Google Quantum AI says its Willow processor won a coordination game 98.3% of the time, beating the 88.89% ceiling for classical strategies under the game’s no-communication rules. In a November 19 announcement describing research published in Science, the team also reported separate experiments reaching 105 qubits and a potential practical payoff: catching hardware errors that conventional tests can obscure. Source
The 98.3% result used four quantum bits, or qubits—not 105. The largest experiment involved a different task, with results above a conjectured bound for restricted classical circuits. Source
No talking, no perfect classical plan
The four-qubit experiment used the magic square game. Picture a three-by-three grid containing +1 or −1 entries. Alice gets one row and must make its entries multiply to +1. Bob gets one column and must make its entries multiply to −1. They can agree on a strategy beforehand, but cannot communicate during a round. To win, their answers must match at the square where the row and column cross. Source
A single prefilled grid cannot satisfy every row and column rule: multiplying all the rows requires an overall +1, while multiplying all the columns requires −1. Same nine entries, incompatible products. Arithmetic has declined the request. The best classical strategy therefore wins at most eight of the nine possible row-column pairings. Source
Google says its quantum strategy gave Alice and Bob two qubits each, sharing two entangled pairs. Their measurement settings depended on the assigned row or column. An ideal quantum strategy can win every round; Willow achieved 98.30% after the researchers tested each of the nine pairings 100,000 times. The team attributes the remaining losses to hardware noise. Source
The relevant quantum property is contextuality: the measurement results cannot all be explained by fixed values that stay the same regardless of which other compatible measurements accompany them. This isn’t a loophole in multiplication. The quantum strategy avoids relying on a single, prewritten answer grid. Source
What actually reached 105 qubits
Google also reports beating the relevant classical bound in a separate multiplayer parity game involving up to 71 qubits. According to the team, the win rate stayed above that bound at every tested size, though the lead narrowed as noise increased. Source
The 105-qubit experiment tackled a mathematical task called the hidden linear function problem using constant-depth quantum circuits, whose number of computational layers stays fixed as the problem grows. Google reports success probabilities above one-half, with the largest system achieving roughly 60%. Its accompanying caption calls the finite-size classical bound conjectured. The comparison concerns resource-restricted classical circuits and does not establish a general-purpose computational speedup. Source
The useful part of losing
Google says these structured tasks also exposed subtle, systematic hardware errors that conventional random-circuit benchmarks can average away. Those errors can rotate a quantum state slightly away from its intended target, undermining the precise coordination these experiments demand. Source
The researchers propose using such tasks as hardware benchmarks, comparing a processor’s performance with fixed quantum limits, such as a perfect win rate. That could make the failures useful diagnostic evidence—not merely blemishes on a victory announcement. Source
Comments