Or a cheaper Chinese model will undercut the American models, in the same way that the current American AI models have ruined other businesses, which would be a delightful irony ![]()
hahah Claude from Temu!
Another experience to share.
A story about overcoming AI with AI.
This happened just a few days ago. I wanted to overcome some limitations of the old zita-convolver — namely, the latency introduced by buffered FFT processing, which seemed to slightly soften the attack when using cabinet IRs.
AI recommended a certain library, so I gave it a try. It worked well and performed reasonably well, too. It was a piece of vibe-coded by a Claude. But when I started looking through the code, I found myself confronted with excessive comments, layer upon layer of defensive programming, and template overhead that I simply couldn’t justify. That was the moment I really understood what vibe-coded software could look like. I just couldn’t bring myself to keep using it.
I “told AI” that the code was too messy and that I didn’t want to use it. It offered to help me develop something the way I wanted. So I put together a detailed specification, built a minimal test harness, and went through rounds of trial and error with AI. Before long, I had a working implementation, and I’ve been happily using it ever since.
The first 128 samples of the IR are processed directly in the time domain. This guarantees zero-sample latency, regardless of the host’s buffer size. The tightness of the pick attack and the immediacy of the response when playing guitar are noticeably better than before.
I only just saw this thread now. I use Ardour to produce music quite regularly and am learning programming on the side. I am against AI for coding and do not trust it at all for that purpose.
- There is an illusion of it being good, but unfortunately it is confidently wrong to a sufficient degree that any generated code is thus untrustworthy from a code quality standpoint. I’ve seen other people use it in action and have caught out errors made by even simple coding assistants on VSCode.
- It is unlicensable. AI-generated code always has snippets of someone else’s code in it. You can rewrite it with your own code, but it still has those snippets of someone else’s code in it. It’s no different than browsing Stack Overflow for code to transplant into your repo and maybe rewriting it to look a little different.
- It fails to solve existing problems around time and effort and introduces complexity. Rewriting it to blend your own code in also negates the time and effort saving benefits supposedly associated with it. It also requires understanding how correct it is, which adds extra difficulty to rewriting it to be correct. It is better to write the code yourself.
- It smells like a multi-level marketing scheme. People who use it fall for its appearance of correctness in a seemingly quicker amount of time then try to sell it to others as if it is a better solution to existing programming problems than other solutions are. It’s almost like they become addicted to it, like as if it was an illegal narcotic. They also use the fallacy of appeal to popularity to market it too, but just because it is seemingly popular does not mean it is good. It is just another way of outsourcing your own brain to someone else, except worse than just getting someone else to program it for you because it doesn’t intrinsically understand how it writes code, among other issues.
- Simple handwritten algorithms can generate code, especially repetitive grunt code that has the same patterns repeated throughout. This negates the argument that AI is the only way to write more code in a quicker amount of time with less effort, since knowing the algorithm would make it easier to rewrite algorithm-generated code.
- It is environmentally expensive. The water and electricity bills, whether through some Claude subscription or home-based setup with a solar panel, will be more expensive than some people with combined diet and laptop electricity costs.
- Slow isn’t necessarily bad. I am comfortable with sticking with older versions of Ardour much longer than other people would be comfortable with. Additionally, despite Ardour’s slowness, I see how Ardour takes ideas that work from other DAWs like Ableton Live’s cue tracks and reimplements ideas into Ardour, slowly but surely. I also appreciate Ardour’s rawness and honesty when it comes to functionality, simply showing things as they are with regards to functionality rather than embellishing them with hype graphics and ornamentations like mainstream and vibe-coded DAWs like FL Studio do. Even with feature restrictions, I can still work with Ardour. Therefore I would be in favour of seemingly slower development but with 100% human involvement.
- Stem separation isn’t really something I think about doing that much. I can just rewrite and retake parts if necessary. I’m generally uncomfortable with doing covers and remixes of music made by other people.
- It reflects “Embrace, Extend, Extinguish” philosophies. I suspect big tech’s attitude to Free, Libre and Open Source software is precisely this, and AI LLMs are very much the “Extinguish” stage of this as their usage is a step to delegitimise free libre and open source software by pushing/sneaking prompted code to many FLOSS repos, especially if undeclared. This Triple E spirit is very dishonest, and in this case they broke various laws in multiple countries in pursuit of trying to extinguish FLOSS this way. It would be bad especially on this front for Ardour to adopt such technologies that go against the spirit of FLOSS and thus associate with this philosophy.
On the flipside, I have learned that multiple prominent dangerous software bugs like the recent PHP bugs and the musl SSH bug were discovered using AI, I think. I do not know enough about that use of AI, and I do not know if it is sufficiently separate from the programming side of developing Ardour that involves writing the code. I think it is possible to do that without AI, and I don’t really know how much AI would create false positives when finding bugs. This is an area of AI usage I don’t know enough about, so I can’t say whether using AI for this purpose is good, or bad, or both. I suspect some third-party could run an AI on Ardour for this purpose to reveal bugs and then publicise them regardless of how Ardour chooses to use AI or not.
Otherwise, I am against AI and I don’t foresee myself changing my views on this for the foreseeable future. This is what I think.
A smarter approach that only processes the minimal required part (if any) in the time-domain can be found in x42 Zero Config|Latency Convolver (which is also built on zita-convolver).
Take that Claude ![]()
It is curious that it did find that code, and I am interested to see how much it copied literally. We already won one such case, where Claude copied @zamaudio’s and my code: Potential copyright infringement · Issue #3 · sebedard-creator/pt_api · GitHub (it required a DMCA notice).
Hi Robin,
Thank you for taking the time to reply. It is a huge honor to get feedback from the author of zita-convolver
Regarding the time-domain head:
You are absolutely right that zconvo’s approach of processing only the minimal required part is more elegant. In a2ir, I deliberately hardcoded a 128-sample time-domain head (which matches our first FFT partition size) to guarantee strict sample-zero latency regardless of the host’s buffer size, while keeping the code path simple.
Regarding the copyright concern:
I take this very seriously. The a2ir engine was written from scratch in this project. it use pffft (MIT) for the FFT backend, and I implemented the NUPC schedule, AVX head convolution, and lock-free atomic swap myself. I did port the mathematical formula of x42’s fftgain algorithm (ap^0.3 * pp^0.2) for loudness normalization, but I did not copy the C++ implementation from zconvo.
However, since this code was developed in collaboration with an AI (chatGPT), I cannot completely rule out the possibility that it regurgitated patterns from its training data. I am currently reviewing the codebase line-by-line against zita-convolver and zconvo to ensure there is no literal copying or structural plagiarism. If I find any similarities, I will rewrite them immediately.
Thank you for the heads-up and for your contributions to the open-source audio community.
Because that’s how it works ![]()
Ironically those bugs often tend to be in bits of the software / code stack that have existed for a long time and no-one pays much attention to or ‘understands’ any more - which both illustrates that AI itself isn’t unique in generating code that may not age well - humans have been doing that quite well on their own - but it’s also a warning about would could happen once we have a huge legacy of AI generated code which no-one cares to understand either. Sometimes the cure is worse than the disease ![]()
So there’s definite proof from your own experience that it is a theft machine? And i’m arguing this everyday with my surroundings, and they all think i’m nuts, and are convinced generative AI is this magical new tech made of stardust & alien plasma… despite being introduced to many artist who claim on-line & in public that they were being…well… basically robbed.
I’m claiming it mixes up all the available info (connected in various ways to the specs of the prompt) and spits out a new result, but basically made of stolen pieces, and they are claiming it produces new stuff (aldo they don’t have the theory how).
It’s just a stochastic parrot, it’s not magical. What people find incomprehensible is the sheer size of the data set required to make these things do anything useful.
The problem with pt-API was that they re-licensed LGPL code in terms of the MIT license.
In your case the code is GPLv3, so that’s fine. more or less. You need to state (C) holders, otherwise the license is not valid.
This is one of the issues mentioned in the OP why we cannot trivially accept LLM generated code into Ardour.
It is perfectly fine to redistribute and modify free/libre software if the result remains free/libre software.
Yeah i know, but that’s not the point
.
How does any generative AI tool create any “new” code, how does it create images, how does it create catchy melodies etc? My suspicion leans more to the theory that it steals whatever it can gather from the net (if it’s not clearly forbidden to do it, by bigger force, by money&influence) then assembles some new Frankenstein made out of someone else’s effort, spits it out and serve that to the oblivious “creator” who prompted something that fits the Frankenstein’s description. (Simplified description of the process, but basically something like that).
If it is not like that, how does it do what it does?
Even the person who came up with the term “stochastic parrot” doesn’t think that’s a useful phrase any more, or a useful description of what these systems can do or how they work.
And I think they really are magical, but not because the systems are magical, but our written language (as a representation of what we know and how we know about the world) is.
I like to point people at this “paper”:
Some folks have a problem that this came from Anthrophic themselves; while I agree that’s a negative, I still think that the paper’s overall conclusions/insights can and should change minds about how to think about these systems.
I’ll read it, maybe that’ll change my mind about it (but i doubt that) . Never ever i did say anything about this technology not being impressive. It is impressive, but that’s not the point. Here’s just part of “the point” presented in soft core manner…in gloves basically
:
What real artists have to say about this is not for faint hearted, but you can all guess it
.
@paul … here’s (to me) one logical subquestion. If these models are so advanced and magical, why didn’t Claude do its magic in case of said Robin’s code? Why did it just copy snippets of already existing code? Why didn’t it write completely unrecognizable code, create his own structuring and reasoning, did variables and values differently, redo the syntax in its own style?
Was it just easier, faster?
Was it just because it’s not dangerous to fiddle with floss code?
Maybe there’s only one way to do it?
Maybe it cannot come up with its own solution?
Maybe it knows Robins solution is good, so it picked that one right away?
How can we know the answers to any of these questions, and i do think those are important questions…as the time goes by and the number of these AI platforms grows, those questions become more and more important. You know, many people are already wondering what is their art/creations still doing on-line?
Given that no-one (not even the companies who created them) completely understand how these models actually work, one also has to be careful not to anthropomorphise their characteristics and then look only for evidence that supports that conclusion.
@paul In this thread and others I often oversimplify things deliberately to make a more succinct point, rather than because I lack a more comprehensive and detailed subject knowledge, but I will read the paper with interest.
Maybe the same reason I’ve sometimes done that with StackOverflow ?
It doesn’t strictly ‘create’ at all, it ‘reuses’ potentially in different ways than it originally was used. This guided by the instruction to develop something different, causes it to find pieces that may work and reassemble them in new ways to solve the problem. That is the gist of it (With some serious broadstroking going on honestly).
It isn’t a thief persay, what might be considered thievery might be the training it is trained on, but even there it isn’t clear cut, as what I just described above, is often not to dissimilar to how an inventor might approach a problem, identify the problem, identify the parts and pieces of the problem, and determine options that solve each of those parts and pieces, and assemble those options into a working solution that solves the problem.
But here is the thing to me, especially when talking about open source. I have been a firm believer in the power of open source as a movement for at least 2 decades, probably 3 at this point, maybe more. But I find that people have differing opinions of what the open source movement accomplishes. To me it is about the sharing of knowledge, finding the solution and them making it available to everyone intentionally to be able to reuse it so that it can be used to solve their problems, and we as a society and race improve.
This is a tool that allows more than just the ‘programmer’ to utilize this. It expands the ability to solve these issues beyond those with specialized training. Of course, we as a society are not used to that freedom, and everyone thinks that because something solves their particular itch, it must solve everyone’s without really doing the needed testing for stability etc. that most production solutions require, and as a result we get slop.
Now all of this waxing poetic does ignore the downsides, that wasn’t the point of the post, but rather to say that just because we reuse something doesn’t make it ‘stealing’. There is a lot more nuance to it than that, and for the most part the open source movement was specifically to allow for reuse in a lot of ways along with other freedoms.
Seablade
I don’t think of these things as tools, a hammer is a tool, even a compiler is a tool - generative AI is a service, its literally asking someone / something else to do the job for you. Thinking of generative AI as a tool is a mental sleight of hand which allows those who use it to believe they still have some ownership over what it creates, whereas I would contest that - and is at the heart of why I have no satisfaction in using it.
If I want to create something, then I do so precisely to learn about the process, or to express some form of creative individuality. if I just want the thing, the plugin, the DAW or whatever, I can just go and get something someone else already made.
Except that for the case of a Wayland backend for GDK2, this isn’t available.
I’m leaning away from any use of LLM for “functional” code for many of the reasons outlined in this thread.
However, for “infrastructure” code, I’m not so sure. We don’t have the resources/the resources don’t exist to do some of this stuff (GDK2/Wayland being a good example), and even then the actual cost is a barrier. Even if you scale up the cost of using Claude by two orders of magnitude, we would never find a human programmer anywhere that would do that work for that cost, and if we were to spend more, it’s a questionable use of the limited resources we do have.
Which would mean either just saying “Nope, there’ll never be a Wayland version”, or accepting a more limited use of LLMs for certain kinds of things. I don’t know which direction to move there (yet).
Exactly.
That’s also the reason i use it sometimes just for work. You know, if the team&customer wants to do something with AI, i don’t mind. I’ll help, i’ll create fx/amtosphere/music whatever they want me to do with AI. It’s a job, i don’t take pride in it, i do it for money etc…no worries.
But when it comes to something i wanna show to others and sign as a author - no way. Generative AI is not allowed there.