DISCUSS reaktor developer glass ceiling shattered with chatGPT superpowers

wanted to start a new thread to put out a general alert on a (possibly late to the party) realization of how revolutionary this new AI stuff stands to be for our relatively niche corner of the internet

one fairly regular type of post here (and the old forum RIP) is someone posting a link to a research paper and asking around hoping someone has even the faintest clue about how one might go about implementing in reaktor, often times no one does

well, it took someone posting a paper related to my main area of interest to finally understand that now pretty much any research paper from your heart's desire can be rendered legible. at least in my case it has no trouble turning the algorithm from the paper into code and then explaining that code to arbitrary detail until it makes sense how one would go about building in reaktor. it even seems to have some ability to break down in natural language how one might build it step by step in core (though it does tend to hallucinate)

this is huge!! the future is now!

Find more posts tagged with

Comments

nanotable

.ens or it didn’t happen! 😛

ANDREW221231

https://community.native-instruments.com/discussion/comment/61075#Comment_61075

already basically at the last step!

colB

https://community.native-instruments.com/discussion/11831/discuss-reaktor-developer-glass-ceiling-shattered-with-chatgpt-superpowers

this is huge!! the future is now!

Hah, chatGPT is already old tech. The paper released a couple of weeks back about GPT4 is scary! Fully general purpose (not just a language processor), can teach itself to use tools, and has already developed theory of mind!

ANDREW221231

https://community.native-instruments.com/discussion/comment/61109#Comment_61109

yeah i've been following this stuff pretty closely but have been slow to use it for much, especially anything technical. have you used GPT4? seems we're pretty much closing in on human level in many domains and if progress stopped right now what already exists is enough to completely transform civilization going forward. only moderately existentially terrifying. the theory of mind thing is especially crazy since it seems to be emergent just from scaling

still even with just the base 3.5 model im floored at the ability to summarize and breakdown academic papers. chatGPT doesn't have access to the internet but since the paper was from 2017 it had the entire thing memorized. though guess it remains to be seen whether the instructions its giving me will lead to an algorithm working as advertised

Cal Scott

I think GPT4 is multi mode now, meaning it can look at visual stuff you feed it, and also generate video, maybe it could be used to suggest variations on an existing ensemble structure. But it needs the extra mode of listening before it becomes a true master of the Universe...maybe thats gonna be in GPT5.

Still its mega powerful already and has given me instructions to generate millions of samples from my collection of 20000 reaktor one shots... party on.

I had a laugh and told it i was way better at building in reaktor and it replied... yeah, but can you calculate Pi to one million decimal places...so funny!

ANDREW221231

https://community.native-instruments.com/discussion/comment/61122#Comment_61122

yeah someone else was saying they were using it for their midi scripting which had improved their music

as far as the multi modal thing.. I was thinking 2d spectrograms as pictures could be one way to input rudimentary audio input:

is the multi modal picture input released/implemented yet?

I had a laugh and told it i was way better at building in reaktor and it replied... yeah, but can you calculate Pi to one million decimal places...so funny

when the AIt starts giving lip 😬

Cal Scott

is the multi modal picture input released/implemented yet?

I think so but can't confirm. Might just be in beta testing. I cant access the official one due to location, so i use derivatives that dont need a key, but they tend to be limited. Probably some in the making as we speak though, i will let you know if i find one.

Spectrograms are good idea for feeding it for sure. I am looking forward to translating some wild chaos research papers into core if possible.

Some things will need a dataset first so the model can it train itself first.

Why did the soliton go to the gym?

To work on his wavelength....😀

gentleclockdivider

https://community.native-instruments.com/discussion/comment/61122#Comment_61122

Yet at the same time it can't accurately calculate the distance between two cities .

HAve a read , here's a verry interesting article by ( genius ) Stephen Wolfram of you know ..the wolfram language .

https://writings.stephenwolfram.com/2023/01/wolframalpha-as-the-way-to-bring-computational-knowledge-superpowers-to-chatgpt/

Sûlherokhh

https://community.native-instruments.com/discussion/comment/62732#Comment_62732

I love this guy. Physicist are just the right kind of craaazyyy. And he went the computational route, which makes all of it just that certain bit more comprehensible... 🧲

Haha, bit, you get it...? 😄

Edit: Apologies for going overboard...

tetsuneko

No disrespect intended, but calling inability to learn Reaktor building a "glass ceiling" sounds a bit hyperbole..

colB

https://community.native-instruments.com/discussion/comment/62845#Comment_62845

I think there was maybe a little hint of intentional hyperbole in the thread title intended for comedic purpose?

...but even then, in the context of Reaktor, it does make sense to think of a barrier existing between two levels of user base.

The whole point of Reaktor is that it's a development environment that is accessible to non-programmers. However, since the introduction of core, it has been possible to use some programming ideas and DSP techniques. So for users who are non-programmers, there is a metaphorical ceiling above which it has not been possible to climb without becoming a programmer, and to some degree, mathematician(yeah right, lol) and and DSP engineer, which is a massive commitment in time and effort, and just very difficult or even unachievable depending on your skill set.

Modern AI offers a glimpse into a future where that need not be a barrier - just have an idea and the AI can do the implementation for you. It's definitely not there yet though.

But I do enjoy @ANDREW221231's enthusiasm for the idea :)

---------------------------------------

...The problem is that GPT is not able to create, only to 'understand' language. So if you ask it something well known where papers and tutorials are readily available, then it will be just great. However when you ask it something that you know is possible using existing advanced techniques, but nobody has written on that particular idea before, you will get nothing useful at all!.

Also, with the intermediate stuff, you need to know enough about the topic to tell when the results from the AI are bogus, and in that case, you don't need the AI, you can just do it yourself.

e.g. recently there's been a thread on ring modulation. I'd been refreshing my knowledge of diode modelling as a result, and decide to ask GPT to write some code for a diode model. It did really well, but did add some spurious code to the end of its model. I removed that last 'extra' bit, and ended up with what I already had, which was just an implementation of Shockley's model. Almost a win for GPT, but redundant, and no use to someone who wasn't already familiar enough to spot the mistake. big whoop.

However, when I tried asking GPT to code up a model of a virtual earth mixer with saturation modelling, it didn't even get started, it spewed out various vaguely related things that were no help at all. That's because although there are plenty of sources on virtual earth mixers, and plenty on saturation behaviour, there is no 'body of knowledge' on the specifics of modelling the saturation behaviour of virtual ground mixers. All the requisite theories are well covered, just not used to that specific purpose in the publicly available literature.

...or maybe I'm just not good enough at authoring GPT prompts :)

Kubrak

AI is sort of garbage in, garbage out. Good and relevant answer requires good and competent inquiry. Vague question results in vague answer.... And one has to know the field of question to be able ask right questions.

It is simmilar to searching web using Google. If one chooses right keywords one gets relevant hits, otherwise many irrelevant....

The right question is the key for the success. But not only in AI, also in discussion with experts... And that is why I think, that it is always benefitial to know a lot by heart. If one does not know, he cannot ask competently and is not able to verify, whether answer is possibly right, or not.

Internet, AI and so are great, but to use it effectively and safely, one needs to know a lot.

ANDREW221231

https://community.native-instruments.com/discussion/comment/62845#Comment_62845

who said they don't know how to build in reaktor? I think maybe you didn't read carefully, my thread was directed towards competent-to-advanced builders who may still have trouble implementing algorithms from the legion of academic research papers on the internet

ANDREW221231

https://community.native-instruments.com/discussion/comment/62871#Comment_62871

what's a term for hyperbolic but also 100% serious? 🤔🤪

one thing I've noticed with these large language models is they are absolute trash at generating any kind of new insight , and id guess it'd be the same for pulling together stuff from different places to synthesize something new, as I suspect would be the case with a saturating console/mixer

it could be that my abilities are just at the ideal level to leverage chatGPT as someone who can usually understand how something works if I have the opportunity to pick it apart. haven't worked in it since making this post but there was that paper on phase retrieval that was completely incomprehensible but asking chatGPT made legible what the algorithm was actually doing

I think the most benefit so far is in having something specific (like a single paper) in mind

ANDREW221231

https://community.native-instruments.com/discussion/comment/61463#Comment_61463

Spectrograms are good idea for feeding it for sure

i think most probably a NN would have to be purpose built for that? I suspect we'll be finding out in two years max

Why did the soliton go to the gym?

To work on his wavelength....😀

I probed Sydney more on the intersection between solitons and audio

apparently shockwaves are solitons!!

😱😱

Quick links

Unanswered questions