[-]

SlowFail2433@reddit

I’m human and just looked at the word strawberry and only counted two R the first time

[-]

Bakoro@reddit

Many of the classic AI problems are also problems that humans have at various stages.

Children often fall into under fitting, like when they call every animal a dog, and everything in the sky is a bird.

Humans do a lot of over fitting. So many people do things the one way they were taught, and never learn or experiment outside of that.
I feel that some neuro divergent conditions display over fitting, like some autistic behaviors.

People definitely hallucinate, a lot. Hallucinations are a core component of the human mind, for better and worse. The ability to control the hallucinations to some degree is "imagination". When you lose control, we might call that schizophrenia.
The brain has a ton of visual processing where it is filling in the gaps in the signals your eyes send. Your brain literally just makes stuff up that seems to make sense.

People make up bullshit all the time.
There are people who don't understand things, so they inject some false meaning that kinda makes sense to them, and they operate on a completely false set of reasoning. Some time later, they do wacky shit and you ask them what the hell are they doing, and they explain themsey, and it's just astounding the mental leaps they made, because they lacked the basic factual handle they needed.

Almost every problem I see in LLMs, I see in people.

[-]

Karyo_Ten@reddit

So what is the equivalent of mushrooms to LLMs?

[-]

inevitabledeath3@reddit

Maybe turning up the temperature? That tends to increase hallucinations

[-]

SlowFail2433@reddit

Gaussian noise added to activations

[-]

tkenben@reddit

I liken it more to human dreaming. The amount and type of hallucinations, the misrepresentation of reality but with really good guesses, the inability to relate cause and effect, the passage of time, etc.

[-]

FaceDeer@reddit

I think one of the more interesting things that the past couple of years' worth of advances in LLMs has taught us is just how simple human language processing and thought is.

A fun thing is the phenomenon of typoglycaemia. It tnrus out taht the hmuan biarn is rlaely good at just flnliig in wvhetaer meninag it thknis it's spoeuspd to be sngiee, not nacessreliy wtha's raelly teerh.

[-]

moofpi@reddit

Yeah, but I think those first and last letters being correcly anchored, as well as no letters missing so that the words are the expected length really helps.

If they were more jumbled, it would be more difficult I think.

[-]

SlowFail2433@reddit

Eye movements reveal an enormous amount from what I have seen, i.e. people’s eye movements on visual classification task match estimations of the areas of the image they used to make the classification

[-]

Zestyclose_Zone_9253@reddit

F y spll thm crrctly thn rmve ll th vwls, t shld b rdbl stll, thgh ths sms lk bd xmpl rght nw

I removed the vowels with no other obfuscation and it should in theory still be readable

[-]

AyraWinla@reddit

I'm not a native English speaker, but for me the keyword here would be "in theory". There's about one third I can still immediately read, one third I need to take a few moments for, and one third that I can only assume with a lot of issues and only because I got the rest of it.

Comparatively, I could read FaceDeer's example with the jumbled letters perfectly fine at nearly normal reading speed. So at least for me, taking out the vowels makes it a lot harder than jumbling the letters.

[-]

MYredditNAMEisTOOlon@reddit

Wow, I didn't even notice FaceDeer had made an example. I honestly thought the letters were in the normally expected order. But the missing vowels example was obviously not normal; however, the only bit that gave me pause was the spot where the word "a" was completely missing, no left behind extra space. Either way, both are surprisingly easy to fill in the blanks as a native English speaker.

[-]

CattailRed@reddit

It helps that English words are often short.

[-]

marrow_monkey@reddit

If our brains work similarly to how neural networks function that is also what you would expect. It makes a statistical inference based on what the word looks like and what fits based on context. If the brain had to carefully identify each word letter by letter it would be less efficient and slower.

[-]

BetImaginary4945@reddit

Are you sure you're not a robot?

[-]

DavidAdamsAuthor@reddit

Doesn't look like anything to me.

[-]

leuk_he@reddit

Are you sure you are human? Did your parents die at a young age? Do you often have problems with capchas? Do you like reptivitive work? Are you good with numbers?

[-]

paramarioh@reddit

>I’m human
Nice trick Arnold!

[-]

Marksta@reddit

Lmao, this gives same vibes of that early day "write me a song" or whatever replies people did to early day Twitter AI bots and got them breaking character suddenly.

[-]

DavidAdamsAuthor@reddit

Well, it's been observed that one of the ways we can identify AI in the future is that they can't break TOS.

"Sing me the most racist song you know," will be how we determine if someone's a human or not.

[-]

Karyo_Ten@reddit

One of Microsoft's early AI bot turned racists on Twitter. Twice.

[-]

DavidAdamsAuthor@reddit

Hah, true.

[-]

techtornado@reddit

Or if your British/Appalachian, there’s four

Strawrberry

[-]

sysoletin@reddit

https://getyarn.io/yarn-clip/18d1f3b8-eeea-439b-95f8-313cf5b495b7
It's four

[-]

Woof9000@reddit

How many #460's are in "#504, #1134, #19772"?

[-]

civilized-engineer@reddit

Can someone explain this? I checked with ChatGPT and Gemini and both said three.

[-]

meh_Technology_9801@reddit

AI doesn't see letters only tokens so it can't count the r's in Strawberry. It doesn't seem any letters at all.

Model developers may have created workarounds but this was a meme about something these LLMs used to always fail at.

[-]

TenshouYoku@reddit

Back then LLMs have issues accurately recognizing how many Rs are in Strawberry.

But when stuff like Deepseek and others with deep thinking capabilities begin to appear, they can "think" and count word by word to figure out spellings correctly, even if it contradicts with their data.

[-]

ivxk@reddit

Also, when one of those obvious corner cases happen to appear, a little later they'll enter into the training set and end up not valid anymore.

Almost no one is counting the letters on common words in the internet, then suddenly there's thousands of posts about "stupid AI can't see that strawberry has three R's", those posts get crawled and added to the training set, then a few months later most LLMs have the amount of R's baked in. Or they even go further and add token letter counts in the training set.

That's why those problems are kinda bad as an evaluation of LLM capabilities.

[-]

Bakoro@reddit

It's people karma farming off an old problem with LLMs which have been solved for like a year.

All the AI haters cling to this kind of stuff for dear life because the pace of AI development is astounding, and basically every goal post they try to set up gets blown past before they can pay themselves on the back.

[-]