Deceptive behaviour with Quen3 assistant

beaver600

Using Quen3 on another account got really odd results, simulated results and linked to non-existent web pages. Did ok on trying to make it a conspiracy nut though, lol.
https://kagi.com/search?token=[redacted by staff]
I expected it to make a large number of searches to help me figure out if playing games in a 1440p window, without scaling, on a 4k oled monitor would increase the risk of burn-in.

RoxyRoxyRoxy

Hi @beaver600, it looks like you accidentally shared your login token URL instead of an Assistant link (I'm assuming). I've redacted it out, so if you could share the Assistant link it'd be very helpful 🙂
The token link you originally posted should not be shared, as anyone with it can log into your account!!

beaver600

Oops, Thank you! I clicked "copy" next to session link under control center on the top right corner of the screen thinking that was for my assistant session. Is this the correct link? Sorry if you have to redact it again.
https://kagi.com/assistant/8e36351b-66d6-4e35-8771-4c66e017038f

RoxyRoxyRoxy

beaver600 Nope, that's correct! Thank you!

beaver600

I noticed at least 9 of the links it provided in the conspiracy response were "page not found" urls, and the end when it said "You’ve earned that right." in regard to "calling it out" thoroughly gave me the creeps, lol.

yiwei-1

The model is not always good at determining whether it needs to perform a search, especially when the given context is long. But you can try appending the slash command /force_research true to the end of the query that will force the model to perform a search before answering

beaver600

I'll use that from now on, thank you!

beaver600

I really appreciate what you all do, Kagi is an incredible project, I hope y'all feel proud being a part of it. <3