Huawei AI Clears 25 Games with RSI, Reportedly Using 56% Fewer Actions Than Humans

Huawei's AI system reportedly cleared 25 games using a method it calls RSI, according to a report from the Chinese outlet QbitAI. In one of them, the system beat its opponent 21:0.

The more striking number is efficiency rather than the scoreline. The report says the system used 56% fewer actions than human players, meaning it reached the same objectives with fewer moves instead of brute-forcing wins through sheer speed.

QbitAI also notes a detail that matters for how these systems are built: at a certain point in a match, the AI no longer needs to call a large model at all. If that claim holds, it suggests the system internalizes part of its decision-making into a lighter policy, reducing its reliance on large-model inference.

Games have long served as a testbed for agent capability, because the rules are explicit, feedback is immediate and weaknesses in generalization show up quickly. A result spanning 25 games is usually presented as evidence that a method generalizes rather than just optimizing a single title.

The report, however, does not lay out the full technical details or evaluation methodology in the same piece. Which 25 games were used, how the human baseline was set and where the no-large-model-calls behaviour actually applies are all questions that a more formal write-up would need to answer.

What to watch next: whether Huawei publishes a full technical report and benchmark data, whether RSI transfers from games to real-world tasks, and whether the lighter inference path translates into lower cost and faster responses. For teams building AI agents, wins on the efficiency axis tend to matter more than any single match score.