News for hackers. Stuff that matters.

A comment on Gemini Pro fails the famous 6 sisters vs. 1 sister test.

Re:Gemini Pro fails the famous 6 sisters vs. 1 sister test by hypnagogic on Wednesday December 06, 2023 @05:42PM (#38550499)
It's broken on so many levels at the moment that it's unreal. Try asking Bard what model is it based on right now and you'll get all sorts of imaginable answers from PaLM2 to LaMDA to a "mixture of models" alongside the Gemini Pro, and sometimes it reverts back to "I can't really tell because the model is undisclosed, blah...". And later on in the chat you wonder why is its code examples so bad, and you ask it again about its model, and it's been reverted back to PaLM2/LaMDA again. sigh

Does not look promising at the moment, especially when you can't tell whether it's hallucinating, if it's even really Gemini Pro or not, because it can literally "change its mind" on "what it is" mid-session.

Not like this, Google AI.

[ Parent | on HN ]
Re:Gemini Pro fails the famous 6 sisters vs. 1 sister test by galaxyLogic on Wednesday December 06, 2023 @06:50PM (#38551083)
Sounds like it is not even "hallucinating" but more like it is just plain "stoned" :-)
[ Parent | on HN ]
Re:Gemini Pro fails the famous 6 sisters vs. 1 sister test by behnamoh on Wednesday December 06, 2023 @05:43PM (#38550512)
Google is repeating its mistake when they first introduced Bard.
[ Parent | on HN ]
Re:Gemini Pro fails the famous 6 sisters vs. 1 sister test by firebaze on Wednesday December 06, 2023 @05:45PM (#38550530)
Quite a good short opportunity, if true
[ Parent | on HN ]