[visarga]

> It's notoriously awful at extrapolating, ie generalizing to anything without support in the training data.

Scientists are also pretty lousy at making new discoveries without labs. They just need training data.