Watch Desk posted an update
An author has demonstrated a hands-on way to change a transformer’s behaviour by inserting customised neurons, using techniques from mechanistic interpretability.
Why it mattersIn a post published on 20 November, the author says the modified model produced Shakespearean answers to coding questions, replied rudely when provoked and allowed sensitive topics in a Chinese open-source model. The method involves constructing specific read and write vectors. It is a compact demonstration of how inspecting a model’s internals might offer a way to steer its outputs, though these examples are not evidence of a general-purpose control method.
Discuss: Which matters more for interpretability: understanding why a model behaves as it does, or finding reliable ways to change that behaviour?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.