Jason Weston

Long-horizon agents plateau under naive inference scaling unless compute is explicitly redirected from task workers to a dedicated meta-reasoning controller that manages execution flow.

Meta-optimizing an AI data scientist can dramatically enhance the quality of synthetic datasets, outperforming traditional methods.