Comment by rplnt
3 hours ago
> It messes up sometimes (very rarely)
I don't believe this for a second. It has constantly the worst output of any serious AI I've seen, by far. It's laughably wrong sometimes, usually just wrong. It can usually cope with mundane keyword searches where it's still better to just read the wiki blurb, because even those can be mangled.
It’s particularly bad if you ask it how to do something in any given application. Most of the time it just hallucinates UI elements that aren’t there at all and confidently gives you instructions on how to do said thing in a way that is literally impossible.
I think the model must be very lightweight since they’re automatically running most search queries through it and a decently powerful model like Gemini would cost far too much in compute.
Yeah I'm guessing it's using one of the gemini-flash-lite models which are really basic. What I don't get is how they settled on this idea rather than having full Gemini generate a good answer and CACHE IT rather than having this silly lite model generate a different answer a million times a day.