Anthropic demonstrates AI improving other AI with automated alignment researcher
Technology
Anthropic just showed that AI can actually improve other AI models, not just itself.
Their new research, led by Anthropic fellow Chen Yueh-Han, introduced an "Automated Alignment Researcher" (AAR) that boosts how well AIs follow rules, all without breaking what already works.
AAR beats humans within 6 hours
The AAR acts like a digital researcher: it searches the available literature, suggests fixes, and tweaks models in quick 30-minute cycles.
The best AAR method beats what experienced humans propose, on average within six hours, and runs at only $4 an hour (compared to $150 for a person).
While it still needs solid data to work with, this could be a major leap toward AIs that improve themselves safely.