Post

That is actually awful. Level of accuracy with these models just isn’t good enough to have justified releasing tbh.

Yeah agreed. I think this style of notifications is where it falls over - each notification being completely unrelated. It works nicely on threads of messages, as they’re all contextually related. My guess is, it’s been built on that assumption.

Was trying to look at why this happened and expect if you’re only using headlines, that may be really really hard? Headlines often get grammatically buttered to keep under the limit relying on reader knowledge to disambiguate, which LLMs won’t have/will have had removed.