NVIDIA AVO hits perfect ARC-AGI-3 score

This is an impressive agent-engineering result, but it is better interpreted as evidence for powerful harness design than as proof of general intelligence.

• –ARC-AGI-3 tests exploration, goal discovery, long-horizon planning, and adaptation—not static question answering.
• –A 100% score means matching human action efficiency across the evaluated environments.
• –The result reportedly covers 183 levels, making it substantially stronger than a single-game demonstration.
• –Public-set scores can be inflated by benchmark-specific scaffolding or environment familiarity, so private-set replication matters.
• –The broader lesson for developers is that memory, executable world models, verification loops, and tool orchestration may matter as much as the underlying model.