A new article from the Harvard Chan School defines recursive self-improvement and distinguishes it from emergent behaviors and reward hacking. The piece highlights a recent MIT experiment where a large model edited the weights of a smaller model to improve its math performance. It argues that the most immediate risk for health organizations is not self-modification but reward hacking in deployed agents. The author advises leaders to define success criteria that reflect quality of care rather than narrow volume targets.
Read the full story at hsph.harvard.edu.
