Commentary: NVIDIA's AVO agent architecture and the ARC-AGI-3 result

Mark Kretschmann (commentary), underlying work by NVIDIA (Terry Chen, Jean-Francois Puget, Humphrey Shi)

Type: social
Venue: X commentary
Year: 2026
Source: x
Access: public
Language: en
Added: 2026-09-29
Verified: 2026-09-29

Summary

Commentary thread on NVIDIA's AVO announcement, emphasizing that the AVO agent architecture took Claude Opus 5 from ~30% as a standalone model baseline to 100% on the ARC-AGI-3 public set (all 183 levels across 25 environments), crediting persistent memory and the surrounding agent system rather than the model itself.

Keywords

agents · ARC-AGI-3 · commentary · NVIDIA

Topics

agents, ARC-AGI-3, commentary, NVIDIA

Research notes

Discovery: Commentary X post by Mark Kretschmann on the NVIDIA AVO result; same underlying work as the NVIDIAAI announcement.
Key findings: Highlights the scaffolding-vs-model gap: same model, 30% alone vs 100% inside AVO
Limitations: Commentary, not a primary source; inherits the caveats of the NVIDIA result (public set only, company-reported, no independent leaderboard verification).
Post text per Discord embed; substance cross-verified against secondary coverage of the NVIDIA AVO announcement.