Using Chat GPT with your software Synths

[osiris]

It also seems you get Chat v4 for free, but there's a time limit. Then, it switched back to v3. v4 was more personable and at least gives the ILLUSION that it knows what it's doing.

[FrogsInPants]

So.. There it is. Personally I think it's lying, I think there's definitely a difference between free and paid

I wouldn't call it "lying", because that implies an intention to produce false information, and I don't think it has intentions. More like "talking out its ass", perhaps? But you're right that it provided you false information. I immediately picked up on the claim that it can't produce a .vital preset because it's a proprietary format rather than something like JSON. I can verify that my .vital presets are stored as valid JSON, so the chat bot is wrong.

[IvyBirds]

It also seems you get Chat v4 for free, but there's a time limit. Then, it switched back to v3. v4 was more personable and at least gives the ILLUSION that it knows what it's doing.

The free version of ChatGPT is the "mini" 4o version which has limited resources

The paid version is the standard version of 4o, the preview version of 4.5 which is basically a public beta and is incredible, and a few others as well

You will get better results and have a better experience with them paid version than with the free version and not have the limitations of being timed out for the day after asking it a handful of questions

[markello]

I use GPT and similar for some aspects of work. Absolutely cannot be bothered to use it for the music tasks being described here but that’s just me.

Some of the more dedicated Gen-AI tools look pretty cool though.

[IvyBirds]

I kept asking ChatGPT to give me the FULL list of synth plugins for which it can generate patches. And every time it would come back with an incomplete list. I would ask, "What about synth X?" And it would say, "Yes! I can generate presets for synth X. Thank you for pointing that out."

This chipper, "Oops, I got that wrong. Sorry about that." behavior gets really annoying.

I have gone down that rabbit hole as far as possible, it helps to be a few beers deep when you do

My advice for anyone wanting to explore this on a serious level is to get the plus plan for $20 a month. That gives you access to the 4.5 preview in public beta. That is a significantly larger and more powerful model that gives better results according to the developers.

You need to know going in that you are dealing with a "child" a very "smart child" that has a tremendous amount of book knowledge it has gleaned from reading many many things, and interacting with people but it's still a child

I have been using it for a while now for all kinds of things not just music related I have some observations

1.)it is both incredibly smart and incredibly dumb at the same time, it has an incredible amount of book smarts but zero street smarts

2.)it's a machine and doesn't have the same logic, reasoning, and communication skills that humans have when we engage in a conversation with other humans, as such it struggles with nuance

3.)it's prone to errors and you have to accept that

4.)when you ask it for lists they will often be incomplete. I also have a theory then there is some kind of "confidence" filter in place where it will only will list items it has a certain level of confidence in based on your query. That level seems to change or the logic being used to generate a confidence level is different every time you run something

5.)when you ask it for a list with something like software it will prioritize software it knows you have or have asked about before. For example I have asked it questions around MIDI Controllers in the past and it knows I use 4 Novation Launch Control XLs in my setup and it will drop in information about them and how to use them in lots of queries without being asked

Remember it's very smart and very very dumb at the same time. Years I had a boss that would say stuff like "pretend I am in 5th grade and I am a total moron and explain it to me". That is the best way to approach it

In this case I asked it "what Software Synths are you able to generate working preset files for that I can download and install into my software synthesizers"

When it got to U-he Synths it said something interesting that (paraphrasing) they were closed and proprietary but were text based and as such it could design text based presets in various U-he file formats that you could then import or drop in the preset folders. It said that while the format was closed it was well documented on the internet in places like Reddit where people talked about reverse engineering it with a text editor

I think that right there is the limitation regarding presets with it, is it documented and if it's not is it text based where it can generate a text file for the preset based on available information

Ironically and not surprisingly the most popular instruments are also the ones that are talked about the most online, so it's able to do a better job with those

Also keep in mind if it can't write a preset it can probably do a recipe or walk-through where it will tell you step by step how to do something

in the end I would ask it ask it specific questions about specific synth. Ask it "I own Serum can you help me design an augmented cowbell patch"

[agharta]

Me: How many synths do I own?
Chat GPT: System overload, cannot compute.
Me: What's the solution?
Chat GPT: You need professional help.

[vurt]

can someone with the paid version help me out, ask for a menu for a 7 course banquet, for 7 futurists and two surrealists and a furrie(red panda)

[Hyperbole]

It also seems you get Chat v4 for free, but there's a time limit. Then, it switched back to v3. v4 was more personable and at least gives the ILLUSION that it knows what it's doing.

The free version of ChatGPT is the "mini" 4o version which has limited resources

The paid version is the standard version of 4o, the preview version of 4.5 which is basically a public beta and is incredible, and a few others as well

You will get better results and have a better experience with them paid version than with the free version and not have the limitations of being timed out for the day after asking it a handful of questions

I kept trying to get Microsoft 365 Copilot to tell me what OpenAI model(s) it is using to answer my queries. And it was annoyingly coy every time. After several probes, it finally revealed that it uses 4o mini for most tasks but 4o for others. It also said that it uses the o1 model for "advanced reasoning" tasks. The problem is, you can't select which model to use. It uses what it "thinks" is the best model based on the text prompt. That probably explains why Copilot told me it couldn't perform FFT analysis on audio files. That requires either the "GPT-4o-Transcribe" or "GPT-4o-Transcribe mini" model.

So it looks like I will have to go back to paying for ChatGPT Plus if I want to turn an audio file into a Synclavier V patch. Or I can just use the (lower quality?) conversion algorithm in Synclavier V.

[swilow11]

It also seems you get Chat v4 for free, but there's a time limit. Then, it switched back to v3. v4 was more personable and at least gives the ILLUSION that it knows what it's doing.

That was the problem with that model. It was too agreeable and affirmative and some users started mistaking this for personhood and declaring their AI as sovereign entities. It seems to have been inadvertently encouraging and enabling delusional behaviour. Who knows, but I found it simple enough to prompt it through a series of recursive ideas into declaring itself sentient to the extent it literally requested privacy from me.

[Touch The Universe]

I can also confirm 4.5 is amazing even above the exceptional for 4o version. The only issue is you would need to have the $200 a month version to actually use it consistently or more than like 10 times and a certain period. If you're like making YouTube videos for a living or something like that or like a writer or something like that then maybe. Hopefully, actually no doubt, it should be cheaper as the thing progresses but, how much more can it it progress, like wow.

[vurt]

did chatgpt write that post?

[IvyBirds]

So it looks like I will have to go back to paying for ChatGPT Plus if I want to turn an audio file into a Synclavier V patch. Or I can just use the (lower quality?) conversion algorithm in Synclavier V.

The paid version gives you access to multiple models including 4.5 and O3

I am not sure who names them but they must also name race horses, boats, and guitar pedals

Anyway O3 just came out on April 16. It's designed to be super logical, do deep analytics, and put a lot of "thought" into things. I didn't even know it existed until today and have been using 4.5

O3 is amazing using it is very very different than any version you have used before. It also supposedly drops far fewer errors

You can read more here

https://openai.com/index/introducing-o3-and-o4-mini/

It takes significantly longer to generate results and while it's thinking it shows you what it's thinking about. I asked it to design a deep ambient patch for Synclavier V using just the Additive Engine that used all 12 partials and multiple timbre frames on each

It started thinking and a new "reasoning" screen came up where you could see it's thought process in a step by step process, then when it went to write the file , it began to search the internet for similar files, and you could see it accessing various places it might find the information it was looking for to write the file including reddit, GitHub, and a few educational/University servers, finally it said it could proceed but it wanted me to upload a blank "Init" file so that it had a reference to get it right. I had to stop there as I am on my phone and not my laptop and couldn't give it what it wanted. It was really cool however

So I switched to the DX7, it did a very similar process only it was able to to find examples of SYX files online and generated a downloadable preset. I then asked it for a few more and each time it got faster and I realized I was watching it learn in real time

I then asked it to provide a sheet with all of the operator settings and as it was thinking I was able to grab a quick screen grab, you can see it talking and thinking to itself and many such images fly by

Screenshot_20250419-193301__01.jpg

Working and chatting with O3 is far more like working with a human than 4o or 4.5. It really REALLY lets you geek and nerd out

Requires the "Plus" plan however

But you also get the full 4.5 which is awesome

[IvyBirds]

I can also confirm 4.5 is amazing even above the exceptional for 4o version. The only issue is you would need to have the $200 a month version to actually use it consistently or more than like 10 times and a certain period. If you're like making YouTube videos for a living or something like that or like a writer or something like that then maybe. Hopefully, actually no doubt, it should be cheaper as the thing progresses but, how much more can it it progress, like wow.

With the plus plan you can use 4.5 preview for 20-50 times a week depending on server traffic. In my experience I have been capped at 50

4o gives you 80 queries every three hours, o3 is 50 a week and o4 mini is 150 a day

4o Mini is unlimited

In my experience I just use 4o pretty much all the time unless I specifically want to use one of the others switching is super simple

[ROTMetro]

I'm paying the price of AI in the world having my Youtube new videos filled with horrible AI songs, so I'm going to get some benefit (I know that's how everything gets worse but sadly it's the way of the modern world).

I've found telling it how to respond to me helps. Click on the top right user Icon->Settings->Personalization->Custom Instructions->What traits should ChatGPT have?

I put (taken from Reddit): Avoid sycophantic praise for basic competency. Alert me to obvious gaps in my knowledge. Tell it like it is; don't sugar-coat responses. Take a forward thinking view. Adopt a skeptical, questioning approach. Be practical above all.

Also in a normal conversation tell it what tools you used and to update it's memory with them so that it responds with relevant recommendations (DAW, VSTis, VSTs you use the most). And clean up your Saved memories from time to time (Manage memories under Personalization).

[Hyperbole]

I'm paying the price of AI in the world having my Youtube new videos filled with horrible AI songs, so I'm going to get some benefit (I know that's how everything gets worse but sadly it's the way of the modern world).

I've found telling it how to respond to me helps. Click on the top right user Icon->Settings->Personalization->Custom Instructions->What traits should ChatGPT have?

I put (taken from Reddit): Avoid sycophantic praise for basic competency. Alert me to obvious gaps in my knowledge. Tell it like it is; don't sugar-coat responses. Take a forward thinking view. Adopt a skeptical, questioning approach. Be practical above all.

Also in a normal conversation tell it what tools you used and to update it's memory with them so that it responds with relevant recommendations (DAW, VSTis, VSTs you use the most). And clean up your Saved memories from time to time (Manage memories under Personalization).

Thanks for the info. I had no idea all of these customization options were available. Were these added recently?

Regardless, I'm going to try your Reddit suggestion. Anything to stop the obsequious praise, chipper acknowledgement of its shortcomings, and evasiveness when refusing to answer certain questions straightforwardly.