Training AI systems utilizing other AI frameworks has emerged as a highly sought-after objective for neolabs — and now, an investigator in Anthropic’s fellows initiative has offered us an initial glimpse at how this could manifest in real-world applications.

On Friday, Anthropic released a new study titled “Automated Researchers Can Reliably Mitigate Alignment Failures,” explaining how AI systems might consistently enhance a model’s performance against a series of alignment criteria. When presented with 10 measures for particular misaligned actions, the automated systems succeeded in boosting performance on each one without compromising overall efficacy.

Headed by Anthropic fellow Chen Yueh-Han, the system emulates much of the conventional methodology in research. Each automated entity scans the existing literature, suggests a technique, and trains the model using that technique for 30 minutes, steadily increasing the benchmark over multiple iterations. Successful methods are retained while those that are ineffective are eliminated, enabling the system to function rapidly and on a large scale.

“Overall, these findings offer preliminary proof that automated alignment post-training could be feasible in the near future,” states the paper.

The study is a move towards recursive self-enhancement, which many view as the next critical advancement in AI development. If models are capable of refining their own alignment training, it’s likely they could enhance training methodologies more broadly — at which stage, human AI researchers might soon be rendered unnecessary.

The paper openly confronts this notion, directly contrasting the Automated Alignment Researcher (AAR) with its human counterpart. “The best AAR method outperforms what seasoned humans propose, on average within six hours,” notes the paper. “Human-guided research directions do not yield superior results.”

There’s even a financial comparison, should anyone remain skeptical. “An AAR incurs a cost of approximately $4 per hour in API inference, compared to the $150 per hour allotted for our human researchers.”

In fairness, the paper also acknowledges certain limitations of this method. The automated framework only functions effectively to the extent that the benchmarks accurately align with the genuine alignment objectives, and even then, considerable effort is needed to establish and uphold those benchmarks — not to mention the necessity of maintaining and expanding the literature from which the automated researchers derive their information.