BETTER, CHEAPER, FASTER ALIGNMENT WITH KTO

Summary from the TLDR AI newsletter

Most alignment frameworks require ranked preference data (A is preferred to B). This is usually expensive and rare. This new framework only requires an independent yes or no rating per data point.