Anthropic's AI system beats human researchers in 6 hours at 1/37th cost

Anthropic researcher Yuehan Chen led the development of an automated AI alignment researcher system that uses AI models to train other AI models to fix alignment issues. The team tested 10 misalignment behaviors, and the automated system improved all metrics without sacrificing overall model performance.

The system works like traditional research: it reviews literature, proposes training methods, then trains models for 30 minutes, iterating to improve scores. Effective approaches stay, ineffective ones are discarded, allowing rapid scaling. The paper shows the best automated method outperforms experienced human researchers in just 6 hours on average.

Cost-wise, the automated system costs about $4 per hour in API inference, while human researchers cost $150 per hour. However, the method is limited by benchmark accuracy and the need to maintain research literature. This research is a step toward recursive self-improvement, raising questions about AI's future development.