Training AI models using other AI models has become a highly sought-after objective for neolabs, and now a researcher within Anthropic’s fellows program has provided an early glimpse into what this could look like in real-world application. On Friday, Anthropic released a new paper titled "Automated Researchers Can Reliably Mitigate Alignment Failures," which explores how AI systems could consistently enhance a model’s performance across a series of alignment benchmarks. When presented with ten benchmarks targeting distinct misaligned behaviors, the automated systems managed to boost performance on each one without compromising overall quality. Spearheaded by Anthropic Fellow Chen Yueh-Han, this system mirrors much of the conventional research methodology. Every automated system scours the existing literature, devises a strategy, and trains the model using that approach for 30 minutes, progressively raising the benchmark over multiple cycles. Successful methods are retained while ineffective ones are discarded, enabling the system to function swiftly and on a massive scale. "Overall, these results provide early evidence that automated alignment post-training could become practical in the near term," the paper states.

This work represents a step toward recursive self-improvement, which many consider the next major milestone in AI development. If models can refine their own alignment training, it’s conceivable they could enhance training practices more broadly—at which point, human AI researchers might soon find themselves obsolete. The paper doesn’t shy away from this notion, explicitly drawing comparisons between the Automated Alignment Researcher (AAR) and its human counterpart. "The best AAR method beats what experienced humans propose, on average within six hours," the paper notes. "Human guided research directions do not lead to stronger performance." There’s even a cost breakdown for those still skeptical. "An AAR costs roughly $4 per hour in API inference against the $150 per hour we pay our human researchers."

To be fair, the paper also acknowledges certain constraints of this approach. The automated system is only as effective as the benchmarks themselves, which must accurately reflect the alignment objectives. Even then, considerable effort remains in establishing and maintaining those benchmarks—not to mention continually updating and broadening the literature that the automated researchers draw from.