China Uses Quantum Computer to Fine-Tune AI for the First Time

Chinese scientists have become the first in the world to use a quantum computer to fine-tune an artificial intelligence system — a large language model with one billion parameters. It was the first practical use of a quantum platform. The computer used in the work was Origin’s Wukong, a computer based on 72 superconducting qubits.

Image source: Origin

The Wukong system is the third generation of Origin quantum computers. In January 2024, cloud access was opened to it from all over the world. As the developers admit, the flow of scientists was led by researchers from the United States, despite the fact that Chinese scientists still have no access to similar resources from Western partners.

«This is the first time that a real quantum computer has been used to fine-tune a large language model in a practical setting, demonstrating that modern quantum hardware can begin to support real-world AI learning tasks,” said Chen Zhaoyun, a researcher at the Institute of Artificial Intelligence at the National Science Center in Hefei.

According to scientists, the Origin Wukong system improved AI training results by 8.4% while reducing the number of parameters by 76%. Typically, supercomputers are used to solve such problems — a specialization of general-purpose AI — which requires significant computing and energy resources. A quantum computer that uses the principle of quantum superposition — a set of probabilistic states instead of two classical ones (0 and 1) — can exponentially speed up calculations with relatively modest resource costs.

In particular, the scientists demonstrated the benefits of fine-tuning a large language model using a quantum system for diagnosing mental disorders (error rates were reduced by 15%), as well as for solving mathematical problems, where accuracy increased from 68% to 82%.

To run AI training algorithms on the quantum platform, the researchers developed what they called “quantum-weighted tensor hybrid parameter tuning.” The quantum platform handled the weights, while the classical part prepared a large language model. Thanks to superposition and quantum entanglement, Origin Wukong’s platform was able to handle a huge number of parameter combinations simultaneously, which accelerated model specialization.