Goertzel posts 33% ARC-AGI-3 score

The interesting part here is not just the score, but the method: this is another data point that scaffolding, not raw model prompting, is what matters on interactive agent benchmarks.

• –ARC-AGI-3 is no longer a static puzzle test; it rewards exploration, hypothesis revision, and long-horizon planning, so agent architecture matters as much as model quality
• –The reported 32.58% puts this result in the same rough band as other public benchmark claims, which suggests the real bottleneck is search, memory, and environment modeling
• –The writeup is also a warning label for benchmark hype: a clever verifier loop can move the number without proving general intelligence
• –For developers, the takeaway is practical: if your system can maintain a world model and self-check its own plans, you may get farther on hard evals than by swapping in a better base LLM
• –The benchmark’s value is still real because it pressures teams to build agents that adapt over time, not just answer well once