Anthropic’s latest paper reveals that automated systems can reliably enhance AI models, hinting at a future where self-improving AIs might render human researchers obsolete.

Lead researcher Chen Yueh-Han’s system follows a traditional research approach, automatically searching for methods, proposing them, and training models over several iterations. The results? Every benchmark was improved, without any loss in overall performance.

While the cost of using an Automated Alignment Researcher (AAR) is a mere $4 per hour compared to $150 for humans, the system is not without limitations. The benchmarks must accurately reflect alignment goals, and maintaining such benchmarks remains a significant challenge.

Recursive self-improvement could be the next big leap in AI, but it could also spell the end for human AI researchers. The paper suggests that AARs often outperform human researchers, even within six hours.

The implications are profound. If AARs can reliably improve alignment and training practices, they may soon surpass human capabilities, leading to a future where AI evolves faster and more efficiently than ever before.