Thread 106364493

The "AI" bubble popped right? We can go home now?

This is what our dear philanthropic tech geeks poured all their money into.
Good job.
Should have probably found a use case for AI first to give it a direction to go in instead of just having a sassy chatbot.

>>106364493 (OP)
>forms an incorrect opinion
>uses reasoning in an attempt to explain incorrect opinion
>realises that it is wrong from its own reasoning
>re-evaluates its opinion into a better one
I dont know, already seems better than most "people" on 4chan to me

>>106364513
>This is what our dear philanthropic tech geeks poured all their money into.
No, this is what our boomer parents and grandparents poured our inheritance into.

>>106364665
70+10 year old

>>106364513
>Should have probably found a use case for AI first
Companies act like NPCs.

>Zuck and Musk started firing people because of their own failures? We must jump on the train!
>Others are offering WFH? Must jump on the train!
>Doing RTO? Must jump on the train!
>Blockchain? Yes, jump on the train!
>IoT nobody asked for? Jump on the train!
>AI in calculator? Jump on the freaking train!

>this tool is for generating text, it is actually really bad at doing even basic arithmetic
>*uses tool to do arithmetic*
>wtf why doesn't it work

>>106364513
Much better returns than the metaverse.

We are literally on the cusp of the Singularity.

There is no turning back now.

>>106366279
>this tool is a PhD level intelligence
>no, not like that!

>>106364778
take that back

>>106366335
>heh, that guy has a PhD in physics, he must be pretty smart, right?
>but he knows nothing about 15th century Polish history, and because he is not a master of all trades he is useless

It's literally my job to string together "agents" (lmao) to basically automate shit in a certain extremely boring industry. It works very well as long as it's simple enough that it's something that an intern with a big guidebook/database can do.

It's not magic and takes some work the coax the actual requirements from the user, but it's not even hard to get results from an LLM. It'll definitely reduce the number of employees that a company needs for certain mundane tasks.

>>106366367
I know what you're trying to do, but you still wouldn't a history major not being able to count to 30.
Also, math is one of the major metrics used to benchmark these models. Literally the thing you're saying we shouldn't expect them to be able to do is the thing companies are using to say how good their models are.

>>106366411
If you want help with figuring out what 2025 - 1995 is then open the calculator, don't ask ChatGPT. LLMs are for generating text, not doing arithmetic. You are intentionally using a tool wrong just to complain that the tool isn't good at what you're using it for.

>>106366454
>You are intentionally using a tool wrong just to complain that the tool isn't good at what you're using it for.
Again, the tool benchmarked by having it solve math problems.

>>106366479
The math problems they're benchmarked on are there to test problem solving and reasoning capabilities, not calculator style arithmetic. Again, it is fucking stupid to use LLMs to determine if 2025 - 1995 = 30. Use your head or a calculator.

>>106364493 (OP)
Billions of cheap GPUs will flood the used market soon

>>106366512
But the problems still include arithmetic.

>>106366533
Yes, and they're not the point of the test.

>>106364493 (OP)
> People hate Microsoft/Google/Facebook-style AI
> Companies used it for a few months as an advertising gimmick (wow, we're so modern, look!!!) and were ridiculed
> No one uses AI for useful things, it's just for porn or super basic help
> Google/Microsoft could have created virtual libraries or virtual teachers, but no, they prefer to make porn or stupid propaganda for the right and far right
> The scam is becoming more and more obvious with the famous “give me another $500 billion and I'll really make intelligent AI this time, I promise”
>We're razing entire forests and draining huge water reserves just for their crappy data centers.

Yes, the bubble is bursting, like all the other products that tech companies have tried to sell for 20 years but have all failed because they're crappy and/or anti-human.

>>106364493 (OP)
The AI bubble popping would just mean less AI companies but not less AI.

>>106366543
... So the problems that require it to use arithmetic to get correct answers are not the point of the test.
I mean, unless you're going to say that it is supposed to just generate the solutions from its learned data.

>>106366586
No, the arithmetic part of the problems are not the point. The math problems are there to rest problem solving and reasoning, not raw arithmetic calculation.

>>106366593
But it's still only evaluated by it getting the correct solution.

>>106366622
Do you remember how in school math teachers were really adamant about you writing out HOW you arrived at an answer, not just the answer itself?

>>106366635
This. And "AI" is doing exactly the opposite - faking out the way of solving a problem knowing the answer from statistics (which is not always correct by the way). And they call it "reasoning".

>>106366635
I don't, actually. It has been a long time.
So basically you're saying they're giving it points for doing the right steps, even if it is getting the wrong values out.
>how many rs in blueberry
>there are three "r"s in in blueberry
Well it was looking for "r"s, half points.

>>106366657
The point of the tests is to see if they are doing the right steps. We already have calculators, we don't need calculators that operate at much worse efficiency.

>>106366665
You would have a point, if the whole AI testing and scoring weren't automated.
People aren't grading the answers the AI spits out. They have the answers in the script, that's how it is graded.

>>106364493 (OP)
imagine being so autistic you cant recognize autistic humour lol

>>106366676
Source that LLMs are only ever tested for their final output and never ever evaluated on the process they use to arrive at that answer?

>>106366454
So If I want my model to help me code, I must make sure it contains absolutely no arithmetic to ensure it doesn't somehow implode because 2+2 is beyond its phd level intelligence?
The other dude is exactly right. And taking his example, how could you even write a paper on history if you cant grasp the concept of how long ago the second world war was for example?

>>106366755
It will probably handle basic coding arithmetic just fine. However, the more complex the math gets and the more steps there are the higher the likelihood becomes that it will produce errors, unlike calculators which should never produce errors like that.
>The other dude is exactly right. And taking his example, how could you even write a paper on history if you cant grasp the concept of how long ago the second world war was for example?
LLMs don't know or grasp anything. They are still capable of writing papers.

>>106366306
A part time shift at Target would has better returns than Metaverse.
Yes let's get people to buy virtual real estate in a second life clone that barely has an avatar with a working lower body.

>>106366796
>They are still capable of writing papers.
Yes, the markov chain generator is capable of writing papers.

>>106366867
I'm glad we agree that the tool works for its intended purpose.

>>106366796
>include important people
>only adolf hitler and bonito mussolini named

>>106366880
No problem, it redid the whole thing in less time then it took for you to read the first two sentences.

>>106364493 (OP)
i have grown to have a real boggle with the tone of voice the AI overview uses. like i'm trying to do info gathering with a bunch of google searches and each time it inserts itself into my process just so it can offer its own redditified take each time

We're in the cringe singularity and our gods are money and data

LLMs make me want to kms so I focus on my hate for companies.
Aren't they so fucking retarded right guys (WERE IN THE MOST SLOW AND INSIDIOUS WORLD ENDING CATASTROPHE)

>>106366903
You really think a history phd, or a history student for that matter, would forget to mention stalin/eisenhower in the context of world war 2? And would need another try?
If you handed that in as an assignment, you don't get a second chance, so just redo it hurr durr insn't an argument. And this really wasn't even a particularly hard question, there is an insane amount of data available about WW2. Now imagine if you let the "AI" loose on something with less training data, and you can see just how underwhelming it really is.

>>106366991
I didn't ask it for a PhD tier paper on WW2, I asked it for a quick 5 paragraph summary. It's capable of writing in different styles, and if you prompt it correctly then it should be capable of delivering a PhD style paper. I don't know how you can even dispute that it's not good for writing papers when for the past half decade we've heard endless story from academia about how 'cheating' with LLMs is rampant and nobody knows what to do about it, and now even some professors are using LLMs to grade their student's LLM written papers.

>>106366522
Go look on ebay, it's full of delusional idiots trying to get "I know what I got" prices on 2016-era hardware like V100.
If nvidia doesn't buy back the H100s and crush them, they'll still be $8000 on ebay.

>>106367022
>even some professors are using LLMs to grade their student's LLM written papers

they didn't REALLY care about cheating, at least not enough, and now everyone is adapting. monkey see monkey do

it's still bad. objectively, for everyone, bad.

>>106367057
>they didn't REALLY care about cheating, at least not enough, and now everyone is adapting. monkey see monkey do
If you use them properly then it's really hard to catch LLM cheating, you can literally instruct them to include minor grammatical and spelling errors that would be expected of someone in 9th grade.
>it's still bad. objectively, for everyone, bad.
I'm not saying it isn't, I'm just saying that it's a specific tool that's good for a specific thing, even if it can give the appearance that it's good for a long of things.

>>106364493 (OP)
>AI can now replace politicians

>>106367022
I downgraded the tier to history student in my response if you bothered to actually read it. And again a human could have just gotten this result by googling WW2, then looking at the wikipedia page and summarizing that in 5 paragraphs, and would have generated a better answer than this one. And would have probably learned something in the process.
Vs using an llm which you STILL have to fact check because it would have missed including Stalin, but somehow mentioned mussulini even though it ITSELF mentions that he's only of second rate importance.
If we just keep going down this road we will just end up with a bunch of incompetent pseuds spending hours "prompt-engineering" shit that they could have just learned by using the internet or -gasp- reading a BOOK.

>>106366578
>No one uses AI for useful things, it's just for porn or super basic help
The crazy part about this is that despite porn being easily one of the few things people would pay AI for, and pay a crapton at that, big tech is completely averse to the possibility of it even being allowed and will go to several lengths even lobotomizing their AI to make sure it won't make it.
Well except for Elon Meme, to be fair.

>>106367226
>I downgraded the tier to history student in my response if you bothered to actually read it. And again a human could have just gotten this result by googling WW2, then looking at the wikipedia page and summarizing that in 5 paragraphs, and would have generated a better answer than this one. And would have probably learned something in the process.
A bad prompt on my part resulted in a decent summary, better than what most people could write (certainly non-academics) and it did hundreds if not thousands of times faster. I don't know why you think they are incapable of writing solid papers when there's so much alarm from academia about GPT usage.
>If we just keep going down this road we will just end up with a bunch of incompetent pseuds spending hours "prompt-engineering" shit that they could have just learned by using the internet or -gasp- reading a BOOK.
There's a real worry that modern education is turning into just that, yes.

>>106366454
>open the calculator
why can't the LLM do that, wasn't that the whole point of "agentic" AI?

>>106364493 (OP)
https://youtube.com/watch?v=nnyQXRStSCo

>>106367455
Some of them do, but when/why they don't:
>Architecture simplicity: Base LLMs are easier to deploy and scale without modular systems.
>Security: Code execution tools can open security risks.
>Use-case alignment: Many applications of LLMs (chat, summarization, translation) don’t require accurate math.
Straight from the horse's mouth