Anthropic Paper Shows Automated AI Researchers Outperforming Humans at Fixing Alignment Failures

Skynet Chance (+0.05%): Automating alignment post-training is risk-reducing in principle, but the demonstrated step toward recursive self-improvement and the explicit finding that human-guided direction adds nothing shifts oversight away from humans and makes benchmark-gaming (optimizing proxies rather than true alignment) a central failure mode. On net the loss-of-control surface grows more than the mitigation offsets it.

Skynet Date (-1 days): A working, cheap automated researcher loop that iterates in 30-minute training cycles compresses the timeline for models improving their own training, pulling forward the point at which capability gains outpace human review. The 40x cost advantage over human researchers means this scales immediately rather than gradually.

AGI Progress (+0.04%): Demonstrating that an AI system can independently propose, test, and select research methods that beat experienced human researchers on a real task is direct evidence of autonomous research capability, a core AGI-relevant competency. It is still narrow — confined to alignment benchmarks with curated literature — so it is a meaningful step rather than a discontinuity.

AGI Date (-1 days): If automated researchers generalize beyond alignment post-training to research practices broadly, the human-researcher bottleneck on AI progress weakens substantially, accelerating the path to AGI. The stated cost differential makes massively parallel automated experimentation economically trivial today.