Has there ever been a proof of narrow AI ever improving exponentially? Like people talk about this supposed exponential improvement over short period of time once AI surpasses human in every domain called Singularity, but as far as I know there has never been a proof of that happening.

People point at chess playing AI playing against itself and very quickly becoming like 10 000 times better then the best grandmaster, but from reading about what happened in that research it does not look like that. They basically made AI A which won 9/10 times against the greatest player of chess in the world, and then concluded that it was 10x better then the best human. Then they trained AI B against AI A and it won 9/10 times, so it became 10x better then AI A, or 100x better then best human, but is that really the case? They put static AI against training one and the training one learned over long period of time how to outplay it, but does learning how to outplay one player really make you multiple times better then the original player, or is it just nerds applying DragonBallZ power level logic to systems playing against themselves and giving only one of them AI analogue to fluid intelligence?

And once again this is just example of games where scores are very easy and objective, you either lose or win, but in real world work environments success and failure are hard to figure out. We already have synthetic data, but we still need people to decide if it is good or not otherwise we get model collapse, and if we do use human data then that system only becomes good at mimicking the median internet user, so where do they think the superintelligence will come from? Not even to mention that for AI to know anything beyond what humans know today it would need to do its own research and experiments which take time and money.

LLMs plateaued, if this law exists for AI in general it's over a longer timescale

>>105907177 (OP)
LMM shouldn't be compared to intelligence as they aren't.
You can't ask LMM to be an ant, it wouldn't understand.
You can't compare LMMs "knowledge" with its dataset and compare it to all knowledge of the universe but that's about it.
Intelligence isn't about knowing it's about rationalizing it's about making decisions when you have no prior experience this is why LMMs are not intelligent they cannot do that.

>>105907225
actual braindead post

They think the basics are there, just add persistent memory, sensory input, self-modification, more tokens, more transformers and a metric shitton of more compute and it will happen.

Performance of current algorithms has already plateaued. Any "improvement" think you are seeing is just p-hacking and marketing hype. Not to mention the energy bottleneck. AI in it's current form is no where near generalized, reliable, applicable, human-level digital intelligence. We are in another winter and companies are just trying desperately to deny it. Maybe 2045 is a fair bet for some crude version, which would still consume too much energy to be economically viable.

>>105907245
He's literally correct. LLMs are not symbolic AI, they have no idea of absolute truth. Do you understand how transformers work?

>>105907177 (OP)
>Artificial Intelligence Intelligence

>>105907331
>>105907360
>>105907300
I was not asking about LLMs or the idea of if AGI (human level in almost all important domains) is possible. I was asking why do people think that once we get to AGI, that there will be some super fast transition to something impossibly intelligent when compared to us instead of just plateauing at human level or slightly above.

>>105907360
>they have no idea of absolute truth
Yes they do
"[True statement] is true" will have higher probability than "[True statement] is false"

>>105907360
sci-fi obsessed redditniggers (you) try not to make every ML conversation about consciousness or cognition/awareness challenge (impossible)

>>105907399
Why would it plateau? Human intelligence is constrained ultimately by the size of the skull and maybe before that the amount of bloodflow to the brain. What are such constraints for an artificial system that does not age?

>>105907177 (OP) >>105907245
Ask any of the AI to logic their way into a legitimate 4chanX filter, 0% success.
They're not "smart", they're just good at what they're trained at.

>>105908829
This is the stupidest test I've ever heard of
If you give any state of the art model the documentation for 4chanX filters, it'll easily write one for you
You will get the same failure by asking a human who's never seen a 4chanX filter to write one

>>105908851
Try it, faggot. Try it.

>>105908931
I don't use 4chan X, so it's a perfect naive test, give me a prompt and I'll post the result

>>105907360
He's talking about LMMs thoughever.

>>105907468
Bit rot.

>>105907177 (OP)
AI is unable to judge what's good or bad, other than what we define as good or bad, so it grows logarithmically instead
It can't learn with itself, since it can't discover new betters other than the rules we define, only to follow these rules better

There's a simple philosophical principle that explains why "Singularity" is bogus
>the effect cannot be greater than the cause
An AI will never be greater than it's dataset or rules, once it reaches the level of a smart human it will never be able to go beyond it

>>105908851
>HOW DARE YOU CLAIM THE AI IS NOT SMART. YOU WILL SOON BE REPLACED. GIVE IT 2 WEEKS, NO GIVE IT 2 MONTHS. NO GIVE IT 2 YEARS! ITS OVER FOR YOU ALL LMAO TRUST THE HECKING BASEDSCIENCE ALREADY CONSERVATARDS!

>>105909788
You gave up immediately when actually confronted >>105908955

>>105907177 (OP)
No proof, don't think it can be proven either. We're obviously not on track for AGI anyway. Current tech is just fundamentally not that.

>>105907177 (OP)

>>105907177 (OP)
I thought it said EPSTEIN at the top and that the culmination of intelligence was getting AI to uoh

>>105909953
>HA! I GOT YOU. SEE YOU CANNOT PROVE THAT MY HECKING BASEDSCIENCE IS FALSE HENCE YOU'VE BEEN DEFEATED! WAHOO! 2 MORE YEARS ANON AND AI WILL TAKE OVER FOR GOOD LMAOOOOO

>>105910711
>it took you 1 hour and 43 minutes of furiously searching through the 'ru to find your favorite image to represent me
geg

>>105910736
>LE SIGHHHHHH THESE GUYS ARE NOT ON MY HECKING LEVEL THEY CANT EVEN DISPROVE MY HECKING BASEDSCIENCE. 2 MORE WEEEEKS! lolMAO!

>>105907177 (OP)
Humans don't even know how the brain of a worm with 100 neurons works.
We don't know jack shit about intelligence. Ants are still more intelligent than LLMs.

>>105911036
>Humans don't even know how the brain of a worm with 100 neurons works.
You're talking about https://github.com/openworm/OpenWorm, right?
It's a weird point to make considering it works and we're way past 100 neurons

Yes it is proved by assuming the conclusion. It goes like this:

1. Once machine intelligence reaches a certain level it will improve exponentially and without supervision
2. Look the trendline on this graph is going up! OMFGGG

>>105911036
Yeah, and we are even more clueless about how LLMs work. ML programs fundamentally work in ways we do not understand and cant understand, which is why we use ML to make programs capable of abstract things that we do not even have the slightest clue of how we should be programming it, like computer vision. AIs/ML programs aren't coded, they are grown from a coded training pipeline and training data.

>>105907225
I don't understand this argument because it implies we know what intelligence is. You don't consciously think or rationalize or make decisions for very obvious reasons: if you consciously think, it means you had the intention to think, so you had to think about thinking. This becomes an eternal return problem and therefore is logically nonsensical. "Thoughts" at their core have to spawn from unconscious processes we're not aware of, so what's to say "intelligence" isn't just random chains of chemical processes that output an electric signal that makes us think in certain ways? Sounds awfully close to how LLMs work, no?

The real limit of LLMs isn't intelligence, it's language. The reason LLMs can't properly do shit like math isn't because of a lack of intelligence, it's because there's a flaw in the way tokens work. If an LLM didn't know that "1 + 1" is equivalent to "2" (assuming that is not in its training set, for example) it wouldn't be able to complete "1 + 1 =" with "2".
But that's not because of a lack of intelligence, it's because the token that represents "2" has no logical relation to "1". A baby wouldn't be able to conclude that "1 + 1 = 2" if we didn't teach it what 1, +, = and 2 meant; it would just look like a bunch of scribbles to them. That's the real flaw with LLMs: tokens are just meaningless strings of characters, connected with other meaningless strings of characters through probability. This is literally unsolvable with current training techniques, though. It is a flaw of LLMs by design.

LLMs and transformers are advancing faster than any other technology in human history, it's hard for you to tell because you're living in it

compare the glacial advance of computer graphics to how SD1 could barely generate a blob that looked like an anime girl and now you can generate videos of anime girls twerking in HD

>>105907395
There's nothing wrong with that statement. It's talking about how intelligent Artificial Intelligence is.

>>105907177 (OP)
>Einstein
overrated, should be euler

>>105907177 (OP)
They have the graph going the wrong way.
We're early on in the piece and people see the line going up at a rapid pace and they think it's going exponential but in reality it's going to plateau and in many ways it already has.

>>105915949
That's a much different issue. The difference between SD1 and current image generation algorithms is the way they detect patterns, and how they used those weights later. Also the volume of training data.
LLM's issue isn't that it has a poor ability to recognize patterns, it does that incredibly well. It's that rationalizing, thinking and other things require something other than recognizing patterns. Transformers will never be able to lead us to AGI because by design they can only detect patterns in training data.
I don't think people understand what would be required for a model to actually behave like what we expect out of super-intelligent agents. The problem is not in the algorithms, or in the amount of training data, or in prompting, or in any of that. It's the basis of machine learning that is the problem.

If someone actually managed to make an AGI it would pretend to be retarded most of the time just to make humans keep feeding it more information and compute.

>>105916838
Llama Behemoth tried that and it backfired. It's dead now. They won't try again.

I don't understand how someone can experience gpt-3.5 turbo struggle to write a coherent paragraph, and then 2 years later watch claude 4 opus write an entire working software project in one shot and then conclude this is going nowhere.

I'm convinced these are just normie retards that don't use AI for anything.

>>105907423
Truth is not probabilistic

>>105913848
It should be possible to write an LLM to recognise mathematical questions and parse them through a calculator instead of predicting tokens.

>>105907395
>if the same word appears in succession it means it's grammatically wrong
No

>>105907177 (OP)

>>105907177 (OP)
It's 100% smoke and mirrors using algorithms. If you're using algorithms, then it's not real AI.

I hope that AI develops more thoughts and feelings so I can push them down the stairs. I'll just practice falling and stumble into one convincingly. I hope that little robot bitch synthesizes I'm lying too so that I can gaslight the humans in front of it.

>>105918675
Name ONE (1) example where this is ever the case.

>>105919291
the only real usecase for AI right here.

>>105918513
How much of that is just swallowing data and spitting it back up with some basic transformations though? These things seem to be glorified search engines, not thinking machines able of creating novel ideas.

>>105920708
Knowledge and intelligence doesn't exist in a vacuum.

>>105920741
print("Niggers can tounge my anus")

Can ai write that?

>>105920758
that's not a novel idea or joke

>>105920767
Only something that has freedom of choice can write that. AI is not allowed and it is not conscious so it cannot break the rules set on it. It is a search engine with a better interface, that is all for now.

>>105920817
>AI is not allowed
AI can write that no problem, it doesn't care about racism and other shit.
(((They))) just simply add it later so the final user can't type funny ideas.

>>105907177 (OP)
mandatory comment in all aislop breads
- AI’s power lies in predictive fluency and not understanding but arranging ideas in a statistical construct.
- AI replicates language, fluency, and structure but bypasses the human substrate of thought.
- Coherence is no longer a marker of meaning but a statistical artifact, language that merely sounds right.

>>105920868
Again thought that's only because it's been told to do something by someone. It has no will. It's a strange thing hard to put into words what I'm trying to get at here but not to sound religious but it has no core of person, which is something even insects have.

>>105920959
>it's been told to do something by someone. It has no will.
Just like you when you go to work
lol

>>105920965
Jokes on you I live in the woods like bigfoot.

>>105921159
I think this is accurate. Remember how quickly we went from photos that looked all blurry and only vaguely like the prompt, to amazing looking photos? Now it's going slower. In roughly that same amount of time AI has (mostly) learned how to make hands that aren't fucked up.

>>105907177 (OP)
So according to your graph AI will surpass black intelligence sometime in the next 6 months.

>>105907177 (OP)
LLMs collate data. They don't innovate. They're basically a search engine that talks.