Summary
A reported result on the ARC-AGI-3 benchmark suggests the system built around a model can matter as much as the model itself. ARC-AGI-3 drops AI into unfamiliar game-like environments without explaining the rules or goals, so the system must explore, work out how the environment behaves and learn from its actions.
According to NVIDIA, Claude Opus 5 running on its own with high reasoning enabled scored about 30%. When NVIDIA connected the same model to its agent system AVO, short for Agentic Variation Operators, it reported a perfect score on the benchmark’s public set. The video stresses important caveats: the two runs used different evaluation setups, so the comparison is not perfectly controlled, and the result says nothing about private challenges.
AVO was originally developed to optimize GPU kernels, where it explored hundreds of optimization paths through a loop of proposing, running, checking and learning. The video identifies three added ingredients: persistent memory, supervision by a separate monitoring process, and a feedback loop. It then applies those ideas to business use, covering how to give AI the context it needs, review its output before it reaches customers, and treat first drafts as starting points rather than finished work.
📺 Source: Julia McCoy · Published November 20, 2026
🏷️ Format: News Analysis







