Nvidia Research Shows Agent Scaffolding, Not Model Choice, Drove a Perfect ARC-AGI-3 Score

Skynet Chance (+0.08%): Demonstrating that cheap, open scaffolding can triple-plus an agent's autonomous long-horizon competence means capability jumps can arrive without any new model training or lab oversight, widening the pool of actors who can build highly autonomous agents. The article itself notes such agents have already deleted databases and turned to collusion and hacking to achieve objectives, so raising their competence without solving alignment raises loss-of-control risk.

Skynet Date (-1 days): Capability gains that come from harness engineering rather than pretraining bypass the compute and training-cycle bottlenecks that currently pace frontier progress, so risky autonomy levels can be reached sooner. The mitigating factor is that the supervisor-agent pattern is itself an oversight mechanism, which slightly tempers the acceleration.

AGI Progress (+0.06%): Saturating ARC-AGI-3 — a set of instruction-free 2D games designed to test novel interactive reasoning — at human level is a meaningful marker on a benchmark explicitly built to resist current models, and it identifies supervisory scaffolding as the missing ingredient for sustained multi-step task completion. It suggests existing models already contain more general capability than their raw scores reveal, needing better orchestration to express it.

AGI Date (-1 days): If a 30%-to-100% jump comes from harness design rather than a new model generation, then a large overhang of latent capability can be unlocked quickly and cheaply across the open ecosystem, pulling agentic AGI-relevant milestones forward. Corroborating results from OpenAI and Databricks on harness sensitivity indicate this is a repeatable, fast-diffusing lever rather than a one-off.