HOST: So, what is the reader trying to read?
EXPERT: Okay, so here we're talking about a weight update, basically the little numerical tweak that happens when a model trains on a new fact or a response tendency.
HOST: Does it get to see the lesson that caused the change?
EXPERT: No, the authors just present the update and ask a neutral question. It's like teaching a made-up planet fact, then later asking what changed without ever bringing up planets.
HOST: So how often did it actually get the whole answer then?
EXPERT: On held-out updates, at least one of a hundred attempts passed for about 2% of facts and 16% of behaviors. The authors say complete descriptions are still unreliable.
HOST: Then how does Meta Edit use the reader?
EXPERT: It starts with a written target self report. Gradients show which weight changes affect the reader's likelihood for those words. MetaEdit uses that signal to choose rows to remove or adjust in the original model.
HOST: So what should I remember about the intervention results?
EXPERT: Their measure changes in controlled tests, not proof of a deployed system. The authors didn't test reading arbitrary updates or extracting the examples used for training.