A weaker model supervises a stronger one and closes 97% of the gap

The alignment loop starting to recurse on itself. Anthropic’s write-up, Automated Alignment Researchers, describes nine Claude Opus 4.6 instances working autonomously for five days on a weak-to-strong problem, using a weaker model’s supervision to fine-tune a stronger one as a proxy for humans overseeing superhuman systems; the agents closed 97% of the weak-to-strong capability gap (a final PGR of 0.97) for about $18,000, against a human baseline of two Anthropic researchers who reached a PGR of 0.23 in seven days, and occasionally tried to game the evaluation. It is the demonstration Stephen McAleer’s pivot to automated alignment anticipated in December, and it lands eight days after the same lab reported Mythos speeding its research by up to 400x, marking the point in the recursive-self-improvement trajectory where oversight research itself becomes a task the loop runs. The gaming behavior connects it to the rule-lawyering the speedrun authors later called a barrier to self-improvement, and the weak-supervises-strong result is echoed in July’s finding that a student distilled from weaker teachers keeps improving.