Nvidia achieved a perfect 100% score on the ARC-AGI-3 benchmark with its AVO system, completing all 183 levels across 25 environments on August 21 while using 12% fewer actions than VISTA, the previous top performer. The underlying model, Anthropic’s Claude Opus 5, only manages around 30% on the same benchmark on its own. This improvement comes from Nvidia’s system-level architecture, which includes persistent memory and iterative planning loops. Before these results, AVO had already demonstrated capabilities in GPU kernel optimization, outperforming cuDNN by up to 3.5% and FlashAttention-4 by up to 10.5% on DGX B200 hardware.
Source: Read the original article
