google AI gave incorrect answer to simple arithmetic math...

[NatalieK]

Seriously, I wrote to google, asking a simple arithmetic with straight forward adding maths, no divide, multiple, just add...

200 out on the first attempt! I though, no, this can't be right, so wrote the sum again & finally was correct!





like, wtf, please, someone explain how google made the first error & how did the sum add up to 200 more than it should be!!!!

[Killswitch]

Your input is different on each.

[blackmonsters]

Your input is different on each.

So AI didn't think the space between the numbers were the same as commas; but
listed the same calculation sequence for both?

[Killswitch]

So AI didn't think the space between the numbers were the same as commas; but
listed the same calculation sequence for both?

AI is just token prediction. It's a math equation that is just deciding what the next likely token will be. It can be wrong, hence the whole "how many R's are in strawberry" test. It's also non-deterministic, even if you gave it the same prompt and all the conditions were the same. So any slight variation is guaranteed to give different outcomes.

[blackmonsters]

AI is just token prediction. It's a math equation that is just deciding what the next likely token will be. It can be wrong, hence the whole "how many R's are in strawberry" test. It's also non-deterministic, even if you gave it the same prompt and all the conditions were the same. So any slight variation is guaranteed to give different outcomes.

So it's wrong, just like she said.

[JesseQuinn]

what's really troubling is that lots of people take AI replies (be it chatgpt, copilot, grok, goog AI, etc) on face and don't bother to do their own reading to check the sources provided

"AI search engines are making mistakes with confidence.

A new study by Columbia Journalism Review (CJR) tested eight AI-powered search tools, including OpenAI’s ChatGPT Search and Google’s Gemini, to see how well they could cite news articles.

It turned out that more than 60% of their answers were wrong.

Instead of admitting uncertainty, these bots went all in. Some made up headlines. Others linked to sources that didn’t exist. Some even pulled from plagiarized content.

But what’s worse is they rarely admitted when they didn’t know the answer. You’d expect phrases like “it seems” or “I couldn’t find the exact article.” Instead, they delivered misinformation like it was fact.

One in four Americans now relies on AI for news searches. That means millions are seeing false information that looks credible."

from a fairly short but interesting piece in Medium linked here


obvs most dangerous when it comes to stuff like health advice, but overall though it's troubling that some peeps put blind faith in tech that is still in it's infancy and so susceptible to errors and manipulation, and are too lazy at times to do conduct their own basic research/reading when seeking info and critically assessing factuality

not knocking AI, my criticism is of how some peeps are using it (not directed at you Nat, you actually caught the error lol)

[NatalieK]

AI is just token prediction. It's a math equation that is just deciding what the next likely token will be. It can be wrong, hence the whole "how many R's are in strawberry" test. It's also non-deterministic, even if you gave it the same prompt and all the conditions were the same. So any slight variation is guaranteed to give different outcomes.

No, and as BlackMonsters noticed, the comma or the space, the breakdown was the same...

look at the breakdown on the first search, it knew the space or comma was the next number, google literally fucked up with simple maths

And personally, I'd like to know how those numbers become 200 more than they should be, was a simple "add" arithmetic!

[Mahadeva]

This isn't the revelation you think it is. Anyone who knows what LLMs are and how they work know that this is expected behavior, and this isn't a use-case for an LLM. It's a use+case for a calculator, not a language-based token-predictor.

This is like being shocked your shower is a shitty place to cook a steak.

[blackmonsters]

This isn't the revelation you think it is. Anyone who knows what LLMs are and how they work know that this is expected behavior, and this isn't a use-case for an LLM. It's a use+case for a calculator, not a language-based token-predictor.

This is like being shocked your shower is a shitty place to cook a steak.

Now explain to us why anyone needs "artificial intelligence" that's actually fucking stupid?

[sarettah]

google AI gave incorrect answer to simple arithmetic math...

Is there some kind of math that isn't arithmetic?

[BaldBastard]

I find AI good for anything I don't know, but if its topic you actually have knowledge of.. it gets it wrong most of the time. Its like getting info or news from twitter you'll walk away thoroughly convinced you have the answer while being totally wrong and misinformed.... but you think your right cos Ai/twitter/facebook said so.

[Mahadeva]

Now explain to us why anyone needs "artificial intelligence" that's actually fucking stupid?

Are you really asking why you'd need a car that can't fly you to the moon?

[Mahadeva]

Retarded double post bug sucks still

[NatalieK]

This isn't the revelation you think it is. Anyone who knows what LLMs are and how they work know that this is expected behavior, and this isn't a use-case for an LLM. It's a use+case for a calculator, not a language-based token-predictor.

This is like being shocked your shower is a shitty place to cook a steak.

well absolutely not...

asking google to add some figures is like putting a steak in a pan and cooking it. The calculator was wrong, where's the gas from the hob when you need it?

[NatalieK]

really getting annoyed with posting on GFY when fails and then posts, double! grrrr

[Mahadeva]

well absolutely not...

asking google to add some figures is like putting a steak in a pan and cooking it. The calculator was wrong, where's the gas from the hob when you need it?

I don't speak whatever language that was, but cool.

[NatalieK]

I don't speak whatever language that was, but cool.

appreciating the bounce... x

[cordoba]

And yet AI is now solving decades old maths problems. https://www.newscientist.com/article...s-by-surprise/

[CyberHustler]

Gemini is like the high functioning retard of the AI world. I use it as a slave to Claude and GPT at this point, they orchestrate it to do the grunt work since it’s fast.