Did ChatGPT4 really write this beat?
ฝัง
- เผยแพร่เมื่อ 15 มิ.ย. 2024
- In which we discover whether ChatGPT can bop, by making it write beats as Python dictionaries and routing the output to Ableton.
Support me on Patreon: / marcevanstein
Check out my SCAMP libraries: scamp.marcevanstein.com/
Take my course on Kadenze.com: www.kadenze.com/courses/compu...
Private Lessons: teaching.marcevanstein.com - วิทยาศาสตร์และเทคโนโลยี
Cool experiment. It's like asking it to translate the notation into a language it hasn't seen before until it was prompted. The results were better than I expected.
I like that 7:54 background url easter egg! Really interesting stuff.
Damn that part at 7:54 is so smooth
the video's great but i'm mostly blown away by the song on the website you linked at 7:53. maybe my favorite song ever honestly, so beautiful and real and sad
Nice try, but I saw the quick text.
@@moadot720 why you gotta ruin it, play along instead
?
Confused.
@@moadot720 you should've played along
It might sound strange, but... thank you for the link on Daniel Kahneman. It's really interesting what you were able to do with the code, of course, but this relatively minor part was the thing which I personally wasn't expecting and which nevertheless appeared to be too freaking well-timed for me.
At 7:53 there was a quick hidden link you could see - it links to a form to fill out next video ideas because the link would be seen by only the best of viewers. BTW using the period (.) key and the comma (,) key you can go forwards and backwards 1 frame at a time.
I saw it too! It's a nice little unlisted TH-cam video.
i thought my screen was broken, also how would you kinda like "decrypt" that flashing into a link?
Anyone else see that URL around 7:53? its like embedded into the background near the side, and I could only see it on lowest brightness on my old iphone in the safari browser, I had to pause and type it in manually. Are we really not talking about this website he linked?! This is insane.
I saw that too! It was so hard to see!
what is it?
@@NoMoralCompassOnMe You didn’t see it?!
Nice try, but I saw the quick text.
@@moadot720 what was it ;-;
The system 2 you are describing can be found in several papers. Chain of Thoughts + Tree of Thoughts solve the problem you mentioned (c.f. th-cam.com/video/wVzuvf9D9BU/w-d-xo.html )
In your prompting, you could probably improve by more often providing more often valid examples for your request (it serves the purpose of few-shot training) you did it in some cases but not all, that would probably prevent the misinterpretation of code / index .
I'll check out the link! And yeah, that's a good point about including examples in the prompt. That said, I've done a little experimentation with fine-tuning with GPT-3.5, and it has a much harder time learning some formats that other formats, so I still think it's important to align output format with its prior training data.
Thanks for the great info.
I was going to bring up something like AutoGPT that uses the LLMs to validate their own work by subdividing the tasks. But now my input seems silly 😅
i remember watching gpt with the melodies and kinda "procedural" it sounded (at least with the one i watched)
generally surprised how it actually got pretty funky with the parameters, i guess that could be because its a lot more linear than something like a phrase
i wonder if melodic stuff leans more into system 2 or what not, im not sure
I can't believe the url at 7:53 worked wow :OOOO. You are a legend
Nice try, but I saw the quick text.
Fantastic!!! you might consider using the opanai's API directly instead of working through chatGPT. doing the same thing through the API might allow for a little more control on your end -- including producing more consistent/reliable results. You've already got the python code, all you need is an API key! Cheers!
You know, I think they just recently added GPT4 in the API, so I'll probably try doing that going forward 🙂
This is actually really interesting! I like seeing the ways it messes up
It's interesting that you talk about system 1 and system 2! There's actually research going into using additional, differently trained LLMs to check the initial layers work and modify it accordingly as a way to compensate for its lack of a system 1, as LLMs are great at generating something with certain goals *and* at checking the work of something they learn, but can't do both at the same time on their own
What's with the weird url at 7:53 ?
I saw that sneaky little URL at 7:54
Glad I typed it in, taught me a whole lot about music!
I love the song at the url at 7:53! really cool!
Gotta love it when ChatGPT hallucinates that 4 = 5 😂
did anyone else go to that link at 7:53? its crazy
XD the url at 7:53 caught me off guard. it's really hard to see but the website was actually really cool
I’ve hit the exact same problem with GPT. Asking it to repeat something almost exactly is just not in its capabilities. And it often doesn’t understand the implications of the changes it’s making.
hidden url embedded in background? 7:53 7:54 the website it brings you to has a bunch of rad tunes on it, and some weird writhing blob of red stuff. cool easter egg.
7:53 love that easter egg. Always like these little nuggets for attentive people. And it's unlisted
wow that link at 7:54 is nuts
Nice try, but I saw the quick text.
I do agree with the comparison. Personally I've called it lack of introspective, but this is also a great way to express it.
An idea I had for ai is this:
What would happen if you made a language model ai that could 1 take input from previous conversations and more importantly 2 have an option to create an output which it uses specifically as it’s own input, and not to interact with the world outside of itself.
Wouldn’t that hypothetically create an at least partially sentient ai? Or at the very least an ai that could do far more than others? I’m not a full on ai specialist, but from what I do know I don’t see why this couldn’t create something really interesting
Yeah, this actually resembles how a human would learn to talk
That URL at 7:53 sure was weird!
I tried inputting the coupon code from near the 8 minute mark, but it didn't work. I guess they all got claimed first? Or is it case sensitive?
Did anyone else see the URL at 00:07:53? It's quite hard to see so it might take a while to spot it.
Holy the scamp logo is so genius
The 1 vs 0-indexing is a great insight, thanks
AT 7:53 THERES A URL IN THE BACKGROUND THAT TAKES YOU TO THE GITHUB PAGE FOR THE MUSIC
really love ur vids
I can't believe how good the website is that 7:55 linked to. I won't spoil it.
Loved your comparison of working with ChatGPT to the two systems
7:53 oml thanks so much it helped soo much
for context theres a link on screen put ur video to 0.25x speed to see it guys
I’m curious if you would’ve seen better results if you didn’t continue the same thread for so long. Pretty sure GPT produces worse results the longer you drag it along the same conversation. Also would be cool to turn this into the new “GPTs” app format!
i thought the "final project" was bouta be the jelly castle melody
2:53 Am I the only musician who noticed the lack of key signature, making the sound not line up with the notation? The notated scale is Lydian (C major but starts on F) while the sound plays F major or Ionian.
WHAT???? the link at 7:54 is fire.
istg the "url" at whatever that time stamp is is just gonna be a prank from their patreon or smth
What's with the URL at 7:53?
7:21 gpt tryna be aphex twin 💀💀💀
The link seen at 7:53 is down.
dude wtf how did u find the link at 7:53 thats so cool
the url at 7:53 is a masterpiece i wanna unsee it and resee it
try using CAPS and exclamation marks!!! to force GPT to take it as a firm rule! :D interesting video!
So it listens better when you yell at it? 😂😂😂
@@emilyrln what can I say, it learns too much from Reddit 🤣🤣
@@stepkurniawanoh, dear…
this comment section full of bots or something wth
Abstraction vs extrapolation
The link at 7:53 is cool!
ode to joy ❌
code to nowhere ✅
Bro what is the url everyone’s talking abotu
7:53 if you go to the Url something fun will happen
What url
@@ngprenteeseriously what URL? Everyone’s time stamping this and I don’t see it
@@GoldJerryGold same bro I legit can’t see the url it’s probably a joke the patreon supporters are doing
8:46 reminds me of jack stauber tunes
Try making rules capitals that can help. Also gpt its closer to the auto correct that a brain so yeah it's skills to make music work different. I personally find ia and how they work very fascinating and i don't get where im going with this. The video was very interesting thanks for the video interesting.
crazy interesting! :)
Interesting link at 7:53
Wolframalpha integration is exactly that system 2 for mathematical problems
The link at 7:52 is nice
wow
guys click on the link at 7:53 its insane
Use all cpaps for orders maybe
heh nice easter egg at 7:52
7:54 👀
4:04 no i used , and .
I don't know what to think about that 2-systems comparison. I mean, we do have 2 systems, but they do not work as a generator with a filter for quality. Veritasium's video "The Science of Thinking" explains it best (th-cam.com/video/UBVV8pch1dM/w-d-xo.html ). They are two different systems, each with their own pros, cons and behaviours. System 1 (S1), when needed, informs System 2 (S2). S2, with insistence, can change S1.
S1 is not a generator, it's an autonomic processor of incoming information based on previous patterns engraved by S2. It's FAST and simple, cannot catch mistakes. Almost like a system that can choose a couple of relevant small functions (written by S2) on a LUT and quickly apply them in the needed order. Choosing the wrong order, the wrong functions or the wrong arguments can and WILL happen (see Veritasium's video).
S2 is not a filter, it's a slow, methodical, expensive, logical processor. It does catches mistakes. Almost like a system that, every single time, has to write its own custom function, evaluate if it's correct, try again, repeating until it finds the correct "code". And when that happens, needs to rewrite and refactor in several functions so they can fit on S1's LUT. And writing it there is expensive, slow and needs a lot of repetitions to make sure it was correctly written, and done so to a long-term memory part of the LUT. Of course, it still can use the functions at S1's LUT, but much less likely to use make mistakes on its choices, ordering and arguments.
We also have two modes of thinking: diffuse and focused. These modes, as far as I understood, only occur on S2. S2 is the creative one, capable of learning, error-checking, etc. Diffuse mode is creative, exploratory, almost like trying to create a simple and dirty prototype so you can better understand the problem space. It could even allow for some connections between completely separate, seemingly unrelated, already learned stuff. Focused is when it is full on coding mode. Actually sitting down, trying to write the correct code, being efficient with time and space complexities, then refactoring and placing it in the S1's LUT.
____
S1 / o \ S2 diffuse
o | ↕ |
\ o / S2 focused (only conscious part)
-----
With, of course, S1 and S2 are always running in parallel. S2 focused and S2 diffuse do not. You're either running with S2 on focused mode or S2 on diffuse mode.
So, your comparison is somewhat accurate but calling them generator + filter is not correct. Maybe S2 diffuse could be seen as a generator? Sure, but I think it's an oversimplification of what it actually does.
DISCLAIMER: I love to read and study about psychology, neuroscience and philosophy. But I don't have a degree in any of them. I am but a really curious computer engineer. So take this with a grain of salt, as I just may have written the most polished turd ever.
Interesting! Where did you read about the diffuse/focused thing? Because as a composer/programmer, I run into this *all the time*. It's hard to switch between the diffuse, creative thinking needed to make art and the focused logical thinking needed to code.
If you have time, I'm curious what you think of my experiment trying to write a real piece with ChatGPT: th-cam.com/video/GM07a760gow/w-d-xo.htmlsi=fv2ZtrPcSryHgoo2 It was definitely a challenge trying to get it to be, for lack of a better word, creative.
I agree with the comparison with system 1 and system 2, that ChatGPT is system 1 only, and no system 2.
However, Kahneman's a bit stupid. There's also a system 3 that's even slower that provides the calculation space for system 2.
System 3 is actual intuition, while system 1 is better described as instinctual.
System 1 is quick estimations; system 2 is the shaping of those estimations into answers; system 3 is the space used to shape the estimations, which is also the space used to estimate, which is controlled in a deep complex way.
The brain is an advanced analogue computer. All components are analogue, not digital. As such, the most influential thing is analogue tuning of the analogue interpolation algorithm.
System 3 is analogue tuning of analogue space interpolators. System 2 is connections tuning. System 1 is usage, or direct input to direct output.
To presume that the brain doesn't use analogue tuning interpolation is *profoundly stupid*. It's akin to removing a dimension of the brain. It's like believing that AI is in any way close to human-level thoughts. I know lots of people believe that, but that's only because they don't understand any of the basic principles involved. For someone to have figured out 2 of the 3 basic principles involved in the brain, but completely missed that there's a 3rd is... well it's poetic in a silly way. (there's also a 4th system, but modern psychology is nowhere near figuring that out yet).
Or to put this into AI terms; system 1 is the input and output variables. System 2 is the evolutionary training algorithm and training sources. System 3 is the intelligence of the creator, who doesn't know anything about system 3 and assumes system 2 is enough.
I wonder if the reason GPT violates the logic of music even when given hard constraints is because of the “hallucination” factor
So I guess there’s a tension between GPT feeling organic and original and it’s ability to follow basic rules
7:53 dont act like I can't see that URL
12:18 liga bals
11:42 "If 5 machines can produce 5 widgets in 5 hours, how many hours will it take 100 machines to produce 100 widgets?"
12:05 "You'll realize that each machine takes 5 minutes to make a widget, so 100 machines will make 100 widgets in 5 minutes"
If it takes 5 hours for 5 machines to produce 5 widgets, doesn't that mean it takes each machine 5 hours to make a widget? And so 100 widgets by 100 machines would take 5 hours?
Yeah it looks like he switches it from hours to minutes for some reason.
chatgpt should always follow the rules, make him understand that somehow
imagine teaching someone from a nation with no technology and who lives in a completely different environment. you have no analogies, you have to explain every basic that thing has for him to understand and learn. chatgpt progressed a lot from that point, but still has a long way to go until he understands every basic
hooooly 7:53
0:15 YOOOOO HAHAHAHA
SICK AS FUK
guys is this guy trying to rickroll us there's clearly a rickroll link at 7:53
Hi
no one else saw that at 7:53??
7:54 there is a secret
I like 6:34
The core issue you're running into is that chatgpt doesn't understand music. It knows which 4 letter sets to string together to give you a plausable answer, but it doesn't know how variables like complexity or intensity relate to eachother or how to make intensity better or what intensity even is. It can make shots in the dark really well, but that's about it.
Holy shit, theres a URL at7:53, 7:54 ish. Why did you hide something so cool in the background, everyone go check it out
You listen to J Dilla?
Well, I read the book (mostly)! And so I listened to some through that. Any favorites?
if you want a language where shit starts at one try lua
thanks for the url at 7:53, it was awful
Or you can just skip the LLM and make your algorithmic music in a domain specific language. Given that you have to spend effort to spelunk encoded musical knowledge out of the massive mess of statistics that is ChatGPT --- knowledge that you seem to have --- you could code that knowledge in reusable functions in the language of your choice for less efforts and more reproducible results, i.e. you would control the source of randomness instead of leaving it to the implementation of ChatGPT.
Dude, don't do that. Like just don't. Do you know how long it took me to find the hidden text at 7:53.041667???? Single frame inserts are unfair. o.O Really interrupts the flow of a video 'cause I'm having to go back, press replay, and pray I find it based on a visual twitch in my peripheral vision.
AI generated breakcore?
AI is starting to implement a "System 2" very cool stuff we got these days
7:54 god damn
At 7:53 there was a quick hidden link you could see - it links to a form to fill out next video ideas because the link would be seen by only the best of viewers. BTW using the period (.) key and the comma (,) key you can go forwards and backwards 1 frame at a time.
i saw it too