NVIDIA hits 100% on ARC-AGI-3 public set via agent harness
mhmazur · x · 2026-08-22
Regarding NVIDIA's claim of "solving" ARC-AGI-3 with a 100% score on the public dataset, commentators clarify that this achievement is largely due to their AVO agent harness rather than the Opus 5 model itself. AVO utilizes persistent memory across context windows and inspection capabilities to boost scores from 30% to 100%. It is also noted that the official ranking uses a more difficult semi-private dataset, so public set performance does not equate to AGI.
Related event: NVIDIA's AVO Agent Scores 100% on ARC-AGI-3, Sparking Debate(14 posts)→
More from coding & agent
- User builds auction game app in 8 hours using Grok via phone — iannuttall · 2026-08-22
- SpaceX AI employee shares Grok Bot optimization tips: self-check, explicit goals, and structured output — elonmusk · 2026-08-22
- Codex CLI Update: Instant Startup After Lifecycle Redesign — charliermarsh · 2026-08-22
- Best Qwen 3.8 Variant for RTX A2000 with 12GB VRAM? — MrMrsPotts · 2026-08-22
- The $249 Pyramid-Shaped Autonomous Intern 2 Is a Tiny Computer Built for AI Agents — Matt Wolfe · 2026-08-22
- Dev ships a full site in 8 hours, 75% from his phone using Grok Bot and Cursor agents — iannuttall · 2026-08-22