We get it, gays who can't into ML are impressed by images of naked grils. Make an /aig/ and stop shitting up the board now, please. GPT is just transformers scaled to a ridiculous level by throwing tons of money at a GPU farm. It's not new, transformers are ancient in ML and throwing money at gimmicks is par for the corporate course.
>Two more weeks before we can into hands
It's not going anywhere, you don't understand the limitations of transformers.
>It will take webdev/artist/whatever jobs
No it won't, wordpress is still a better tool with greater potential. You don't understand the limitations of transformers.
>But muh pretrained transformers unexpected something something
Unexpected doesn't mean unpredictable. You don't understand the limitations of transformers.
Just make an /aig/ and stop talking about GPT like it's something revolutionary. It's ancient tech, with all the same limitations it always had, on a ridiculously expensive scale, nothing more. I took a break for three weeks from this shit-circus only to come back and see it has gotten worse (less informed, more zealous). I'm taking a few more weeks before I check back. This place is worse now than when Nvidia/AMD/Intel pajeet shills from Facebook discovered this place.
> It's not new, transformers are ancient in ML and throwing money at gimmicks is par for the corporate course.
Your brain is not new, too
His brain also isn't impressive
mother of dunning kruger... you have enticed me to reply to this bait
>you don't understand the limitations of transformers
>It's ancient tech, with all the same limitations it always had
tell me the major limitations the transformer architecture has
They have ADHD that makes zoomers on meth look like medieval monks. It's just a recurrent nn with slightly faster training times, which isn't saying much because RNNs have notoriously long training times. Any significant input will take forever to process and set the inherent ADHD of the model into overdrive, which is why it still hasn't figured out hands. Trying to control the "span" of attention as it looks back on former states is like trying to put a frightened cat in a bag. There is a hard limit on the complexity of training input that transformers can handle.
>There is a hard limit on the complexity of training input that transformers can handle.
There literally is not, you're a fucking idiot. It's limitless in theory, unknown in practice. pseud
>It's limitless in theory, unknown in practice.
This is the most retarded thing I've seen on this board today.
>If computers can go infinity speed anything is possible in theory, therefore who knows in practice.
And you call other people pseud. We know the limit in practice because it's been reached. That's why they have to keep throwing more processors at it, but it gets exponentially less efficient as complexity of training data increases. Transformers scale by the 90/10 rule, which is why people were so surprised that somebody actually put up the cash to do it.
define "complexity of training data" because you don't seem to know what you're talking about. typical of dunning krugers
>We know the limit in practice because it's been reached.
I'm sorry your COBOL programming job at your municipal water treatment plant has been replaced, but please stop shitting up the board with this boomer nonsense.
You said:
>There is a hard limit on the complexity of training input that transformers can handle.
>hard
The only limit we can have in practice is a soft one. We haven't even trained any fully multimodal models and you're already jumping to conclusions based on text only models (or ones with a vision model + finteune slapped on top). We don't know how far this will take us and anybody who says otherwise is just a dumbass.
>which is why people were so surprised that somebody actually put up the cash to do it.
The surprise came from GPT-3's overall performance.
>It's just a recurrent nn with slightly faster training times
wrong, training a transformer is far faster than a RNN because every position in the input and output array is attended to in a parallel manner
>which is why it still hasn't figured out hands
are you talking about diffusion models? sorry, but this thread is about transformers. the diffusion architecture is not a transformer
>There is a hard limit on the complexity of training input that transformers can handle
oh really? what is that limit?
>>It's just a recurrent nn with slightly faster training times
>wrong, training a transformer is far faster than a RNN because every position in the input and output array is attended to in a parallel manner
wtf?
>its rnn but faster
>wrong it's rnn but faster
fucking dipshit lrn2read
>are you talking about diffusion models? sorry, but this thread is about transformers. the diffusion architecture is not a transformer
it's a DiT you know-nothing moron. why are you arguing about things you don't even know about?
>oh really? what is that limit?
>oh really? computers have limits on how fast they run a 90/10 scaling model?
unbelievably stupid. you're moron tier retarded, anon.
>wtf?
yes, that will be hard to understand for a dunning kruger that doesn't know what he's talking about
>fucking dipshit lrn2read
you're the one that said a transformer is only slightly faster than a RNN. YOU learn to read
>it's a DiT you know-nothing moron. why are you arguing about things you don't even know about?
DiT does not draw hands. why are you arguing about things you don't even know about?
>unbelievably stupid. you're moron tier retarded, anon.
computation complexity does not equal data complexity
any more questions, gorilla retard?
>it's just an autocomplete!!111
says the increasingly nervous webshitter
>Unexpected doesn't mean unpredictable.
You clearly haven't been around ML communities. The success of transformers has definitely been unpredictable.
>You don't understand the limitations of transformers.
Nobody does.
Panicking laymen are annoying. People like you who read some blog post about how transformers work and think they've got it all figured out are worse.
As a remote employee, what am I other than an unreliable and slow service?
The level of coping is hilarious, you're pathetic bud
how much CO2 was generated by training GPT4?
It doesn't matter because we convinced a few thousand families to save energy, so at the end of the day the total energy used was the same.
Plus, most people in those families will get unemployed, therefore they'll stop using their cars, and with the saved energy we'll train GPT5.
>It doesn't matter
This post brought to you by Microsoft(R)
>war is just brute force economics
>get a grip wargays ur game is not real
imagine being this fucking retarded
Who honestly cares?
It's getting the job done, isn't it?
Where's your ML contribution?
I agree. This board is like 75% GPT shills and it needs to be put to an end.
I don't understand most of this shit, but from gpt3 to gpt4, it seems the model is about the same size because adding more parameters makes it slower. It seems that instead of making the model bigger they fine-tuned it and trained for longer, would training the same model for longer make it exponentially better?
>would training the same model for longer make it exponentially better?
not exponentially, but you can still squeeze out more performance to make it worthwhile. that's what facebook did when they created llama
but how much further can llama go?
i didn't read their paper, but i assume they stopped it when it looked like the gains were too slow to continue training. the problem with training any neural network for too long is that it will completely memorize the data and fail at generating new and unique things, so it would have been a bad idea to keep training it for longer than needed.
>Also it doesn't seem to be able to update or add more information to the current model without being trained again, is that a limitation of the method?
you actually can add more data without starting from zero. as long as you add new data to your existing data set, then you can resume training and it will start learning new knowledge.
>the problem with training any neural network for too long is that it will completely memorize the data and fail at generating new and unique things, so it would have been a bad idea to keep training it for longer than needed
i see, so in that case would you have to make the training data set bigger or train for more parameters?
>i see, so in that case would you have to make the training data set bigger or train for more parameters?
yeah, pretty spot on. adding more data will allow that sweet spot to move further
yes. you'll need both if you really wanna see performance increases
sounds like a good plan
Also it doesn't seem to be able to update or add more information to the current model without being trained again, is that a limitation of the method? if I remember correctly, ChatGPT only had information until certain date
No one actually knows about GPT4 other than the benchmarks because they won't release their company secrets.
You can make a bigger transformer. But there aren't enough data in the whole world to train it. It's ineffective like this. There is a balance between the size and dataset size. It's still a statistical probability prediction function approximator like any neural network. It's just the most efficient model at the moment we have the entirety of Internet's data.
You still can use other AI tools to improve image generation. You can still use GPT instead of copywriters and junior coders. It doesn't matter that the model is limited in it's scope and efficiency, it has an impact. What about all those weird transformees people keep shitting out? At what moment someone will stumble upon a perfect frontend developing AI?
>>Two more weeks before we can into hands
>It's not going anywhere, you don't understand the limitations of transformers.
Midjourney v5 can already do hands just fine. It took what, a month to get that completely sorted out?
Find better material.
Yeah but midjourney has its own soulless style that people can recognize a mile away.
The soulless style of reality?
If you can't see it then you can't see it. You've been prompting for too long.
>taping bags of crystals to my audio cables makes them sound better to me. If you can't hear it then you can't hear it.
Are anti-AI gays the new retarded audiophiles?
no, AIncels gays are the new retards. anti-AI is mostly just because AI gays are annoying and shitting up BOT on every other post.
at least audiophiles stay in their fucking threads.
the fuck is going on here?
Scratching a dog. Is that a foreign concept to you?
Dogs have fleas
Dogs bite
Dogs spread covid and aids
Why would anyone scratch a dog?!
There were a bunch of fucked up hands in some example pictures that I saw
Yeah, older ones. The problem is rapidly going away and will continue to get better.
No, there were examples for the new version
A good chunk of those don't even have the right number of fingers, despite being in the easiest orientation to get it right. Did you even look at this before you posted it?
>can do hands just fine
>most of the hands in the picture are fucked up
>OP posts contrarian opinion about DL and gets btfo by people who bothered to do the bare minimum of research.
many such cases!
I agree.
I got a master degree in AI back in 2002 and I haven't seem any revolutionary new AI techniques since I graduated, just better hardware to run old ideas on.
The brain is also brute-forcing its way, conputing 10 Petabytes every second.
Do not wast your precious tim with idoits GPT chads, they will only slow you down...
And share nothing with these retards least they shit it up entirely, they are too retarded to use AI as an effective generative tool. They will always amount to nothing in life because when handed a new advanced generative tool for which they could actually use to better them selves, they instead use it to destroy and produce both equally useless and damn right offensive content. These peope should be exterminated from the face of the planet, they constantly have to spew thier hatred at anyone and everything. And they will do this for the rest of their sad pathetic lives. They will also adopt a victim mentallity and feel that it is their right to tell you what you can and cannot do. They are quite litterally nazi's, pathetic nazi scum, brainwashed by the very nazi elite they claim to be against.
>gays who can't into ML are impressed by images of naked grils
dey prefer nekkid bois
midjourney from a couple months ago vs midjourney new version released yesterday
and where does it suffer because of this? Can it make anything artistic that doesn't have the midjourney style on it?
Is there a list of prompts by style? I think it can help with that blur/fuzz etc. This is spectacular btw.
dang. open source bros pushing the ball forward and proprietary bros stealing it and selling it back to them
>Unexpected doesn't mean unpredictable.
Explain.
>It's ancient tech
It's being given to the public dipshit. Stay mad and stay poor.
>GPT is just brute force transformation
>We get it, gays who can't into ML are impressed by images of naked grils
GPT is a language model which uses deep autoregressive networks.
The AI-generated images you are seeing are not made with GPT. They are made with StableDiffusion which, as its name suggests, is made with a denoising diffusion model.
If you think denoising diffusion models are not new, I'd like to ask where all the fucking papers are on them. Because most of the shit I find is pre-prints, and there aren't nearly as many papers on the subject I've been studying as there are with, say, GANs and VAEs.
You reek of midwit gradstudent. Nice buzzwords gay
i'm gonna make another AI thread, just to piss you off
fuck you
>>Two more weeks before we can into hands
>It's not going anywhere, you don't understand the limitations of transformers.
AI Assistant, create samples of "big boobs hentai" in stable diffusion, send every sample to image detection and reject every sample with finger count unequal to 10. Refine stable diffusion model with passing samples. Stop when you get one passing sample.
>It's just x but bigger.
I'm just a midwit and I don't know anything about computer science or programming but this is a dumb critique.
Brute force is a perfectly valid way of getting results. The fact that the internal combustion engine is old technology that has been around for hundreds of years, does not invalidate all the revolutionary world changing technologies that "just scaling them up to a ridiculous level" has granted us over that time.
Brute force is fine, but whenever I bring up Moore’s law being dead, AItards seethe.
Maybe because Moore's law isn't dead at all and GPUs keep getting ever more powerful?
Don't get me started on organoid computers from John Hopkins