danwitt

It's not magical and you won't get good results if you ask a question that is that vague. Instead you will want to decompose it in to decidable questions that make up "is urgent" and if you find a case that corresponds to "is urgent" you add a new question to the panel that surfaces that.

Personally I want to see the obnoxious comments it’s leaving to poison future sessions.

I think the real contract being violated is a bit narrower than that. More precisely, the pre-LLM assumption was that producing a piece of content took more effort than consuming it, so the content existing at all signaled that someone cared enough to put some time into it.

AI can definitely be used as a tool in a workflow that's mostly driven by human effort. But as a reader, it's hard to tell whether a particular piece of AI-assisted writing came out of that kind of workflow without investing the effort yourself, and the prior of AI Writing == AI Slop is pretty strong. Attribution seems like a partial solution, since someone putting their name behind it will suffer reputational harm if they're producing nothing but slop.

K3 is relatively slow, but 12 minutes points to the harness not the model IMO.