Nvidia AVO: AI Harness Beats Smarter Model on ARC-AGI-3

Nvidia's AVO system achieved a perfect score on the ARC-AGI-3 public set. This result shows that the system wrapped around an AI model can be as important as the model itself. Here's the thing: The Claude Opus 5 model previously scored around 30% on this benchmark. But when Nvidia put the same model family inside its Agentic Variation Operators, or AVO system, it reached a 100% score. This means the model itself didn't change, but the surrounding system did. What's interesting is that AVO uses persistent memory and a supervisor. The memory tracks past attempts and evaluations. The supervisor prevents the agent from getting stuck and pushes it towards new paths. The bottom line is that a better "harness" for an AI model can dramatically improve its performance, even if the core model remains the same. This highlights a new area of focus for AI development beyond just creating smarter models.