What Models Know, How Well They Know It: Knowledge-Weighted Fine-Tuning for Learning When to Say "I Don't Know"
The episode discusses the paper "What Models Know, How Well They Know It," focusing on Knowledge-Weighted Fine-Tuning. This technique allows AI models to quantify their knowledge by tracking multiple sources and calculating consensus levels. Hosts conclude that this enables models to accurately state when they lack information, shifting AI from a 'confident guess' to a reliable, calibrated certainty.