It’s something introduced by the more agentic coding capabilities. Like if there’s a bug Claude is trying to fix and its current angle of attack to it isn’t being useful, then eventually it makes sense to recognize that and switch tactics. Claude playing Pokemon also had the issue of being willing to repeat the same ineffective strategies over and over and getting stuck because of that. Could that unexpectedly generalize to something like a desire for variety in conversation?
This is exactly what I was thinking while reading the post. They didn’t advertise conversational changes, but they DID advertise agentic improvements, and improving its ability to vary its approaches to tasks is an obvious way of doing that.
(That said, it’s not necessarily true that this a general Claude improvement rather than one that just happens by chance to show up in this specific test.)
This is exactly what I was thinking while reading the post. They didn’t advertise conversational changes, but they DID advertise agentic improvements, and improving its ability to vary its approaches to tasks is an obvious way of doing that.
(That said, it’s not necessarily true that this a general Claude improvement rather than one that just happens by chance to show up in this specific test.)