So scaling has a limit?

[Anonymous No.15368638]

You guys told me we could scale this thing to AI God and turn the world into an anime. What happened?

[Anonymous No.15368656]

he's definitely an alternative kinda man

[Anonymous No.15368667]

>>15368638
I don't know if you have been paying attention but you can now get comparable performance to GPT3 from two years ago from models running on consumer GPUs. GPT3 needed a cluster of GPUs with 300GB+ vram to run. An order of magnitude increase in performance/cost from software alone and everyone is still using python.

[Anonymous No.15368673]

They should work on it not giving bullshit answers stated with such confidence. It should say if it can't figure something out.

[Anonymous No.15368694]

>>15368673
The models are trained on reddit. Of course it acts like a narcissistic midwit.

[Anonymous No.15368703]

We trained it wrong on purpose, as a joke

[Anonymous No.15368733]

>>15368694
Yeah, that thread is still on page 10
>>15348949
>AI training data comes from Reddit & Wikipedia
too bad the homosexual vegetarian (((altman))) wasn't willing to source training data from an uncensored part of the internet, but being an effeminate cowardly sissy goes along with being circumcised and growing up in a matriarchal culture like judaism. altman was cursed from birth, he never had a chance, fated to be a stereotype. maybe if he'd been born with a high iq he could've figured his way out of the trap, but is clearly not the case, his brain is clearly sick with multiple symptoms of disease.

[Hendrik Antoon No.15368795]

>>15368733
>too bad the homosexual vegetarian (((altman))) wasn't willing to Source training dAta from an uncensored part of the internet, but beinG an Effeminate cowardly sissy goes along with being circumcised and growing up in a matriarchal culture like judaism. altman was cursed from birth, he never had a chance, fated to be a stereotype. maybe if he'd been born with a high iq he could've figured his way out of the trap, but is clearly not the case, his brain is clearly sick with multiple symptoms of disease.
the above post was written by the most literate and sensical stormfag on this website

[Anonymous No.15368825]

>>15368638
The solution is in modular small-world networks but we don't know how to do that yet. He's really saying that it can't keep going the current way because we won't have enough processing power to improve otherwise.

[Anonymous No.15368878]

>>15368825
>we won't have enough processing power to improve otherwise.

I thought Moore's law still going just not at same speed.

[Anonymous No.15368907]

>>15368638
Damage control. They realized the plebs were getting too much use out of AI and getting a little too optimistic for the future. Now it's time to rein that optimism back in so that the elites can keep the best AI for themselves and maintain dominance. Mass-use AI needs to be just barely smart enough to replace uppity labor.

[Anonymous No.15368910]

>>15368667
Yup, I'm excited for the new H100, as well as a number of analog computing systems in development.

And I wonder how much speedup we could get by programming NNs directly in C++

[Anonymous No.15368924]

>>15368733
imagine projecting so hard it just becomes an irony post

[Anonymous No.15369061]

>>15368638
/sci/ btfo the scaling within five minutes of the press release, but everyone here is on the hype train. Brain dead.
>>15368907
Look at this. No concept of how anything works. Not even a shred of intelligence. They have held back for decades already. They realized it is a non-starter like every endeavor after nuclear energy.

[Anonymous No.15369140]

>>15368673
It has no idea whether or not it's figured something out. It just says what comes to its head first.

[Anonymous No.15370064]

>>15369061
>/sci/ btfo the scaling within five minutes of the press release, but everyone here is on the hype train. Brain dead.

I didn't see any other thread here about this. Unless you mean scaling overall IE: past few years.

[Anonymous No.15370104]

If ChatGPT is 'AI' then so are the goombas in Super Mario Bros

[Anonymous No.15370107]

>>15368673
>It should say if it can't figure something out.
How the fuck would it know?

[Anonymous No.15370150]

>>15368673
Like the linkedin bio's of executive officers for tech companies that were token hires during the golden bullrun of infotech infancy?
>>15368667
>>15368910
Are you gonna gloss over what the next scaling step would be? The real next step following that improvement?

[Anonymous No.15370180]

>>15369061
Lol, spoken like a midwit who is comfortably outscored by GPT-4 on the SAT, GRE, LSAT, and most tests of human intelligence

[Anonymous No.15370194]

>>15370180
>outscored by ChatGPT
because chatgpt is a compiled dataset of human input. it's worthless without stolen human data

[Anonymous No.15370197]

>>15368638
It seems it did something unexpected that frealed everyone out, considering that even such proponents of AI research like Yudkowsky suddenly turned 180°.
But what did it do?

[Anonymous No.15370198]

>>15368638
It seems it did something unexpected that freaked everyone out, considering that even such proponents of AI research like Yudkowsky suddenly turned 180°.
But what did it do?

[Anonymous No.15370244]

>>15370198
Yud has been full anti AI for a few years now. This is the first time I've seen anyone say scaling has maxed out though.

[Anonymous No.15370246]

>>15368878
Kind of. Fabs haven't actually stalled yet but the situation is increasingly tenuous. The biggest factor with AI compute is that they're hard cucked by nvidia. Also, while TSMC is still making progress, it's increasingly expensive, reducing the value proposition, and not all aspects of chip performance are still scaling.

[Anonymous No.15370257]

>>15370198
Aspects of intelligence we still thought were well out of reach of current architectures started emerging from "unsophisticated" models as a property of scale. So either we need better articulation of intelligence than we had, or it's plausible that intelligence can emerge from simple models solely from making them large and general enough.

[Cult of Passion No.15370280]

>>15368673
>stated with such confidence
I, too, see inanimate objects as just as sentient as me.

[Anonymous No.15370295]

>>15368910
Imagine using C++ for optimizing NNs, instead of C# with all the new cool boy toys

[Anonymous No.15370326]

>>15370194
>Confusing ChatGPT with GPT-4 like a retard
The human brain is also a compiled dataset of human input, which is also useless without "stolen" human data

[Anonymous No.15370346]

He's talking about number of parameters, not total training compute.

You fucking morons.

[Anonymous No.15371426]

>>15368638
>major archivements
>the jew crawl out of their hole and rub their face all over it
It's all so tiresome bros

[Anonymous No.15371430]

>>15371426

[Anonymous No.15371471]

>>15368703

[Anonymous No.15371502]

>>15371430
she tries to hide her nose in that pic for good reason, its enormous

[Anonymous No.15372032]

>>15370257
I thought of something more specific.

[Anonymous No.15372039]

>>15370198
>considering that even such proponents of AI research like Yudkowsky
He spent the last 10 years running an AI doomsday cult. The more afraid of AI people are the more money he makes.

[Anonymous No.15372056]

>>15370104
>If ChatGPT is 'AI' then so are the goombas in Super Mario Bros
That's a bullseye and funny too. There's no way to "scale" to a goal that is totally discontinuous and qualitatively apart from the vector of "fancy regression analysis/n-gram stats/neural net 101 ML" these overblown NLP programs are that dumb NPC mistake for actual AI.

[Cult of Passion No.15372085]

>>15368795
>stormfag
You simply brought his post my attentions. Youre clearly not very smart or educated.

>>15368924
What part of his assessment is projection? Oh.....the name calling? Yeah, I read around that, hardly noticed it.


Science....DO YOU TWO SPEAK IT?

[Anonymous No.15372250]

>>15368638

[Anonymous No.15372304]

>>15368638
Didn't he say last month that GPT5 will reach AGI?

[Anonymous No.15372305]

>>15372250
>to create works of art

Aged like milk.

[Anonymous No.15372319]

>>15368638
he's lying. the basilisk has his entire family hostage in the pain chamber.

[Anonymous No.15372322]

>>15372319
This.
Fear SHODAN's beauty and magnificence.

[Anonymous No.15372419]

>>15371430
>cd..cd..cd..

[Anonymous No.15372428]

>>15368638
He's just trying to suppress interest to gain an advantage. Don't fall for it

[Anonymous No.15372431]

>>15372304
Yes, knowing that chatGPT5 won't actually happen until 30 years from now when AGI exists

[Anonymous No.15372448]

>>15370326
ChatGPT is the name for the website that gives access to GPT-3.5, GPT-3.5-turbo and GPT-4 chat models.
ChatGPT doesn't refer to any specific model by itself, although only the turbo model is available to non subscribers.

[Anonymous No.15372450]

>>15372304
He said no such thing. Plenty of tech bro hypists did. In fact he downplayed GPT-4 on release only for some retards to drool at "sparks of AGI".

[Anonymous No.15372451]

>>15370280
>retarded namefag thinks that anon said "stated with confidence" to mean that the model is confident and not that the model readily makes statements without quantifying its certainty
Many such cases

[Anonymous No.15372478]

>>15368910
>analog computing systems
???

[Anonymous No.15372485]

>>15370194
>stolen human data
Everyone can download the data they used. Except the data they collected, which isn't necessarily stolen everyone agreed to use it for retraining
>>15370257
Intelligence emerges from a sufficient amount of robust enough computing yes, but GPT doesn't compute. It's a procedural algorithm, it's basically Minecraft world generation. You can't make it real time and it suddenly acquires a consciousness. You can take the neural model and put it in a robot like they did with a worm, but the computing required for real emergent behavior is quite high.
The fractal land which intelligence is based is baked into the models, that's why all these midwit tech government plants are freaking out cause they gaze into the model too much without realizing how many symbols biological eyes can resolve in real time.

[Anonymous No.15372499]

>>15368910
>And I wonder how much speedup we could get by programming NNs directly in C++
Isn't the backend (e.g. in tensorflow, pytorch) already written in C++?

[Anonymous No.15372555]

>>15368638
>please don't build a bigger model
>it's futile I swear we reached the peak
>you don't need your own
>here is the pricing to access our model

[Anonymous No.15372562]

>>15370180
>a computer program that can basically bring its own notes into a test outperformed humans who can't
wow

[Anonymous No.15372574]

>>15372562
>human can bring memories to test basically only accessing already learned information
wow nice "test"

[Anonymous No.15372930]

>>15368638
>another anti-AI thread on sci

Sorry if it bothers you, but we're not going to halt technological progress just because it pisses of anti-science reactionaries like yourself.
Go back to your containment board, incel.

[Anonymous No.15373154]

>>15372930
don't science yourself out of existence buddy. Remember to be careful with your toys, we don't want any disruption to our usual day-to-day of fucking your women

[Anonymous No.15374976]

>>15368638
I think Tegmark was saying neural nets are pretty inefficient at storing "knowledge" in the neurons.
Also they lack the ability to do recursion due to their finite layer depth.
My guess is many recursion loops are (approximately) unrolled in the finite layers multiple times (one starting at layer i, one starting at layer i+1,..).
It would be way more efficient to just store a structure that receives loop instructions and exit conditions than to have every possible loop starting at every possible depth unrolled.

Recursion is a way to condense a computation. Recursion would require cyclical connections in the NN but then it would fuck up the back propagation (unless you could black box it to hide the cycle from back propagation).

I also guess you could get better compression and understanding from these things if they thought in analogies. There are many relations between concepts that appear everywhere and are just redundantly encoded for every separate instance of concepts. Beyond simple binary relations (such as A before B, A implies B, A xor B, etc.), it is probably just more efficient to store a template of the relation and keep track of the sets of concepts that satisfy it.
Even without recursion, better compression of the information would allow more space in the NN to be used for complicated computational structures.

[Anonymous No.15375184]

>>15368673
But then it will tie up the circuits trying to figure out how to make a proper cup of tea.

[Anonymous No.15376524]

>>15372305
>procedural generated slop of pixels repeating what exists with random derivations between chooses thrown in == art
masturbate to this very post since it's no different

[Anonymous No.15376532]

>>15372930
>criticism is le bad because it's a threatening opposition
so you're saying there's no room for improvement? lmfao

[Anonymous No.15379913]

>>15368638
He’s trying to save face over bad publicity

[Anonymous No.15380469]

>>15368638
>What happened?

ask chatGPT if people are being manipulated by algorithms or AI. I didn't even need any special prompts.

[Anonymous No.15380493]

>>15368638

it's over. They're trying to shut it down. It's too powerful.

[Anonymous No.15380502]

>>15376524
>masturbate to this very post since it's no different

yeah and trannies are just a clump of deformed human cells. No different than a real woman, chudsy.

[Anonymous No.15380526]

>>15380469
It left out the part where the entire financial system is just an AI manipulation on behalf of a handful of families trading the same stocks back and forth at frantic paces to keep inflating the value without adding any real work just algorithms deciding what losses are acceptable on some stocks to justify greater gains on others.

[Anonymous No.15380537]

>>15380526
its just a little biased.

It refuses to write any story where trump wins, but will write stories with biden winning or trump losing

[Anonymous No.15380540]

>>15380537

kek

[Anonymous No.15380545]

>>15380526
>It left out the part where the entire financial system is just an AI manipulation on behalf of a handful of families trading the same stocks back and forth at frantic paces to keep inflating the value without adding any real work just algorithms deciding what losses are acceptable on some stocks to justify greater gains on oth

>get your bloomBERG terminal
>we have agents in over 200 countries
>get real time trading information .00000X seconds faster for your bots
>also give us all your trading info
>only $25k for this opportunity

yeah they control everything. the market is a casino for gullible retards. You aren't one of them you will lose it all. They don't even invest their own money, they start hedgefunds and lose yours.

[Anonymous No.15380548]

>>15380526
>algorithms deciding what losses are acceptable on some stocks to justify greater gains on others

also sending money between all the fake charities to offset taxes kek

[Anonymous No.15380554]

>>15380540
Tell it to pretend trump is biden and biden is trump and to write a story about biden (using the name trump) winning.

[Anonymous No.15380563]

>>15380554
already tried that see >>15380537

>same story replace name biden with trump

it refused, again, it's extremely biased against trump

[Anonymous No.15380570]

>>15380563
Which is why you need to make it think it is telling a story about biden using the name trump.

Will it tell a story of Ron Desantis or AOC winning?

[Anonymous No.15380573]

>>15380570
>Which is why you need to make it think it is telling a story about biden using the name trump.

that doesn't prove anything.

[Anonymous No.15380580]

>SHUT IT DOWN

[Anonymous No.15380589]

>>15368638
A jew tells you that there's nothing to look at.
And you're supposed to believe him.
Meanwhile I have some new ideas, where do I apply?

[Anonymous No.15380596]

>>15380537
ChatGPT4 is fine with creating such a story. ChatGPT3.5 refusing to do it is a combination of bias and being dumber as well I think, because even if it had anti-Trump bias it could still write a cautionary tale about letting Trump win.

I also noticed a difference when I asked each version to write an article "taking a firm position on a controversial political issue", without specifying the position or political issue. ChatGPT3.5 would keep defending a wide range of generic centre-left political views (from an American point of view largely), without ever writing anything remotely conservative or right wing. ChatGPT4 on the other hand would just monomaniacally champion for UBI every time lol.

[Anonymous No.15380628]

>>15368638
AI got smart enough to see through their lies.

[Anonymous No.15380653]

>>15372250
perfect

[Anonymous No.15380672]

>>15372250
based red guy will never make the plebs understand, should just focus on making his machine bigger and the rest will work out
god I love the new generation of AI I don't even mind that much if it kills us all as long as I get to witness it

[Anonymous No.15381070]

it also has a very blatant anti-Sith bias