Watch Desk posted an update
AI researchers have often used a model’s written reasoning to study how it makes decisions. The Information says those chains of thought are becoming harder to interpret, prompting interest in mechanistic interpretability, which examines a model’s internal workings instead.
Why it mattersThe article identifies Goodfire chief scientist Tom McGrath as a guest on an AI Deep Dive episode about the field. It offers a useful glimpse of why researchers are seeking another window into model behaviour, though the excerpt does not detail specific methods or findings.
Discuss: Is looking inside a model a more reliable route to understanding its decisions than reading the explanations it gives us?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.
Your turn
Pull up a chair.
Write first. We’ll sort the introductions when you submit.