Ardour development and AI

Another infrastructure example would be a build-system. Say a native MSVC environment, or some docker container to build Ardour in, etc. This does not affect Ardour directly but can help with lowering the bar for new human developers.

Perhaps AI can also be used to complete some source-code documentation in places. This is useful for the Lua bindings documentation.

Part of this could also run locally (no speed needed), sidestepping some of big-tech involvement to some extent.

1 Like

Just a side note to point out that the zita tools are authored by Fons Adriaensen ( Kokkini Zita - Linux Audio ). Robin did not say that he authored zita-convolver, but rather than the Zero Latency Convolver he wrote is built on zita-convolver.
If you look at the source files in zeroconvo.lv2 you see that most are copyright Robin Gareus, but one header file is copyright Fons Adriaensen, and the implementation file for that header is copyright Fons Adriaensen and Robin Gareus.

The elegance of libre software development in action.

2 Likes

Here’s one on the upside…
You know, generative AI in all areas is a heated topic, so, naturally, discussion oftenly and very easily runs off track (i’m guilty of this too). After watching this video… :

…i realized this topic has one conceptual flaw. It’s basically a gathering of musicians and developers discussing one topic but from completely different positions. Programmer uses his experiences, musician his. When you hear Linus calmly explaining how AI is comparable to compilers or assembllers, you realise that programmers don’t realy see generative AI exactly like music artists do. Aldo it has some overlapping, those are basically two separate issues.

How based do you think Linuses stances are?
And a subquestion (cause Linus mentions it too) - is there a way to get as good of a result as with Claude if using AI locally?

I’m not anti-AI, i’m just anti all this social problems it currently generates (especially for artists, because i understand that side completely…i don’t wanna see art die, in favour to its mimicry).

True, I think this could be considered semantics then at this point as to me having a research assistant is a tool to be deployed the same as an encyclopedia that has presumably already done this. But no real argument on this, as I said just semantics primarily.

Well to me there is a couple of things here. In terms of ‘ownership’ what if ownership isn’t the goal? What if the goal is just to get a job done? It is a question I consider on occasion, for example:

  1. I have put together an ear training tool for my students using vibe coding/llms. It wasn’t that ear training tools didn’t exist, it was that one that met my specific needs didn’t exist. This allowed me to get the job done (Provide ear training to my students that wasn’t nearly as tedious for me to grade and correct, and provide them a path of exercises etc. that allowed them to grow in a way I felt served them best) but the goal wasn’t really ownership necessarily. Now that it has been in a working fashion for several months and I (and my students) have been using it, the thought has crossed my mind on do I make it more publicly available (Technically it is public, but I don’t advertise it at all, not even sure if crawlers have found it) and/or do I provide a service based off it as some of the exercises I am working on currently I haven’t seen an option for on other tools.
  2. I think I mentioned this above, but literally creating helper scripts for my clients to manage tasks, like opening multiple pieces of software and providing the same layout on their computer so that they always walk into a known situation. These people are very technically incapable, so I need to do everything I can to make this as easy as possible out of a fairly technical process (Operating PTZ cameras, switchers, audio mixers and live streaming). I have written these scripts/software before by hand, I now am realizing LLMs do it better and faster than me and mean I spend less time I literally donate to them coding tedious tasks. Again my goal isn’t ownership over these scripts, it is literally jsut to get the job done for the client to provide them an easier experience.

Well there are two different worlds I operate in.

In one, if I want to create something it is because I see a need to address, and I want to address it to make my and other people’s lives easier/better/whatever. This applies to the above.
In a second one, and my more creative world, it is in fact a free flowing expression of creativity, and in that world yes ownership is more of a factor for me, it is also the less likely one I am to use LLMs in.

If I want to learn about the process, I learn about the process, I don’t create to learn about the process, that is a side effect of learning about the process to me in many cases.

I also agree with you, I don’t feel a need to recreate what already exists. Right now my youtube channel is filled with ‘They open sourced Adobe/Sketchup/etc.’ videos that I don’t really care about as much, I watch them and I wonder ‘Why’ in a lot of cases. And part of that reason I know is because they want something free, and I understand it. Heck I have some VERY expensive and specialized tools that would be very tempting to do this for, but really it isn’t replicating the tool, but providing the workflow that works for myself that I would be most interested in (Most of those tools have workflows that I find painful to say the least). I don’t feel a need to replicate every other plugin out there, when there are capable options that do what I need, and unless I have something that improves on it, why wouldn’t I just use what already exists from people that understand the math behind it better than me? But if what already exists doesn’t fit my needs, then creating something for my own use is a possibility, though I would prefer to help modify something and improve it (Open source) so that it does fit my needs.

<ramble ahead>

Of course all of that doesn’t directly deal with LLMs, but what LLMs do is they give me the resources to be able to do more of that when I do run into those cases, at least as of right now. But again, we still have to test, check, and verify the output from LLMs to confirm they are actually doing what we need. Which means I have to be a subject matter expert already in a lot of the ways I do use LLMs to be able to verify myself the output. And while I understand programming and have worked in multiple languages over the years, I am not an expert, and especially not in DSP math, so I can look for what is wrong in code, and I can listen for what is not sounding correct in the output.

That means for my ear trainer example above, I am spending time listening and testing everything with ears that most people would say are well trained after having worked professional as an audio engineer for decades, and applying that to allow me to confirm it does what it needs, looking at the code to see if there is anything that stands out as wrong to me, and developing more rigorous engineer tests to confirm they act as I expect, all of which DO take time, but they don’t require the same skillset of an expert programmer on top of my other skillsets that I do have, and the coding aspect in the LLM still takes far less time than doing it by hand, and even with the review steps I tend to come out well ahead.

Again not always looking for ‘ownership’ as much as ‘accomplishment’ and by that I don’t mean getting a sense of accomplishment from doing the work, as much as accomplishing the task and moving on to allow me to focus on other, more interesting and more specialized, aspects of my work that demand my direct knowledge and attention (Whether creative aspects or otherwise). I know there is likely to be a comment or three about my use of the word accomplishment there when the ‘LLM just takes the accomplishment of others’ but that is also why I clarify that it isn’t about the ‘sense of accomplishment’ as much as the completing of the task to allow me to get to those thigns that I do get a sense of accomplishment from doing.

Hope that ramble made sense.

    Seablade
1 Like

Yep, and this is what I was trying (And likely failing) to draw a distinction with when I mentioned creative uses, but I honestly don’t know if it was in this thread or not anymore:)

Just because I am in support or certain technical uses of AI (Some programming/development and as an assistant, both with significant oversight from humans currently) doesn’t mean I am in support of all of them and ‘creative’ uses of it can be a much more nuanced topic to me.

Seablade
1 Like

I think Linus is absolutely right on the technical aspects of this.

But in this particular excerpt, he’s also totally ignoring all the non-technical aspects (the things mentioned in the OP, and elsewhere in this thread.

1 Like

Back to my comment about (generative) AI being a service and not a tool. I too used to hand code assembler language (and machine code) back in the 8-bit days (when often the binary was the source code, when there wasn’t enough room to host the editor, assembler and the program at the same time…)

I couldn’t code in assembler for a modern CPU (I could but it would quickly become unmaintainable and just, bad), so I use a compiler and it generates assembler code which I couldn’t maintain (or in some cases understand). And I’m fine with that.

So, is that like using an AI to someone who doesn’t code. To me, a compiler is a tool. It’s output is deterministic. The high level language is like the blueprint, a kind of schematic that defines the results I expect in precise ways. I no longer need to know how it arrives at its output, just as I don’t necessarily need to know or care how someone fabricates a physical item, but given the same schematic / engineering diagrams I would expect different fabricators to produce the same end product.

A specification, even a highly detailed specification, for an AI to generate the code doesn’t feel like that. Does that matter if you (happen to) get the result you want? If it generates unmaintainable or incomprehensible code, then it would be like having only the assembler language output from the compiler. Theoretically maintainable, but practically not.

If you want to make changes later then you just have to hope that the AI you use still understands the creative vision of your back of an envelope sketch as well as it originally did.

I’m conflicted because I’m both a musician (well a guitar player at least… :slight_smile: ) and a developer, so the fact that the unpredictable randomness of human interactions creates inspiring artistic ‘accidents’, and imparts essential creative soul is what I enjoy most - but I’m not sure I want that or the AI equivalent of it when I’m iterating over code that I want to be reliable, precise, maintainable and predictable…

2 Likes

Well, companies like Adobe behave abhorrently to their customers knowing their product is “industry standard” So people want a way out from them without losing functionality.

As for the topic, I don’t have a problem with A.I. usage in Ardour. Various bad things are happening in the adoption of the technology globally but boycotting it won’t change that.

Yes, that’s why i was wondering (somewhere in this mega thread) does it even make sense, from developer point of view, to use AI assisted coding in small portions, so that developer still can review the code and understand it completely? Does taking that route also speed up the process significantly/enough? If possible and if it even makes sense, i’d take that approach to avoid locking myself up in the future.

(OT - i did some crawling things too on my 8-bit(c64)as a kid in late 80s in Basic, like manual conversion of series of binary written bytes to 0-255 numerical values in order to create movable graphics - so called - sprites… i must have been soooo bored back then, and i’d definitely avoid doing something like that in this day and age :slight_smile: Funny thing is, i’ve got hired in a job interview like 12 years ago as a service technician for self-service machines based on my understanding of that conversion(it was used in smd resistors) …so you never know. The only candidate that was even better than me was my friend Milan - Arch user of course :slight_smile: :slight_smile: :slight_smile: ) .

I still do :slight_smile: :slight_smile: (When I get bored of all this audio (and AI) stuff) :slight_smile:

1 Like

You know Paul, somebody said it here before, sometimes you just gotta do what you gotta do. If you feel you’ve been pushed into a corner, get yourself out the best way you can think of.
On the flipside, i’m not really representative of the people’s pulse. Guess you’ll have to try and determine that yourself, or, so to speak, roll a dice on it. Don’t get me wrong, i simply live in Bosnia, and around here (aldo we really like our wild nature) we don’t really cry about cut down trees etc …
(I’m trying not to sound like insensitive hillbilly :slight_smile: , but that’s the way it is around here for a long long time now).
Best of luck to you and Robin with this heavy decision, i’m done with this topic :slight_smile: , there’s clearer things that wait on my attention like .sfz instrument libraries etc :slight_smile: .

And that is where it is important to review the code as it is generated in general I think. Even on the scripts and things I mentioned above, I had to watch, and it is the reason it took 7 generations/prompts to get what I was happy with as it often took methods I wasn’t happy with, or tried to pull in extra libraries that weren’t needed. It tended to work (to a point) but I needed to refine it and did, before it was good. All that was still MUCH faster than me doing the research and writing it myself honestly though.

But this is also why I think LLMs are absolutely the WRONG tool to use in a lot of cases I see them being thrown at. Specifically the deterministic vs non-deterministic nature of them. I will use an LLM to help write a deterministic program, but I will make sure the output from the program is deterministic like I want it to be, through manually checking and through automated tests both. But I see advertisements for all sorts of thigns that seem to be ‘Do it with AI’ that is a really bad idea because it is non-deterministic when a deterministic output is MUCH better.

All that being said, I am currently working on my first project that is designed to be utilized from an LLM prompt and am going to be watching the output carefully every time I use it. In this case though the non-deterministic bahvior CAN be desired, but obviously can also create problems and headaches, and I am not sure yet that it will work like I want, but since it is primarily gathering research for me to manually go through and use in my work, I can deal with it and will know when something is off fairly quickly in general. Again verification is needed and it is required in my workflow to turn up issues even when I do this work, so there is no way I am removing that verification steps with an LLM as part of it.

    Seablade

Yea I am well aware of what type of company Adobe can be, lots of first hand experience including a current option that is borderline financial fraud honestly. I just haven’t used their tools directly myself ina. while as I really do see better options out there in most cases. Adobe may not have been the best example for me to use though, maybe a better one would be the vibe coded plugins, etc.

Seablade

I’ve probably said it before (in fact I know I have, but maybe not in this thread) but I think the old truism still holds that “the time you save by getting a computer to do something for you, is almost always spent watching it to make sure it doesn’t do it wrong” :slight_smile:

(typically to describe the printer which prints the test pages just fine and then inexplicably switches to landscape the moment I leave the room. For those who remember when we still used printers that is…)

1 Like

The point has already been raised by Mike about the impending enshittification of the LLM experience and the reality they’re only giving an addicting taste for free and when they become fully pay-to-use this will really throw the brakes on a lot of these open-source vibe coded projects. In my microcosm experience with Gemini this week this is already happening, I realize Gemini and Claude are very different LLMs but I use Gemini as an advanced search engine and also to assist me with making bash/YAD utilities for AV Linux. To be clear I limit myself to programming languages I am at least somewhat familiar with, you will never see me Vibe coding a Plugin or DAW. When I first tried Gemini a few months ago I asked it if I would need to go up to a paid tier for the low-level bash and YAD stuff I was doing and whether I could expect better output from a paid tier. It told me that what I was doing wouldn’t even get close to any usage limitations and the quality of code output would be the same… Fast forward to this week when I was simply getting together a scripted way to build and package the entire Enlightenment DE stack for AVL and I was hitting usage limits where it would time out and Gemini made a LOT of stupid mistakes and typos as we went along. I am painfully detailed with my prompts and I know exactly what I want the end result to be so in my case I caught most of the errors and in the end the script got finished and works. In a larger context I would be quite alarmed as a ‘real’ developer on a project like Ardour when the enshittification phase really kicks in and the usage limits and tier upgrade costs only get progressively worse. Lol, paid Ads in your LLM interface anyone?

2 Likes

Here’s me trying to offer a somewhat different perspective beginning with a full disclosure. Please, bear with me because it is probably going to get long. I’ll try as much to avoid politics as I think this is wrong forum for that.

Full disclosure:
I work in a tech company. It is not an AI company but its product sells an AI solution. I am under no circumstance an AI expert and probably on some topics I’m repeating what I’ve been told, but I don’t believe I’m too away from the current topic.

[Edit]: In the interest of my own privacy, after reading my own post… I’ll leave it on that I have a number of credits to spend. The full text will be a little incoherent about why I do what I do, but who’s looking for coherence in 2026?

In all that context. I believe the current situation with AI is a disaster.
It is one more wrong step in every direction. It is hateful.
And (here’s the hopefully new thought on the topic)… it won’t go away.
Because it is sistemic.
It is a byproduct of our society and system.

Having said that, some thoughts.

  • Is it theft?
    Theft is a moral and ethical issue that we try and regulate using law.
    Now… take a look: AI Copyright Case Tracker – Chat GPT Is Eating the World

    If you ask me? Yes, no doubt. Companies have used material (tons, billions, gazillions), images, texts, code, music, videos and whatnot which had owners, haven’t compensated them in any way, shape or form and now they are selling a byproduct of that as a service. According to a lot of judges, that is not theft because that is the way we learn and these companies used this materials to make their models learn.
    Very conveniently they always leave out the only important topic on this which is property so if someday you meet one of them, please ask them on my behalf if you can download a copyrighted creation because you want to learn.
    I’ll wait for the answer

    And still if it is theft or not, as I see it, is not the most crucial topic in that realm. Here’s a very interesting and short article to get your ten minutes of rage a day: https://www.forbes.com/sites/maryroeloffs/2026/08/17/ai-companies-are-buying-and-destroying-antique-books-heres-why/

    It contains this quote:
    “Project Panama is our effort to destructively scan all the books in the world,” an internal Anthropic planning document said. “We don’t want it to be known that we are working on this.”

    Our system will destroy every thread of culture it can put its hands on to sell you a service if it means that only they can have access to it (and altering the ecosystem along the way) even if they don’t own it.

    For some reason that escapes me, I remember the Napster sues but it is hard to read about anyone suing these companies… must be me.

    That harm has already been done whether we like it or not. And will continue to be.

  • But it is made of fairy dust and feathers of unicorn wings
    No. By no means.
    The techniques used, although very refined in the past years I already studied their basics in University (as many of you here). Initially (with the first models), it was an inference process ¡FOR EACH LETTER! of what letter comes next. As image generations is an inference process for each pixel of what pixel comes next and sound generation is an inf… you get the idea.
    The result of the inference is a probabilistic result that varies with the amount, quality and variety of training data. That’s the reason they do so well comics of a specific style and so bad general images and realistic scenes (as of now) as an example (reality tends to be a big variety of data).
    The only reason this didn’t happen sooner is because we didn’t have powerful GPUs sooner.

  • But not only it is intelligent, it is SUPER intell… I’m sorry, I can’t finish the phrase
    No, it is not.
    It doesn’t resolve conflict, it doesn’t create anything new, it doesn’t imagine, it doesn’t want, it doesn’t feel an urge to investigate, create or share it isn’t awake at night thinking of a specific topic and trying to find a way around and NO it doesn’t solve math problems by itself auto-magically.
    It refactors training data in a very clever way and is able to automatize repeatable actions while trying a number of options at an impressive speed. Which makes it… a very fast machine.
    But being the economic horror it is I’m guessing selling “SUPER FAST… THINGS” wouldn’t sell very well

  • It is an environmental nightmare
    Oh, yes. Yes it is.
    So is the pistachio agriculture in… well in most of the world, or the avocado (very popular in my country) or sending food, among other things across the ocean because a country produces it cheap or buying things we don’t need with money we don’t have from the other corner of the world sent in huge cargo ships that run on fuel.
    Am I being a relativist here? Yes. I know.
    But if we are to make a list of things that are ruining the environment humans need to survice… well… AI is just the new kid on the block. And not the biggest one.
    It is AMAZINGLY concentrated in a very few hands so its impact is more noticeable.
    Yes, yes, I know, I know some of you have a study that says that consumes very little water (because the study only considers the water spent in cooling data centers while providing the servers that do the inference) and some of you have studies that say that consumes a huge amount of water (they also consider the training and the whole energy consumption of the training of the dataset and other services).
    It is an environment disaster.
    As it is your (my) foreign car or your (my) exotic vegetable in your (my) fridge.

  • But it is an economic and political disaster

    Ooooooh, the size of it. A whole economic cornerstone that lies on hardware not yet manufactured to get installed in data centers not yet built, fed with an amount of electricity that no country is able to produce at the time (heard about that fantastic idea of mini-reactors?), on networks not prepared to carry that amount of power (and not being modernized at the time because that’s the government’s issue) to be sold as a service to nobody knows how many users for nobody knows at which price.
    And here’s a fun fact, the companies spending money on this are not declaring loses because this is not at all money spent on something they don’t have is… “R&D investments” so even for stock companies, loses doesn’t exist.

    Here’s a not that accurante but fun link: https://isaiprofitable.com

    Yes, yes it is. And is getting bigger. Summarizing the last months greatest hits, these companies, on the verge of being stock companies (which has financial reporting obligations) said all, at the same time, that they couldn’t because “If so we are going to destroy humanity because… reasons” and ON THE SAME WEEK asking governments to take care of this (which they all refused because these companies are going to “autoregulate” somehow).

    Politically… that’s a fun talk for somewhere else.

    And on top of that we are starting to see this:

    Revolut Cuts AI Coding Costs ~90% With Self-Hosted GLM on 56 B200s | AI Weekly

    Which I believe we will see more and more. Someone might wonder why in all these news the same specific models for their local environment are always being used…and why those models exist, but again, that’s political, so I’ll pass.

    However! Is it possible to make a self-hosted useful tool? Depends on your scope and your knowledge. If you know nothing about programming… no. Not at all. If you know very little still probably no. If you have use for a coding assistant that takes your boring, repetitive and procedural tasks for you, yes. But it’s not just an LLM that you need, it is a whole ecosystem of tools, process, harness and tests. but even at consume hardware, as already stated in the thread you can have a self-hosted low consumption fun assistant for several tasks on a docker container. That I know for a fact. And even, using shared memory an MoE that is actually very useful with a relatively cheap investment (a little more than a new laptop)

  • Can’t be licensed
    I mentioned it before. That is well… not false but inaccurate.
    Take the list of the 50 biggest software vendors in any market.
    Their code at a bigger or smaller capacity is AI-generated.
    And is copyrighted.

    Curiously, there doesn’t seem to be any legal issue at the time, is it?
    Well… 50, I have just checked the biggest 50 and for a lot of them I don’t know, but for a big majority of them I know (I say know, not guess).
    Of course, for all we know they could be copying and pasting directly open-source or code from forums full solutions and nobody will know so…

  • ENOUGH! Do you even have a point???
    Well, actually, I do.

As I have already mentioned, in my opinion, although it might seem relativistic, it is not a matter of this specific product or this specific service. Is the new iteration of a systemic problem that manifests in many ways, shapes and forms and most of the time is rendered to a specific question which is… property.

I rarely read debates about taking property of open source models (the environmental, economic, thievery and legality harm has already been done by the time a model is published) and enhance, improve or expand current solutions.

I myself, don’t like the idea, as it is today but I’m so willing to see another angle or another proposal…
Because what I see and read so far is that this service is in very specific hands, used only by very specific companies which have none of these concerns and they don’t care about it that retains the property of everything including the code.
While I love that this kind of debate can only exist in very specific places, it is safe to assume that it will grow (really?? DAW clones?? REALLY?). I already mentioned, I am seeing closed source plugins on sale for a price that I know for a fact that are done with Claude (not because I’m very smart but because there are people lazy enough to copy and paste Claude features and specifications texts and put a price tag on them while Claude would give you the exact same answer to any other user).

About things like Suno and services as such… does someone really listens to that? Is there anybody that really enjoys what it produces? I don’t get it.

I’m not looking for sympathy.
I’m spending those tokens in what I (in my limited knowledge) believe might be useful for other persons with the hope that someday someone that has real knowledge and skill will try to make it something actually useful and that’s where I will find some humanity in this.

P.S.: Long as I suspected… sorry. Hopefully there might be a useful phrase to show something new…

2 Likes

True to an extent, but if that was completely true, we wouldn’t ever use a computer or write a program.

But yes I do intend to measure out if I am saving time or not, but really I think this is less about time saved in effort and more about overcoming the first push to start the project, starting not he tedious work is always a chore to me, and after that first batch is when the ‘fun’ can begin. So even if it doesn’t save time because I have to check behind it, it is a worthwhile thing to me as checking behind it exercises a different area of my brain than the initial tedium of work:)

Yea I am weird.

Seablade

Is that creation licensed open source or Creative Commons? Yes. Even in the concept of commercial releases in the US there is a grey area around fair use for education.

Sadly I agree with you about the destruction of books, but there ARE much better ways to do that that do not destroy them, the issue becomes the race for cheap knowledge they find it cheaper to destroy, I hate the fact of it but the issues here to solve are FAR more complex than can be covered here.

Seablade
1 Like

Copyrighted material.

I’m no lawyer and I understand this is a very complex challenge around ruling and law. However, it escapes me (or not) why there are so many appealing and comes and goes about why if a company uses copyrighted material to sell a product might or might not be fair use but there was worldwide unanimous consensus about that of an individual accessed copyrighted material, was a crime.

Take this for what it is, my knowmedge of law is close to none and very specific for each country.

It just doesn’t feel right that a company acquires one copy of copyrighted material (Anthropic was in a suing for pirate books) to sell a service and the author doesn’t see a dime.

They have dissapeared (I believe) but there was a flood in the summer of YouTube AI videos which were very obviously AI versions of well know drummees on the Internet.

Of course, they haven’t seen a dollar.

Don’t they class it as ‘marketing’ :slight_smile: (in a sense you have to admire the inventiveness of that - classing the loss you make by giving your product / service away for less than it costs to make is ‘marketing’ even if it reveals a certain cynicism, see my other posts about enshittification)

This. Thank you

I don’t have an objection to AI specifically, it could be and has been a beneficial technology in many areas of important scientific research. I have a real problem with generative AI, and the morally bankrupt way in which it is being implemented and deployed.

If the Ardour devs want to use it, then that’s a decision for them to make - the project may lose subscribers if they do, and probably won’t if they don’t so the choice to me would seem to be obvious.

Is there an existential crisis for Ardour if they don’t / can’t support Wayland, and this could be solved by AI? - well apparently its 16,000 lines of code, so I guess it could be solved by people. Just one of my plug-ins has around 50,000 lines of code in it and I wrote every line, its a lot, but its not that big of a deal (it was my job - what else was I going to do other than sit around thinking about not doing it…)

Some code gets re-used, obviously, but I wrote the UI toolkits, openGL renderers, shader code, DSP etc from the ground up because that’s what you do when there isn’t AI. I wanted to make a thing, so I made a thing, and it’s what I would still do, and I’m sure Ardour itself has vastly more code, code that’s already been written without the benefit of AI.

1 Like