The Perils of Optimizing Learned Reward Functions: Low Training Error Does Not Guarantee Low Regret
May 1, 2025
Evaluating Superhuman Models with Consistency Checks
The last decade has seen an impressive rate of progress in machine learning (ML). Fueled by advances in deep learning and computer hardware, the performance of ML models has been steadily improving with some models even achieving superhuman performance in their area of expertise, meaning that they perform better on average than any human does.
May 10, 2024