Summary
The episode works through four questions that students often find difficult in the first lesson of the course: why users of AI assistants benefit from knowing how neural networks learn, how a model with higher accuracy can be the worse choice, why a model's most predictive feature can mislead, and how a gradient descent step works when the label is zero. It places the evidence on AI use and productivity in context, including what surveys of use can show, who took part in the experiments, and the reversal found on a task outside the jagged frontier. Worked examples include checking an assistant's flawed explanation of overfitting, comparing two hypothetical loneliness prediction models by accuracy, sensitivity, specificity and positive predictive value, and identifying data leakage in a hypothetical model of program dropout. The hosts close by debating whether the fictional Cedar Valley team should use a prediction model to choose whom connectors contact, and they agree on a rule for adding a locally validated model alongside the existing screening route.