TRAINING AI FROM REAL-WORLD TOOL USE

Summary from the TLDR AI newsletter

Perplexity combined rejection sampling fine-tuning with hint-guided self-distillation so its Computer model could learn from both successful sessions and user-corrected failures.