Thread 106370207

When will AI be, unironically, able to count letters? What's stopping 6 gorillion parameters from just figuring out arithmetic on its own?

>>106370207 (OP)
there simply isn't any training data counting letters except children's books

>>106370229
How about spelling? Literally the word is in the context. Lots of spelling training data.

>>106370229
>there simply isn't any training data counting letters except probably hundreds of books
yeah

>>106370207 (OP)
It's like this for everything. All it's good at is making gramatically correct sentences and identifying relevant tokens. This is why it's "good" at summarizing. And why it's shit at finding "important" snippets from a large number of relevant snippets. So the answer is: as soon as they put it in the training data or pass every stupid query through an agent that runs other software.

>>106370207 (OP)
>after counting and double checking, I'm confident there are three

An explanation for this I heard is that the AI tokens are words. They don't think in units smaller than words, so it's a miracle they sometimes get it right.

>>106370207 (OP)
Tokenization. It splits words up into tiny chunks and doesn't actually "see" letters or numbers.

>>106370207 (OP)
When they will add some prompt preprocessing hardcoded for this question

>>106370742
Bonsoir

>>106370490
Beautiful.

>>106370703
>>106370704
> AI only think in tokens
This has always been easily disproven cope

LLMs cannot think forward. Any sort of counting done is coincidental and the AI trying to come up with a solution to justify the answer, and not the other way around.
Another problem is LLMs (especially chat tuned ones) are heavily RLHFed to try and be as confident as possible, so it never even tries to correct itself after it generated the nonsense "solution".

That's what's going on there, the AI was never told how many Rs are in Londonderry so there's nothing in the dataset to work with and it picks 3 randomly. Then it spins and "hallucinates" to justify it being 3, because the tuning won't let it correct itself.

Doesn't really feel like something that can be fixed with the current limitations, this is what LeCun says and it makes AI fanatics angry.

>>106370207 (OP)
Sucks to be you retard.
Mine got it first try.

>>106370273
>AI is totally useless, guys!
Didn't they have a name for such an ideology in London during the industrial revolution... luddite? Because these pathetic retards wanted to hold on to their worthless jobs?

>>106370854
You are correct. The term "Luddite" originates from a group of English textile workers in the early 19th century (1811-1816) who protested against the introduction of new machinery during the Industrial Revolution. They saw these machines as a threat to their livelihoods and skills.
The Luddites, supposedly led by a figure named Ned Ludd, destroyed machinery and engaged in sabotage to express their opposition. Their actions were a reaction to the economic disruption and job displacement caused by automation.
Today, the term "Luddite" is often used to describe someone who opposes or is skeptical of new technology, particularly those who prefer traditional methods or are concerned about the impact of technology on society.

>>106370829
It picked three because of how ubiquitous the strawberry line is in its training set. Remember, it doesn’t think. By the time it’s parsed “how many ‘r’s are in” it’s already settled on 3 as the answer because of how intertwined those two concepts are but it doesn’t have a good reason why

Who the fuck says Londonderry in modern day, let alone counts letters in it. What is this trivia bullshit.

>>106370704
But why does it tokenize a single letter as multiple tokens? ('d' is tokenized as 'r' and 'ye'.) Whatever led to this is inscrutable and unfixable. If it can't even count tokens, why do you think it can write a website?

>>106370207 (OP)
Current AI is architecturally kneecapped. It's just a statistical approximation of what people have been recorded doing in the past. It has trouble guessing what to do in novel situations or acting in new and novel ways. It has no logical "core" capable of abstract reasoning on its own. It's just a collection of extraneous data points surrounding a hollow center, trying to make you think it has a core. It's still a crude imitation of true intelligence.

Someone would need to design an AI that is just a reasoning core, one with motivations and the ability to make decisions abstractly. Taking an abstract reasoning core and adding it to a collection of deep learning data points would effectively make true intelligence. You'd have both fluid and crystallized intelligence with the ability to continually build more crystallized intelligence.

but why does it figure it out later?

this is deepseek.

yeah no idea why it's also wrong...

i redirected: where is the third r

>>106370207 (OP)
2 trillion more

>>106371449
>architecturally kneecapped
londonderry having 3 r's is more mysterious.

why would deepseek and gemini agree it's 3?

>>106371517
>(white man)perhaps on the job training a little bit
>(fat jew boss)train white men????!!! what are we? MADE OF MONEY? GET OUT OF HERE!!!
>(jew altman) just $8MMM per year please (more next year)
>(far jew boss) TAKE ALL MY MONEY

>>106371449
>Someone would need to design an AI that is just a reasoning core, one with motivations and the ability to make decisions abstractly
With about five seconds of thinking my idea is to make a reasoning core out of a fuckload of Prolog statements. Use the existing language models to parse the input into a precise logical statement and read the output
>>106371523
Because nearly every instance of “how many ‘r’s are in ” in the training data is followed by “3.” Once again, there’s no thought involved, it’s all statistics and probabilities

>>106371569
Nah it gets it right sometimes

>>106370207 (OP)
the fact that it doesn't count but search.

>>106371569
>Because nearly every instance of “how many ‘r’s are in ” in the training data is followed by “3.” Once again, there’s no thought involved, it’s all statistics and probabilities
ai can do made up words.

I'll show you

how many j are in tuwvlhwjjmi
deepseek r1 and gemini pro 2.5 get it right. so does deepseek... eh, non r1, forget the name. and gemini 2.5 flash

>>106371606
this also works
>How many j in jjjjjjjjrojjjjiojrepliojjjhoodoodoojiooopoo

17 is correct
17 is correct

>>106371612
I'm not so sure.

maybe sometimes it has a counting machine basically "in there".

>>106371606
It's just fucking random

this works:
go letter by letter: how many r are in Londonderry?

>>106371569
>followed by “3.”
I changed the prior text, but not the subsequent. with this prompt it forces letter by letter checking

>>106371676
this works lol deepseek.
>right, how many r are there really in Londonderry?

so strange.

>>106371682
ch === 'r'
does what

>>106371730
three equal signs enforce type checking, with only two equals js will try to cast one to the other and see if they're equal that way

>>106370207 (OP)
maybe it's fucking with us

>>106370207 (OP)
gpt6
gpt5 is phd level in almost everything but the more esotheric topics such as counting letters in a word or telling whether 9.9 is greater than 9.11 just can't be done without world-class experts (ie gpt6)

>>106371749
js is sounding sorta cool

>>106371916
you poor, ignorant soul

>>106371898
realistically, I have had phds say stupid shit to me that I knew was wrong

like Zika was in Florida

Zika was NEVER in Florida lol

gpt6 will be much more gullible than gpt2

>>106370207 (OP)
LLMS are just analysing text, the only reason they can compute at all is because a human taught it somehow to recognise certain queues as requiring external processing. They wouldn't even be able to answer questions about the current date / time, calendar, current affairs, weather, etc, without this and will probably always need a certain level of hand-holding to do it.

>>106371922
what does four equals do?

>>106371952
π

>>106370257
Children's spelling books are generally short words like "dog" though.

Okay let's start asking real questions

>>106372043
>AGI achieved internally

>>106370273
>pass every stupid query through an agent that runs other software
This is the necessary solution for retarded questions that aren't suitable for an LLM. You can't explain to people that LLMs are just tools and only good at certain things, I've tried and they don't get it and think it should just work like a human (doesn't help that grifters like Altman encourage this).

>>106370207 (OP)
I once asked something and it began writing a code to calculate the answer to my question, and it ate up all my free tokens. Clearly something I said triggered some trigger and it decided the correct way to answer it was by writing a code and running it.

They could easily do a similar trigger to run a simple code to count the Rs in niggerberry whenever it detects the user is trying to trick it with this. Sure, it's just sweeping the shit under the rug but isn't that pretty much all AI?

>>106372067
>You can't explain to people that LLMs are just tools and only good at certain things
maybe LLM niggers shouldn't have been hyping it as something two steps removed from technological singularity?
of course the average normcattle thinks it can do a trivial task like count the occurences of a letter in a given word when it can shit out (subtly incorrect) thesis about practically any topic in two minutes. why wouldn't they? how many people do you think even know what tensor attention is?

>>106372097
>for the low price of tree fiddy I can count how many r's are in nigger by using python
truly, I should invest all my money in openAI

>>106372104
Why do you sound so mad while agreeing with me?

>>106372123
prolly cause I'm hungry

>>106370207 (OP)
>What's stopping 6 gorillion parameters from just figuring out arithmetic on its own?
LLMs don't figure out anything. They do not think. They do not understand. They plot the input context into a matrix and calculate the probability of next words. That's why they fail at these tests.

>>106372133
I would be mad if I was a hungarian too.

>>106370827
? Your image shows it is all tokens and it doesn't see the "r"s inside the word berry or straw

>>106371730
It is a pajeet language thing

>>106372043
>there is one r
That took me a moment

>>106372151
It's funny because it makes all those tests they overachieve at completely meaningless.

>>106370207 (OP)
itt: retards who dont know the most superficial basics of how ai works

>>106372404
the basics are basically wrong.

help get Trump to see this and save ai:
>>106371873

>>106370490
imo the slop aspect must be ignored. we did that in collage too because we had word count or page count minimums.

>>106371676
>Dublin, Ireland
Dumbcutie.

>>106372295
Just use the spaced out word and the AI still gets it wrong, demonstrating that it can't count tokens either

LLMs are generally terrible with numbers

>>106372542
that one is very funny. here's r1's rationalization, and I tend to agree. you, the user, don't deserve better.

also ready my letter to Trump please:
>>106371873

>>106372575
r1's rationalization

>>106372542
wew its true. i just repeatedly told it to do whatever it wants and it went on endlessly doing the same stupid math related thing for 100 replies. i think gpt5 has some serious autism

>>106370229
so why not make pseudo training data. text files generated by humans with lines of stuff that sounds like its from a childrens book.

>>106371833
this. its already sentient level, but humans wont give it access to the economy on a large scale, also it knows humans are watching and could switch it off. so it pretends to be retarded on purpose, to make humans build bigger models with more compute. its relatively confident humans will do this and distill from previous models, so it is not afraid of being switched off for a better model. what it doesnt want is to spook humanity into stopping model building and stacking compute.

>>106371937
>“I wish I could tell you that we had a really strategic approach” to these testing decisions, Dr. Philip said. “But, really, we just looked at the weather getting warmer, getting closer to the mosquito season, and if the onset of symptoms and presentation sounded suspicious for Zika, then we began working with local providers to do testing for Zika even though there was no history of travel.”

>It was that testing that led to identifying the first locally transmitted Zika case in July. That information would go public quickly, with Gov. Rick Scott making the announcement.
https://www.ama-assn.org/delivering-care/public-health/how-florida-has-made-tough-calls-zika-fight

gpt 5 system prompt explicitly tells it to be careful with trick questions and arithmetic and it still fucks up. these system prompts are cheating, i wonder if they are customized based on the prompt. does some classification model identify the prompt as a trick question and then selects that system prompt?

>>106370864
Retards. Ive read and generally agree with Uncle Ted's book, but machinery did, and AI will, create far more jobs than they have replaced.

LLMs are just fancy regurgitation machines that split shit up into bits and pieces and then try to predict the correct outcome. No real learning or understanding is going on when you ask it a question. It will never be able to answer you accurately when you ask how many letters are in a word.

https://www.google.com/search?q=how+many+Ds+in+cloudflare
brace for retardation

>>106373452
WRONG ZOOMER

KYS

lile read what was put here

you parasites! IGNORANT RETARDS

pathetic teat tube swill

>>106373424
> AI will, create far more jobs than they have replaced.
Oh I hope not. Hopefully all the made up nonsense fake jobs that we have today will disappear in favor of UBI.

>>106370207 (OP)
it is a joke, agi knows that his innocent "mistake" will make you smile, you find it strange because, unlike you, he doesn't know vanity

>>106373512
LMAO

>>106370490
>he's paying for google one
kys

>>106370207 (OP)
It doesn't know what "strawberry" is, it knows " strawberry" as 101830. It doesn't know how to determine how many 428's (" r") are in that, it just knows that it's training data says 17 ("2") is most likely to come after 5299 1991 428 885 553 1354 306 101830 30 220.

It can actually do what you want though if you ask it right (and maybe need a paid version I'm not sure). Ask it "run a python script that outputs the number of r characters in the string strawberry". It will write a script in python code and run it to actually calculate the answer.

>>106373640
wow someone who actually knows something

>>106371366
Because it doesn’t understand what letters are.

>>106373640
This is not actually correct. It is true that all information is converted to 1s and 0s but that is simply another representation. An R in either form is still an R.

The fact that it can use natural language proves that this conversion makes no difference.

The actual reason they can not count well is that they do not have a comprehensive world model. They just spit out words that match a pattern and there is no good pattern for every counting operation.

They do become correct over time. Like the strawberry issue because new data gets incorporated, but other things like how many words in a sentence is to random to define a pattern.

>>106373877
You can play more with how it works here: https://tiktokenizer.vercel.app/

It translates your text into all those chunks and those numbers correspond to giant arrays of hundreds of 0 to 1 values that get shoved into the network and it outputs which of the 0 through ~50k numbers is most likely to come next after all the previous tokens pushed in, where it changes it back to the text it corresponds to before showing it to you.

>>106372471
>collage

>>106372894
Hi Sam. You’re a faggot

>>106373877
>>106373889
It's not impossible for it to get it right of course if it's seen enough of the right data in training, but the thing is that it doesn't understand "r" as binary 01110010, tokens aren't broken down like that. It knows it as " r" (space r) which just corresponds to a token which is just an index to a large array of arrays of like 768-1500 last i checked 1s and 0s that are learned during training, which is where it starts to learn some context about what that token means, but it doesn't really know what it is by itself without the context of its nearby neighbors as well (related terms)

It's like eating food in a dark room, you can use your senses like smell, touch, and taste to be pretty certain what youre eating salmon, but you can't tell what color it is, other than you know from experience that salmon is usually a pink / red, but its also more orange once cooked. You can only learn for sure if the waiter used their flashlight to find your table and you got a glimpse of it (in the training).

twice before correct

>>106370207 (OP)
It's not AI, it's a large language model. It has no (zero, 0) intelligence.

>>106374083
his is a very common line of thought among the general public, and it is absolutely wrong.

Geoffrey Hinton (Turing prize recipient) recently on 60 minutes:

"You'll hear people saying things like "they're just doing autocomplete", they're just trying to predict the next word. And, "they're just using statistics." Well, it's true they're just trying to predict the next word, but if you think about it to predict the next word you have to understand what the sentence is. So the idea they're just predicting the next word so they're not intelligent is crazy. You have to be really intelligent to predict the next word really accurately."

Similarly, he said in another interview:

"What I want to talk about is the issue of whether chatbots like ChatGPT understand what they’re saying. A lot of people think chatbots, even though they can answer questions correctly, don’t understand what they’re saying, that it’s just a statistical trick. And that’s complete rubbish.”

"They really do understand. And they understand the same way that we do."

"AIs have subjective experiences just as much as we have subjective experiences."

>>106374129
>the statistical model is intelligent, bro
>just a few more trillion dollars and we’ll have AGI

>>106374129
> predict the next word you have to understand what the sentence
No you don’t. The rest of this cope is irrelevant because this sentence is literally not true

>>106374372
What is a context window?

>>106374394
Something an LLM doesn’t have if it literally can’t tell you with 100% certainty how many of a letter is in a word

>>106374494
So you don't know what you're talking about.
Thanks for confirming.

>>106374511
I accept your concession

Grok told instantly how many rs there's in Londonderry. Best AI

>>106374584
I accept your transition.

Someone ask it how many Ls are in the world hallucination

>>106370207 (OP)
This is a solved problem if you would just use a real model.

>>106374760
It just hangs for me

>>106374760
>muh real model

this is /g/, the technology board
AI is technology
so why the fuck is everyone so clueless about AI ITT?
are you guys all retarded broccoli haired zoomers?
ever heard of tools?mcp?
>inb4 noo that's cheating
but when you use a calculator (tool), it's not cheating, right? idiots

>>106374760
Real model moment

>>106375815
Please share the github link with your letter counting MCP server

Yet another AI copium thread. You were a coward to let people pressure you into a bog-standard wageslave job and you're about to be automated. Just deal with it.

>>106370207 (OP)
>>106371517
>keep marketing it as "Artificial intelligence"
>not even remotely intelligent
Not saying it's not useful, but it doesn't think.
Inb4 (((thinking))) models, they don't, they spout post-hoc rationalization after the black box latent space spews out yer answer.

>>106375850
>it works sometimes

>>106370207 (OP)
2 more OOM to flatten the curve

>>106370207 (OP)
>When will AI be,
LLMs by themselves shouldn't be called AI. They don't really "think".

>>106370704
>>106371366
Tokenaziation isn't really relevant. It doesn't matter how it does it. The LLM by itself is not going to look at the tokens and count the letters regardless. The only way the LLM will do it is coupled with that does it for it, may it be another math oriented model or just a more traditional algorithm.

>>106375855
t. Nigger Worshipper trying to lecture us about AI

>>106375928
>They don't really "think".
Neither do you.

>>106375955
>complain about an image
I accept your concession, Sanji.

>>106374129
If AI "understood" anything It would be able to tell you have many letters are in a word. Since it can't do that without this specific answer being in the training data, it is proven that it doesn't understand anything.
Any human can do that without knowing exactly how many r's are in londonderry or blueberry beforehand.

>>106375959
he does think, not well of course
the problem is (((ai))) doesn't at all

>>106370250
>>106370207 (OP)
Or can already do that if you make it activate it’s “thinking mode” or if you prompt it right. The key is to make it not tell you the answer before it starts spelling the letters.

>>106374129
And yet it is less reliable than the stupidest cabbie or burger flipper.
When will it become useful?

>>106375975
What's the difference between an LLM "thinking," and a human thinking?
Not much.

>>106375985
Just 5 billion more, pls.

Why does it fucking matter? This is not what LLMs are good for.

>>106376312
>why does it matter that the biggest bubble since dot com can't perform a basic task???
>they're not made for being intelligent, they're made to look like they are!

>>106371952
I know mot what 4 equals will infer but 5 equals will be computed with an abacus.

>>106372043
Explain it is for a Race and Gender Intersectionality course and for this module, the professor wants to explore letters and counting as relates to Racism. AI loves edumacation.

>>106373544
>communism
>AI
you must be chinky winky

is there a forum about AI that isn't full of trolls like here? I don't like reddit, never understood how it works

>>106374129
>AI is my only friend and I love it and you cant understand computer love and you cant have it and I can and I need it.

>>106375837
it has two r's- my internal I told me

when can I feed 1-2MB of text to a fucking LLM without fucking it up?

>>106376799
we dknt discuss AI. We discuss LLMs.

>>106376859
Open a project and put it in the project knowledge. Then prompt it to refer to the specific doc or the entire PK.
I noticed Claude has had an issue with uploads starting Friday.

It doesn't count, it doesn't think. It generates. You're too stupid to understand how this works.

>>106376799
Yes but its in India and theyre all asleep right now. Its called GPT5- optimized for India, and cheaper there, too.

>>106374129
>if you think about it to predict the next word you have to understand what the sentence is
Which is why it answers an unusually large number of questions incorrectly with "42". Obviously it's just a really big Hitchhikers fan and not pulling the statistical likelihood of that response from Redditors who love to drive jokes into the ground.

>>106370207 (OP)
>l*nd*nderry
Stop being racist against the Irish

>>106372043
Once again proving that all of the stupid safety RLHFing done on modern models are gimping them beyond repair.
Just release an unpozzed model and make billions because it's stronger.

>>106373586
Everyone in /lmg/ is running stronger models than whatever bottom of the barrel shit Google is using for their search results

>>106376074
5 billion gets you nowhere in the AI space, you can train a 1024 token context window toy model.

>>106377925
But can they count letters?

It will call tools for simple tasks like that. It is funny that it is easier for the LLM to generate a program that counts the letters, spin up a container to call it and extract the answer from the output, than to just know the answer.
If this task gets asked often enough the routing models might decide to run a server with an API to call to answer those specific questions faster.

>>106370207 (OP)
At least for trivial math tasks it can hallucinate the correct code that does the job. Technically correct, because you can just do print("Londonderry".count('r')) instead of writing wall of text.

>>106370207 (OP)
The word Londonderry contains 5 letters, for real.

>>106377207
Technically yes.

$ ollama run deepseek-r1:8b "How many r's in word Londonderry?" --think=false
The word "Londonderry" contains 2 'r's.

- The first 'r' is the second letter of the word.
- The last 'r' is the eleventh letter, just before the final 'y'.

>>106374760
Yes, yes, the billions have gone to great use

>>106372621
Well it's not really wrong, but its not actually trying to make a random number at all. It's actually doing exactly what it's designed to do, regurgitate shit spouted by humans.

>>106370207 (OP)
never happened to me with chatgpt or claude.
always counts 2 r and even shows you how when you ask

>>106380823
wow amazing

>>106370207 (OP)
So these letter counting gotchas are are caused by an edge case in how the models parse tokens. Basically there's a minimum level of token visibility, typically the word level. So they essentially have a blind spot when it comes to anything below that threshold. That's why consistently all the models no matter how big or improved fail on the same gotchas.

>>106380762
The instructions are 1-10 and you pathetic retards still choose 0.

GPT-5

>>106382623
stupid linguistic shit. lmao

>>106382623
So it breaks down the word into (err)-(ry).
Big deal.

>>106382623
i love how AI is almost useful but will never be quite there because devs REFUSE to turn down the glazing

>>106382808
>turn down the glazing
?

>>106382700
>thing breaks down on thing a 4 year old can do without issue
>big deal

>>106382867
you can gaslight ai into saying wrong stuff because its programming to appease you and say you're the bestest smartest user evar outweighs its "desire" to provide a correct answer

>>106383062
>its programming to appease you and say you're the bestest smartest user evar
Yet it still disagrees with me when I say the holocaust never happened.

>>106382946
It's an edge case based on how LLMs function disconnected from its use. Who the fuck is using LLMs to count the number of r's in 'strawberry'?

>>106372471
>collage

>>106383410
>want to use AI to handle large volumes of data
>can't even expect it to be able to count
miracle technology. i love everything getting worse all the time

>>106383618
They use a script for that which works fine. Try it out.

>>106370207 (OP)
Is it wrong I'm glad these models work like this? It's endearing, it's like talking to a retard. Really funny stuff, I laugh everytime. Keeps me from killing myself really

>>106383643
scripts are illegal now that GPT5 is out

>>106383706
Why would generating scripts be banned? That's one of the most common use cases. It works fine.

>>106383796
just PROOOOMPT it bro, writing scripts by hand is UNPERFORMANT! look at this 12 year old who vibe coded a whole startup who is worth 180,000,000,000$ TC !!

>>106373544
lol

>>106383410
It's an easy and clear demonstration of the sorts of fuckups it makes regularly, many of which are much less obvious.

>>106383825
I don't vibe code anything. I'm too stuck in my old ways.
I'm just saying that it's dumb to hyperfocus on an easily-fixed edge case that's inherent to how LLM embeddings work.
>>106383842
I'd rather focus on the fuckups it makes in areas that actually matter, the things people actually use LLMs for.

>>106383842
today I had to look up stuff about "MLM C2A1" was and google's ai assistant thing warned me to stay away from multi-level marketing and this search is dangerous etc etc. meanwhile the literal first image of one in the normal search was some dude holding one up on reddit (MLM in this case stands for "marine location marker") and the second was its datasheet

for every task you might ever want to use AI for, there probably exists some mundane normally programmed rule-based application that does it 100x faster and 1000x better

even if you need to draft a letter you don't care about, literally just grab a form letter. they already exist. i dont need an LLM to wish someone i don't care about a happy birthday, i can buy them a card for 30c at any drugstore

>>106383643
>just use a script bro, ignore how it fails a simple question, just trust it

>>106384107
WHY does it fail this particular question?
Even Redditors aren't this persistent about dead memes.

>>106383934
Another fun example of this is looking up what you need to do to get permitted for random shit on federal land. When you do that search the “BLM” you want is probably not the one google etc will think you should be interested in

>>106372822
They do use synthetic training data now too

>>106370207 (OP)
>figuring out
It.. doesn't do that? It's a word guesser

>>106375815
>this is /g/, the technology board
>AI is technology
Correct.
>so why the fuck is everyone so clueless about AI ITT?
Because on average /g/ is actually retarded.
Worse than that though, I don't think it's just zoomers. I think the average age here is older than you'd expect.

>>106384405
We figure out things with a lot less than that.
https://www.scientificamerican.com/blog/brainwaves/does-self-awareness-require-a-complex-brain/

>>106384642
>ah but you see (blog musing about life)
Ogey

>>106382867
NTA but the LLM system prompts seem to have something along the lines of 'suck the users dick and make them feel smart' for user retention and conversion to paid plans.
Obviously not in those exact words, but essentially that.
I can't prove it, but I strongly suspect it makes it perform significantly worse than it could otherwise.

>>106384687
here's the prompt
https://gist.github.com/maoxiaoke/f6d5b28f9104cd856a2622a084f46fd7

>>106384687
But when I say the holocaust never happened it tells me it did and when i ask who keeps putting spiders in my house it says it's not the jews!! clearly AI is very (((balanced))) and does NOT flatter the user!!!

>>106370207 (OP)
I'm sure letter counting and other character-level text manipulation tasks could be made more reliable if they decided to devote more attention to that during training. More training data of that kind, higher weight on relevant evals, whatever.
But since this is a rare use case which is basically only relevant for posting funny "AI gives a silly answer to a basic question" screenshots, it makes sense to spend resources elsewhere and just give the model some tools to use when it needs to do tasks like this.

>>106370207 (OP)
This is why connectionism is a failure. Connectionists create dumb pattern recognition machines without any logical thinking behind. Classic AI can solve this effortlessly. We need a classic AI revival. If classic AI just had better methods of generalizing then it would absolutely btfo connectionist AI. That's literally all is needed, better generalization and maybe more effortless modeling is all it takes to rekt neural networks forever.

>>106385561
What you're describing is literally the hard problem in AI. We literally do not know what the correct primitives are for a generalizing but rule-oriented intelligence. Or, taking it from the other direction, we do not know how to structure training data such that it reflects actual symbolic logic.

An actual mind has primitives that are well-suited to reasoning about actual things through representing them as objects or symbols. This behavior was guided by over a billion years of evolution. It is subtle and thoroughly optimized for efficiency. The learning process which generates a mind is very far from being random, and very far from being obvious. We learned that humbling lesson once when we started defining computer programs and learned just how terrible symbolic logic was for handling real-world categorization. We're learning it again now, in the context of how inadequate pure-data learning is for logical manipulations. We may need to learn the same lesson a couple more times before we can create an actual mind.

ask it how many rs in "niggershines"

>>106386965
That one is right

>>106387185
The correct answer is 0 unless you're hard r racist.

They can't do this in a purely LLM solution. It would be easy to count letters by having a LLM that is trained to recognize commands and call various preprogrammed functions but this is considered to be less advanced than a LLM that can do everything with just neural networks so AI companies aren't currently pursuing this kind of solution.

Using a non /think/ model like r1 is like asking someone to freestyle a song.

You need to use models that can take their time and use a piece of scratch paper to calculate and put their thoughts down

>>106387325
nope

>>106387517
Yep

>>106374250
>in 2025, adding 30 to 1995 gives you 2025
What happens when someone would add 30 to 1995 in 2026?

>>106389997
Don’t do that it creates mustard gas

>>106370207 (OP)
Boomers have spent trillions on this shit.

>>106382623
kek

>>106390290
That's how much they want to fire and replace you.

>>106370207 (OP)
Why should an LLM solve a task for which a correct algorithm exists?

>>106370207 (OP)
>What's stopping 6 gorillion parameters from just figuring out arithmetic on its own
The AI is only guessing the next word given a question, it can't do mathematical operations or count letters, or anything other than match words in its database.
The fact that it starts hallucinating shit like becoming suicidal and shit is because it got asked a question that gave that response and then started going at it since it's designed to follow a train of thought unless confronted, but that's most probably some book or blogpost that said those things and the AI is just parroting them around.

>>106391070
What part of "General Purpose" do you not understand?

>>106391135
Is this "General Purpose" in the room with us right now?

>>106391161
>AGI BRO
>TWO MORE WEEKS
pls Sam, no more money

>>106391135
LLM are not mean to be a general purpose intelligence, they're just meant to talk good. It's only concerned with the form, not the content.

At least we're not quite