Post by @lauren@mastodon.laurenweinstein.org

@lauren@mastodon.laurenweinstein.org · 6 hours ago

Google AI Overviews vs. "The Wizard of Oz"

Google Gemini does it again. Moron. You know that when it can't get trivial but widely known stuff like this right, imagine the important things it's getting wrong!

Query: "or you hurt the spell"

AI Overview:

"That famous line comes from Miss Almira Gulch in the 1939 classic The Wizard of Oz. [note, Google didn't bother to quote the title -L] She says it to Dorothy while struggling to contain Toto in her basket, muttering: 'These things must be done delicately, or you hurt the spell.'"

Of course the line has nothing to do with Gulch putting Toto in her basket, it's spoken by The Wicked Witch of the West" to Dorothy regarding removing her ruby slippers. Obviously Google AI is SO STUPID that it can't tell the difference between the character of Gulch and her (Dorothy's dream) incarnation as the Witch, and couldn't even separate the two completely different scenes. Insanity.

Image: https://mastodon.laurenweinstein.org/@lauren/116596671260305171

Lauren Weinstein

@lauren@mastodon.laurenweinstein.org · 6 hours ago

#Google Gemini does it again. Moron. You know that when it can't get trivial but widely known stuff like this right, imagine the important things it's getting wrong!

#gemini #google #ai

Pino Carafa

@rozeboosje@masto.ai · 5 hours ago

@lauren once again, with feeling:

LLMs are unable to separate fact from fiction. If it is fed with language samples that link "Gulch", "Toto" "basket" and "hurt the spell" often enough it may spit out such a statement and will have no idea that it is not an accurate representation of what happens in the movie. It has no concept of what a "movie" even is.

Asking LLMs for statements of fact is idiotic.

Lauren Weinstein

@lauren@mastodon.laurenweinstein.org · 5 hours ago

@rozeboosje Well that's the problem isn't it? People ask Google Search for something and at the very top they get total garbage, in some cases horrifically dangerous garbage!

Pino Carafa

@rozeboosje@masto.ai · 5 hours ago

@lauren Yup, that's why I switched to DuckDuckGo *and* I use its NOAI link for searching.

Lauren Weinstein

@lauren@mastodon.laurenweinstein.org · 5 hours ago

@rozeboosje Doesn't solve the problem. The vast majority of people use Google and will continue to use Google, I have no interest in people who choose to go elsewhere in this context, I'm interested in what most people are faced with.

1 more replies

Peter Ludemann

@PeterLudemann@mathstodon.xyz · 6 hours ago

@lauren But regular Google search did ok for me. So, do we have something like the opinion writers at the WSJ who appear to not read the news pages of the WSJ? 🤡

Lauren Weinstein

@lauren@mastodon.laurenweinstein.org · 5 hours ago

@PeterLudemann My example was a regular Google Search, showing the AI Overview that resulted (the AI Overview backend is various Gemini models).

Peter Ludemann

@PeterLudemann@mathstodon.xyz · 5 hours ago

@lauren Seems to be a lot of non-determinism there ... when I tried it, I got correct answers (including https://www.imdb.com/title/tt0032138/characters/nm0002121/) but also a link to an utterly unrelated (except for the title) YouTube snippet. Maybe there's a lot of pressure on the servers, causing timeouts for retrieving critical pieces of information? (I'm not excusing wrong answers; just hypothesizing another cause of "AI" failures beyond just the fundamental technology issues)

https://www.imdb.com/title/tt0032138/characters/nm0002121/

Lauren Weinstein

@lauren@mastodon.laurenweinstein.org · 5 hours ago

@PeterLudemann There is non-determinism all over the place with AI Overviews/Gemini. In my ad hock testing, I've frequently tried the same query multiple times in a row and gotten completely contradictory responses presented as authoritative, often with source links that -- as you note -- have nothing to do with the answer provided (or even provide their own opposite answers --- though usually you simply can't find anything in them that directly relates). Timeouts? I.dunno. I HAVE seen AI Overviews stall for several seconds with messages like "Still working" and the like and another message I can't recall right now when the delay got really prolonged. I've also seen "Something went wrong" on the AIO panel and the AI presentation completely empty.

1+ more replies

Can Acar

@canacar@ioc.exchange · 6 hours ago

@lauren Not sure whether it is better or worse, but it is not "watching" the movie and interpreting it wrong. Confusing Ms. Gulch and the WW from the movie would be hard even though the the same actress played them both.

It seems it used a misattributed quote like this one: https://www.imdb.com/title/tt0032138/quotes/?item=qt2105155
And mixed in the other parts (containing Toto in basket) to match, and no, this is by no means acceptable.

https://www.imdb.com/title/tt0032138/quotes/?item=qt2105155