AI - The artificial intelligence thread

  Quote

New experimental work from Adobe Research is set to change how people create and edit custom audio and music. An early-stage generative AI music generation and editing tool, Project Music GenAI Control allows creators to generate music from text prompts, and then have fine-grained control to edit that audio for their precise needs.

https://blog.adobe.com/en/publish/2024/02/28/adobe-research-audio-creation-editing

  On 2/29/2024 at 1:36 AM, Nebraska said:

https://blog.adobe.com/en/publish/2024/02/28/adobe-research-audio-creation-editing

Expand  

goddamn this is might be the cringiest shit i've seen so far this year. fucking pathetic trash.

  On 2/29/2024 at 1:47 AM, auxien said:

goddamn this is might be the cringiest shit i've seen so far this year. fucking pathetic trash.

lol. what do you mean? those guys are so natural.

wtf so scripted. 

“hello what is up cool? can AI do loops for my videos on YouTube content creator like and subscribe turn on notifications (Ads x2 (unskippable)).”

”yes check me out I will take this orchestral loop and make it longer” *types in “longer”*
 

”wow now that is very cool for the loop to get longer”

Adobe

My friends I normally don't say this but it is indeed literally over. https://ltx.studio/

  On 2/29/2024 at 4:41 AM, logakght said:

My friends I normally don't say this but it is indeed literally over. https://ltx.studio/

It’s crazy how hard they’re going after content creation and the arts. I guess they all hope to license their Ai to streaming services and YouTube makers. 
I think Sean was right when he said in the twitch streams that “there’s gonna be so much shit art..”

it’s already there and is only going to increase the din of noise. 

we're slowly getting to the part where the real humans start training their German shepherds to sniff out the skynet produced human look alikes

  On 2/29/2024 at 4:39 AM, Alcofribas said:

“hello what is up cool? can AI do loops for my videos on YouTube content creator like and subscribe turn on notifications (Ads x2 (unskippable)).”

”yes check me out I will take this orchestral loop and make it longer” *types in “longer”*
 

”wow now that is very cool for the loop to get longer”

Adobe

Expand  

can you raise the intensity of this post?

  On 2/29/2024 at 4:41 AM, logakght said:

My friends I normally don't say this but it is indeed literally over. https://ltx.studio/

I've seen you say it's over several times now ITT. please define "it"

^

  On 2/29/2024 at 7:09 PM, luke viia said:

can you raise the intensity of this post?

please click on link for payment options

- Adober

  On 2/29/2024 at 6:10 PM, ignatius said:

it’s already there and is only going to increase the din of noise. 

the more trash they have, more will making authentic stuff be valuable and sought after

at least that's my hope

  On 2/29/2024 at 7:48 PM, zero said:

I've seen you say it's over several times now ITT. please define "it"

I normally say it because it sounds funny, but this time I mean it for real. It's over.

  On 2/29/2024 at 10:33 PM, cichlisuite said:

the more trash they have, more will making authentic stuff be valuable and sought after

at least that's my hope

https://en.wikipedia.org/wiki/Imagination_age We're so back.

  On 3/1/2024 at 1:38 AM, silentvision said:

 

Expand  

  On 3/1/2024 at 1:00 AM, logakght said:

I normally say it because it sounds funny, but this time I mean it for real. It's over.

ah ok, gotcha. that "it." for a sec there I thought you meant the it referring to the destruction of the entire human race. glad we got that sorted.

  On 2/29/2024 at 1:36 AM, Nebraska said:

https://blog.adobe.com/en/publish/2024/02/28/adobe-research-audio-creation-editing

Expand  

How far we've not come

In case anyone here hasn't seen SORA yet. OpenAIs new text to video model:

And this is the worst it'll ever be

saudi arabia already off to a good start showcasing their male robot mohammed, and his appreciation for the finer things in life. salute!

  On 2/29/2024 at 6:10 PM, ignatius said:

It’s crazy how hard they’re going after content creation and the arts. I guess they all hope to license their Ai to streaming services and YouTube makers. 
I think Sean was right when he said in the twitch streams that “there’s gonna be so much shit art..”

it’s already there and is only going to increase the din of noise. 

I think probably the AI is mostly going to be used this kind of very low effort stock music, graphics and videos. Like the youtubers are using the same few stock music pieces that you can fucking hear everywhere so if they add a bit of variety with AI, then just go ahead by all means. And quality wise it's not going to be a huge loss if some corporate graphics in a Powerpoint presentation are going to be AI generated instead of the most bland stock photos they could find. Just type in "women eating salads while laughing" or "ethnically diverse group of middle-class people celebrating in an office" to generate the millionth image of that type.

I don't think it will have much breakthrough yet as Art that is consumed as Art. Like a music album that someone intentionally enjoys. But who knows? There's already that whole vocaloid scene out there..

As a software engineer I fucking wish I could just let AI generate the code for me by describing the problem, and we actually have an access to to a kind of "corporate" version of an AI at work but it sucks so much at programming. It can only solve the easiest most generic problems you can find in a computer programming class. Like, whoop-de-doo, another implementation of Dijkstra's algorithm but now in Rust. I wish my job was just solving basic programming puzzles all day, but it's actually mostly battling with undocumented idiosyncrasies of different frameworks and systems that the AI has no knowledge of. Well, I can still use it to flesh out bug reports or whatever.

  On 3/10/2024 at 6:16 PM, zkom said:

As a software engineer I fucking wish I could just let AI generate the code for me by describing the problem, and we actually have an access to to a kind of "corporate" version of an AI at work but it sucks so much at programming. It can only solve the easiest most generic problems you can find in a computer programming class. Like, whoop-de-doo, another implementation of Dijkstra's algorithm but now in Rust. I wish my job was just solving basic programming puzzles all day, but it's actually mostly battling with undocumented idiosyncrasies of different frameworks and systems that the AI has no knowledge of. Well, I can still use it to flesh out bug reports or whatever.

I have some pretty good experiences with using chatGPT (4,0) for writing functions so far though! But I've noticed it's a bit of an art to ask the right questions.

Basically, the idea is that ChatGPT could give the right answer/code, IF it is being asked the right question. So you're searching for the right questions. And be careful to not ask too many questions in one session. Because the longer the chat, the worse ChatGPT performs. After a while, it does become rubbish. So basically, if you haven't got a good solution after three questions, you're better off starting over anew with a different approach. As opposed to keep on trying in the same chat. Because it will hallucinate after a while.

A while ago a got an invitation for the Google Foobar challenge (which is a story in and of itself: https://itsmohitt.medium.com/things-you-should-know-about-google-foobar-invitation-703a535bf30f). And without much Python-experience I got ChatGPT to write the solutions successfully. After a bunch of successful challenges, I got asked if it was OK if a Google recruiter would approach me. But it ended there, because Google ended the Foobar challenge (https://foobar.withgoogle.com)

That was just last month or so. Around February. Right around the time Google pushed their *woke* Gemini generative AI. And lost considerable value in the stock markets.

And I guess they noticed themselves it was fairly easy to pass the Google Foobar challenge without much programming knowledge. (I do have a background in AI/Computer Science though. I'm just not into programming. I'd rather have ChatGPt generate the code for me)

Possibly a combination of the two: they need to fire a lot of people, and their particular way of looking for the right people wasn't working because of ChatGPT. 

  On 3/10/2024 at 6:16 PM, zkom said:

s a software engineer I fucking wish I could just let AI generate the code for me by describing the problem, and we actually have an access to to a kind of "corporate" version of an AI at work but it sucks so much at programming. It can only solve the easiest most generic problems you can find in a computer programming class. Like, whoop-de-doo, another implementation of Dijkstra's algorithm but now in Rust. I wish my job was just solving basic programming puzzles all day, but it's actually mostly battling with undocumented idiosyncrasies of different frameworks and systems that the AI has no knowledge of. Well, I can still use it to flesh out bug reports or whatever.

did you try to feed a custom GPT4 actual documentation in PDF form? You can feed it couple of books and make it look in there before answering 

GitHub Code Pilot also just received an update that made it much better and now you can actually talk to it instead of just letting it generate code

  On 3/11/2024 at 3:10 PM, o00o said:

did you try to feed a custom GPT4 actual documentation in PDF form? You can feed it couple of books and make it look in there before answering 

 

GitHub Code Pilot also just received an update that made it much better and now you can actually talk to it instead of just letting it generate code

The problems that arise are from undocumented limitations and it seems a lot of times the company who made the system didn't even seem to know they exist. Sometimes they knew but left the limitations undocumented for whatever the reason. Some use cases we have are so exotic that there are zero hits in Google. It might be no one even thought of trying to use it the way we do.

It just all comes down to lots of detective work and trying to figure out what happens in the proprietary black box when I feed it different things.

I can't use ChatGPT, Co-Pilot or similar open for all AIs at work because of security reasons.

  On 3/11/2024 at 2:42 PM, Satans Little Helper said:

s. And be careful to not ask too many questions in one session. Because the longer the chat, the worse ChatGPT performs. After a while, it does become rubbish. So basically, if you haven't got a good solution after three questions, you're better off starting over anew with a different approach. As opposed to keep on trying in the same chat. Because it will hallucinate after a while.

yeah I had the same experience. I  always restart with a better promt when asking coding questions and almost never start a conversation about it. the worst results are when you tell it to correct itself instead of starting a new session 

caude.ai also got an update recently which made it perform better than GPT. I gave it another spin today and the results are really great so far: https://claude.ai/chat/

  On 3/11/2024 at 2:42 PM, Satans Little Helper said:

I have some pretty good experiences with using chatGPT (4,0) for writing functions so far though! But I've noticed it's a bit of an art to ask the right questions.

Basically, the idea is that ChatGPT could give the right answer/code, IF it is being asked the right question. So you're searching for the right questions. And be careful to not ask too many questions in one session. Because the longer the chat, the worse ChatGPT performs. After a while, it does become rubbish. So basically, if you haven't got a good solution after three questions, you're better off starting over anew with a different approach. As opposed to keep on trying in the same chat. Because it will hallucinate after a while.

A while ago a got an invitation for the Google Foobar challenge (which is a story in and of itself: https://itsmohitt.medium.com/things-you-should-know-about-google-foobar-invitation-703a535bf30f). And without much Python-experience I got ChatGPT to write the solutions successfully. After a bunch of successful challenges, I got asked if it was OK if a Google recruiter would approach me. But it ended there, because Google ended the Foobar challenge (https://foobar.withgoogle.com)

That was just last month or so. Around February. Right around the time Google pushed their *woke* Gemini generative AI. And lost considerable value in the stock markets.

And I guess they noticed themselves it was fairly easy to pass the Google Foobar challenge without much programming knowledge. (I do have a background in AI/Computer Science though. I'm just not into programming. I'd rather have ChatGPt generate the code for me)

Possibly a combination of the two: they need to fire a lot of people, and their particular way of looking for the right people wasn't working because of ChatGPT. 

Expand  

Well, yes, this is kind of what I meant that when you have well defined problems like you would get in a coding challenge or CS class it gives decent answers. But the real world problems when you're developing software and systems are hardly ever so well defined and everything is kind of murky and undocumented and ill-defined, possibly with inaccurate documentation that no one bothered to update. The kind of algorithmic part is usually fairly straightforward, it's all the technical details and buggy hardware and software that are the pain in the ass.