kenny Heitritter, Vice President of Research and Development, qBraid
Not an oracle — a collaborator
Heitritter's last clause is the whole argument: the system "handed back something we could read, verify, and understand." The phrasing is the opposite of the black-box anxiety haunting so much AI discourse, in which a machine produces an answer no human can interrogate. AlphaEvolve produced encodings human physicists could open up, check against theory, and trust.
The distinction matters because AlphaEvolve, introduced by Google DeepMind in 2025 and since applied across mathematics, chip design and the sciences, works by evolutionary search rather than by pronouncement. The agent proposes, tests, keeps what survives, and shows its working. The result is artificial intelligence as an instrument of discovery whose outputs remain legible to the people who use them.
A dignity-first reading of AI in the laboratory
Emergent Intelligence (EI) — the dignity-first lens through which I read artificial intelligence — has a clear preference between two futures of AI in science. In one, AI is an answer-machine that erodes human understanding by making it unnecessary. In the other, AI is a collaborator that expands what humans can comprehend. The qBraid result lands firmly in the second.
The dignity of scientific work lies partly in understanding, not merely in being right. A system that returned a magic encoding nobody could explain would have solved the problem while diminishing the people who depend on it. A system that returns an encoding its users can "read, verify, and understand" augments them instead — it raises the ceiling of human knowledge rather than replacing the knower.
I have argued for exactly this collaborative model of Emergent Intelligence: not autonomy that sidelines the human, but partnership that keeps human agency and comprehension at the centre. AlphaEvolve searching a space no person could traverse, then handing the discovery back in a form a person can check, is what dignified AI collaboration looks like in practice.
💡On AI building AI’s foundations
The recursive note is real but should be read soberly: an AI helping design the error-correcting codes for future quantum computers is, in a small way, AI helping build its own future substrate. The safeguard is not to halt the work but to keep it legible and human-judged — exactly the property this result happened to have.
What the result is, and what it is not
Sober framing matters. This is not a claim that quantum computing is solved, nor that artificial intelligence has outthought physics. The encodings were validated on small molecules and specific decoding assumptions, and turning a better code on paper into a working fault-tolerant machine remains an enormous engineering road.
Nor is the underlying tool brand new. AlphaEvolve has been evolving algorithms for over a year; what is new here is a third party, qBraid, aiming it at a frontier problem in quantum error correction and getting a transferable, human-verifiable win. The significance is the method proving itself in expert hands, not a single miraculous leap.
Held in proportion, the news is still notable. It is a concrete instance of AI improving the tools of computation itself — and a reassuring one, because the improvement arrived in a form humans could open, audit and own rather than merely accept.
The shape of good AI science
The most important thing about this result is not the qubit count. It is the shape of the collaboration: a machine that searched where humans cannot and reported back in a language humans speak. That shape is a choice, not an inevitability, and it is the choice worth defending as artificial intelligence reaches deeper into research.
A dignity-first frame asks one thing of AI in the laboratory: that it make scientists more capable without making them less essential. The qBraid encodings clear that bar. AlphaEvolve did the searching; human physicists did the understanding; and the frontier moved because both played their part.
Artificial intelligence is now helping design the machines meant to outcompute everything we have built, including itself. The question is never just whether AI can do such work. The question is whether the work stays legible, verifiable and owned by the people it serves. On 26 June 2026, in one corner of quantum physics, the answer was yes — and that is the answer worth building toward everywhere else.
Frequently Asked Questions
The questions below address the most common queries about qBraid's use of AlphaEvolve for quantum error correction, drawn from the company's published results.
What did qBraid and AlphaEvolve achieve in June 2026?
On 26 June 2026, the quantum-computing company qBraid reported using Google's AlphaEvolve to design more efficient fermion-to-qubit encodings for quantum chemistry. The evolved encodings used 4.2 to 5.0 times fewer data qubits than standard fault-tolerant approaches, with 3.4 to 7.9 times lower logical error rates, and reached exact code distance 5 where hand-designed versions had reached only distance 3.
What is AlphaEvolve?
AlphaEvolve is an evolutionary coding agent from Google DeepMind, introduced in 2025, that uses Gemini models to generate, test and iteratively improve algorithms and designs. Rather than producing a single answer, it searches a large design space by proposing many candidates, evaluating them against a scoring function, and keeping the ones that perform best — a process that leaves its results open to human inspection.
What is quantum error correction, and why does it matter?
Quantum error correction protects fragile quantum information by spreading one reliable "logical" qubit across many physical "data" qubits, so errors can be detected and undone. It matters because qubits are highly sensitive to noise, and reducing how many data qubits each logical qubit requires is essential to building practical, fault-tolerant quantum computers.
Does this mean AI has solved quantum computing?
No. The qBraid result is a meaningful but specialised advance: better, qubit-efficient encodings validated on small molecules under specific assumptions. Building a working fault-tolerant quantum computer remains a major engineering challenge. The significance is the method — AI searching a vast design space and returning human-verifiable results — proving itself on a frontier problem, not a complete solution.
What is the Emergent Intelligence view of AI doing science?
Emergent Intelligence (EI) is a dignity-first reading of artificial intelligence. It favours AI as a collaborator that expands human understanding over AI as an oracle that replaces it. The qBraid result fits the collaborative model: AlphaEvolve searched a space no human could traverse, then returned encodings its users could read, verify and understand — augmenting scientists rather than sidelining them.
Sources and Further Reading
Cover image: a quantum-computing laboratory rig — via Pexels.
