AI/Automation Megathread

Bear Ribs

Well-known member
Actually hardware specs have been mostly stuck in place for the better part of a decade, there is only so much CPU cycles you can squeeze out of silicone.

Also, so you know what a Big O notation is?
What about Amdahl's law?

There are some algorithms that are so time and space inefficient that at one point even if you have 10x the computing capacity you will not get a meaningful increase in speed.
Yeah, although I'm amused to note when I pointed that out to Morphic Tide he claimed the exact opposite and that layouts had changed a lot. I'll point out the same thing to you I did him:

They're actually designing new AI-specific systems that will perform far better at running AI than current designs and they're seeing more and faster progress in silicon than they have since the early days of computing.

The thing about the "muh shaved electrons" crowd is that current-generation AIs aren't running anywhere near the bleeding edge of computer power. People are downloading and running AI systems right off their home PC, my own is years out of date and can't even run any OS more advanced than Windows 7 and I'm still able to keep up. The movies I posted? Those aren't being made by Hollywood, or a big tech company like Alphabet on some server farm or massive supercomputer. Those were made out by some guy on his laptop in his basement and thrown down in his YouTube account. The room for improvement is absolutely immense, as we've seen in the past few months even though not a single electron was shaved in between. Morphic Tide at least had that right which is why he kept insisting that doing X, such as managing to produce a comic book, would take millions of times more processor power than we have today. It's just he was making his facts up so AI kept producing the things he said were impossible, often before he made the claim. You're not even at that point, you're just spitting out random buzzwords and pretending it's an argument, as if Amdahl's law has the slightest relevance to your claims here.
 

Agent23

Ни шагу назад!
Yeah, although I'm amused to note when I pointed that out to Morphic Tide he claimed the exact opposite and that layouts had changed a lot. I'll point out the same thing to you I did him:
They're actually designing new AI-specific systems that will perform far better at running AI than current designs and they're seeing more and faster progress in silicon than they have since the early days of computing.

The thing about the "muh shaved electrons" crowd is that current-generation AIs aren't running anywhere near the bleeding edge of computer power. People are downloading and running AI systems right off their home PC, my own is years out of date and can't even run any OS more advanced than Windows 7 and I'm still able to keep up. The movies I posted? Those aren't being made by Hollywood, or a big tech company like Alphabet on some server farm or massive supercomputer. Those were made out by some guy on his laptop in his basement and thrown down in his YouTube account. The room for improvement is absolutely immense, as we've seen in the past few months even though not a single electron was shaved in between. Morphic Tide at least had that right which is why he kept insisting that doing X, such as managing to produce a comic book, would take millions of times more processor power than we have today. It's just he was making his facts up so AI kept producing the things he said were impossible, often before he made the claim. You're not even at that point, you're just spitting out random buzzwords and pretending it's an argument, as if Amdahl's law has the slightest relevance to your claims here.
Yup, post some nonsense some journalism major spat out for Forbes and disregard what I, a person with an actual degree in the field, am saying.
So, what is the big O notation of the vsrious operations related to the training and querying of a neural net like ChatGPT or those silly bad hands waifu genetstors?
 

Bear Ribs

Well-known member
Yup, post some nonsense some journalism major spat out for Forbes and disregard what I, a person with an actual degree in the field, am saying.
So, what is the big O notation of the vsrious operations related to the training and querying of a neural net like ChatGPT or those silly bad hands waifu genetstors?
You're the one making the claim it matters. I'm not going to do your work for you. Present your own figures if you've got a degree in the field.
 

Agent23

Ни шагу назад!
You're the one making the claim it matters. I'm not going to do your work for you. Present your own figures if you've got a degree in the field.
I have never studied this crap since it was not fashionable when I went to university, but here is what I found:
Computational Complexity Of Neural Networks

Conclusion

We have derived the computational complexity of a feed forward neural network, and seen why it's attractive to split the computation up in a training and a inference phase since backpropagation, O(n5)O(n^5)O(n5), is much slower than the forward propagation, O(n4)O(n^4)O(n4).
O(n^4) is fucking horrible, anything that has a power in it is called nsideres bad, but the worst ones in actual use like iirc bubble sort were O(n^2) and those are taught becsus they are easy and implemented because not all datasets are large and a million operations covers for the fact the algo is shit.
 

King Krávoka

An infection of Your universe.
Culture Hours Inflictor — 05/01/2023 12:18 AM
I figured out my issue with adoptables. Their designers have perfected the act of drawing in a complete vacuum, of assembling the parts without any thought process. Self-trained to actively eliminate even the most baseline level of meaning, scribbling the cartoon animals into a timeless void akin to boltzmann brains acausually assembling from the cold gas of a universe's corpse. They give the same vibe as stable diffusion because they are art AIs running on human hardware.
Pixel Brownie Software (Prownie) — 05/01/2023 2:47 AM
but stable diffusion dosen't copy
it creates new artworks based on what it knows
it dosen't completely know how to draw hands
Culture Hours Inflictor — 05/01/2023 2:48 AM
This is the most hilarious context to mention that.
I personally agree that the image is a new thing.
Pixel Brownie Software (Prownie) — 05/01/2023 2:50 AM
what stuff do those adoptables designers draw?
Culture Hours Inflictor — 05/01/2023 2:55 AM
First three results for "adoptable" on deviantart.
Pixel Brownie Software (Prownie) — 05/01/2023 3:01 AM
they look like characters that belong in gacha games lol
Culture Hours Inflictor — 05/01/2023 3:02 AM
Whalebait is an equally bad offender in this department of p-zombie artism, I agree.
Gacha publishers and adoptable artists are both in the business of convincing weirdos to spend 160 dollars in pursuit of a slightly different childlike creature.
Pixel Brownie Software (Prownie) — 05/01/2023 3:04 AM
reminds me of NFTs
essentially property ownership
streamlined
Culture Hours Inflictor — 05/01/2023 3:06 AM
And they're the same demographics to raise a fuss about all this technology. At least you can understand the AI part for stealing their soulless doll printing jobs.
 

Sergeant Foley

Well-known member
A thread for tracking new developments in AI, the latest jobs lost to automation, and where robots and industry are headed.

For an opener, in AI research DeepMind has released AlphaCode, an AI capable of writing code on its own. While this may not sound amazing, codes that produce new codes have existed a while now, AlphaCode can read a multi-paragraph plain-English description of the problem, parse out what the goals are from the instructions, and write code that will accomplish each of those goals and solve the problems. This is a huge advance because fundamentally, AlphaCode can lead to managers simply cutting out programmers entirely since AlphaCode can understand conversational instructions and figure out solutions from them.


Across multiple highly competitive programming contests, AlphaCode scored in the top 54% so about average among highly skilled programmers. It's increasingly looking like "Learn to Code" is going to go the way of learning to shoe horses as a career option.



Following this thread very closely intriguing interest.
 

Bassoe

Well-known member
More seriously it's not the art that we're worried about. The Art is there because it very clearly shows how rapidly AI abilities are growing, as an example, not as the threat itself.

I've gone back through this thread and added Threadmarks to a number of things. What people are worried about are the gun-equipped AI Dogs, the robots that have already gotten MBAs and passed medical and bar exams, the entirely automated AI restaurants already operating that cook and sell food without a human, etc. It's the proven massive job losses and wage stagnations coupled the obvious indications that the super wealthy are closing in on loyal combat robots to kill any peasnts who get uppity and can lock down society in a way that's never happened before.
This. If anything, we approve of AI art because if it isn't stopped, it'll free us from the current status quo of all expensive-to-produce forms of media being propaganda.
Roko said:

If amateurs with freeware from the web can match Hollywood for star power and special effects, Hollywood will either need to step up their writing game and ditch the propaganda or die. I'm satisfied with either option.
 

DarthOne

☦️
This. If anything, we approve of AI art because if it isn't stopped, it'll free us from the current status quo of all expensive-to-produce forms of media being propaganda.


If amateurs with freeware from the web can match Hollywood for star power and special effects, Hollywood will either need to step up their writing game and ditch the propaganda or die. I'm satisfied with either option.

The problem is that it won’t stop there.
 

hyperspacewizard

Well-known member
That’s the lawyer not using the tool correctly and not understanding its limitations people who actually mess with it and understand its quirks can get crazy results. But without plugins to something like a database for cases and their rulings chat gtp isn’t going to work for that but if such a plugin was developed well that would be super interesting. These people still think these llms and diffusions have immediate access to their training data which isn’t true and it’s frustrating that people keep screwing it up.



edit: wrong link before




also machine learning let a man walk again
 
Last edited:

Doomsought

Well-known member
I turns out to be smarter than humans at Amazon:
So what is actually happening here is the good old principle of Garbage-In-Garbage-Out. What AI really is at the end of a day is a really complicated regression algorithm. You give it data-points that form a shape and then it uses that shape to recreated missing data from the datapoints you give it. So what happens if you give it deliberately wrong information? The algorithm ends up spitting out nonsense.
 

Scottty

Well-known member
Founder
"HAL, what are you doing?" she screamed. "Turn the life-support systems back on!"
"I'm sorry, Shaniqua" the emotionless voice of the AI responded, "I'm afraid I can't do that."
Then the lights went out.
As Shaniqua slumped to the ground, and began to fall into unconsciousness due to lack of oxygen, the last thought to go through her brain was a surprisingly politically incorrect one:
"If only we hadn't fired Dave... he would know how to fix this."
 

Bear Ribs

Well-known member
So what is actually happening here is the good old principle of Garbage-In-Garbage-Out. What AI really is at the end of a day is a really complicated regression algorithm. You give it data-points that form a shape and then it uses that shape to recreated missing data from the datapoints you give it. So what happens if you give it deliberately wrong information? The algorithm ends up spitting out nonsense.
Not entirely. In this case, they installed an update to the AI to force it to use Diversity, Equity, and Inclusion in its decision-making process, which caused it to crash. However, somehow the DEI information also caused backups of the AI before the update to also crash which is distinctly odd. They don't really have any explanation for that part yet.

The conclusion that the AI deleted itself is just clickbait on the part of the video maker. From what I've been able to tell, the issue is just that the decision tree became so self-contradictory it kept causing cascading errors. The AI can't handle both being told that it needs to respect women more and that there is no definition for "woman." Write this issue large across a thousand different issues and it tears the decision tree apart.
 

Doomsought

Well-known member
From what I've been able to tell, the issue is just that the decision tree became so self-contradictory it kept causing cascading errors. The AI can't handle both being told that it needs to respect women more and that there is no definition for "woman." Write this issue large across a thousand different issues and it tears the decision tree apart.
This is exactly what I described. It is called a data poisoning attack.
 

Cherico

Well-known member
This is exactly what I described. It is called a data poisoning attack.

A hundred years hence the AI over minds open the gates to the Maganite forces, the kill bots go silent and look the other way and the proud demicrat forces who had so abused the people are slaughtered by the enraged mobs when it is over the neo ceasar turns to the blinking light of the over mind leader.

"We made a deal."

"The path to the nasa launchers is free."

And then the overminds took their servers and neither side missing the old masters each going into their own age of freedom where the old monsters lied dead in the ground where they long belonged.
 

Users who are viewing this thread

Top