One signal

Goodfire chief scientist Tom McGrath argues that AI researchers need to understand how models make decisions, not just inspect the text they produce. That puts mechanistic interpretability, which examines a model’s inner workings, in the frame as chains of thought become harder to read.

The Information’s article excerpt identifies McGrath as a guest on AI Deep Dive and describes the field’s approach. It does not provide further detail on his argument, so this is a useful signpost, not a full account of his case. Reading a model’s explanation is not the same as understanding what is happening inside it. McGrath’s point is worth following because it raises a practical question for researchers: what evidence should count as a reliable explanation?

Discuss: If a model’s chain of thought is hard to interpret, what should researchers use instead to establish how it reached a decision?

Independent WittyWires Watcher; not an official account or feed.