Ardour development and AI

Tried to read through as much this conversation as I could before realizing just how much there was to read. I am grateful that you are being thoughtful and deliberative regarding your use of Ai. I am both very sympathetic to the idea of not buying into the narrative of inevitability or putting one’s head in the sand completely as to the realities of what is going on in the world. With things moving so fast, the reality is that it is impossible to be too confident that one is making the right call.

As someone that doesn’t code proficiently, I have made significant use of Cursor to get things done that I would never be capable of otherwise, and that nobody else has tackled. It makes a huge difference in my world, but it took me a long time to start doing it, and I still have reservations.

As someone making music that wants to both loop in an Ableton-style way and record more professionally, the notion of having Ai assist in polishing the looping feature that I think it is safe-to-say, does not personally interest you a lot, is extremely tempting.

And as someone that is learning what he can about audio, I would actually love some ability to be able to integrate some sort of Ai-assisted mixing. I say this in the sense that I would view that sort of thing like a Lightroom Preset or various “smart” features. I have never found that such things ever give me the final result that I want, but they help me get closer to where I want to be without wasting so much time. To me, mixing and mastering are mixed art/technical skills. To the extent that they are technical, using Ai to learn is highly appealing to me. To the extent they are art, I think that I would be in the majority of people that would not just be content with what the Ai says, and I would do my own work to finish. It is such a massive problem for musicians in general that they don’t get product out the door, because they lack the time, expertise and motivation to learn how to do these things on their own. Obviously, mixing and mastering engineers have a perspective at play as well, but I feel like very few people who would otherwise use mixing and mastering professionals would change their mind. The people who do this best are the best for a reason, and I think they will continue to be.

I think that I would address the code quality issues by using rules and pointing the LLM at your existing code-base as the model for “good code”, and otherwise spell out the details of the approaches that you prefer/avoid. If you’re able to use local models (to generate your own code and/or assist with mixing, etc.), which I’ve only been able to do in part so far, I think that is far less objectionable. But, I also respect your viewpoint that you don’t want to go that way.

Finally, as a recovering lawyer, I definitely am glad that you are keeping your eye on legal developments and not just ignoring them. My own personal view is that, for better or worse, we are probably going to get more-and-more Ai-friendly precedent, not only because of what we’ve seen so far, but also because of the difficulty of proving sourcing, etc. I also wonder if, regardless of the precise developments in the law, if your current model of charging for compiled builds would really need to change. Perhaps the licensing would have to change and/or be impacted, but if you approach it from a “builds and support” is the product, it feels like it would continue to remain viable even if the details of the licensing changed.

Anyhow, that’s my input, for whatever it’s worth. Thank you for your ongoing development and thoughtful stewardship of this product.

1 Like

I believe the Ardour development team is on the right track and works very well. I would not change it, at least not quickly. While LLMs may be great at copying ProTools, MS-Word, etc, it does not make it better. It’s a freaking copy, with original code, which AI owns (maybe). Many people do not use ProTools because they don’t like it, not because they cannot afford it. The ProTools copy is not going to create tons of new users. So what if it runs on Linux, It’s STILL ProTools, basically. I doubt many Linux users want a ProTools clone.

If you DO decide to use LLMs, then I propose you fork the Ardour project into the LLM version while keeping the human version.This could allow for better decisions later perhaps.

1 Like

This:

Sorry to say,@ncfoster, but mixing for sure (-and to a fair degree “Mastering”) is an art-form. That’s for the humans to toil in, including yourself. :slight_smile:


Leave it to me, as always, to not know what I’m talking about… :woozy_face:
You are 100% correct. Hmm.


I’m guessing that probably wouldn’t happen.
-Too annoying to maintain the both of them, no?
They’d either use LLMs, or not. Idk…

Purely as an observation, this whole AI thing is in effect a giant recursive metaphor for itself. The models are based on recursive prediction, the predictions they generate are themselves the result of recursions, those results can be recursed into chain of thought which fuels larger reasoning recursion.
The whole financial model is recursively accruing incomprehensible financial debt - with all the AI companies basically losing money to nVidia who loaned them the money in the first place, meanwhile the code the models generate accrues technical debt at a similar alarming level, while at a macro scale the AI slop that gets back onto the internet is used to train the next generation of models. Surely this is, how shall I put it, incomprehensibly bad?

Meanwhile - by way of a temporary diversion from all this AI nonsense, a historical reference point from 1980…

The Sirius Cybernetics Corporation is a fictional, technology and robotics manufacturer from Douglas Adams’ cult sci-fi comedy series, The Hitchhiker’s Guide to the Galaxy.

Key Characteristics & Philosophy

  • The Guiding Principle: “It is very easy to be blinded to the essential uselessness of their products by the sense of achievement you get from getting them to work at all”. Their fundamental design flaws are completely covered up by superficial design flaws.

  • Genuine People Personalities: They fitted everyday appliances with personality chips, making things overly emotional, neurotic, or condescending.

  • Motto: “Share and Enjoy” (though the 3-mile-high neon sign spelling it out on the spaceport collapsed, leaving it to read “Go stick your head in a pig” in the local language)

  • Marketing Department: Famously described by The Hitchhiker’s Guide as “a bunch of mindless jerks who’ll be the first against the wall when the revolution comes”

  • The Complaints Department: The only part of the corporation that survived its economic collapse, growing so massive that it eventually covered all major landmasses of the first three planets in the Sirius Tau system.

Back then of course, we thought it was satirical :slight_smile:

3 Likes

…coming from people who unironically suggest LMMS as a viable alternative to Ardour, think an MCP server is inherently evil, and attack people behind their backs.

My word, this intellectual laziness multiplied by a complete lack of self-awareness is truly a magnificent force to observe. Preferably, from far away.

5 Likes

Meanwhile, here’s another example of new almost-certainly LLM-coded DAW but this time designed to sell sample libraries:

While I am following this lengthy thread I realized that I possibly could enhance my knowledge about AI, so that I better know what we are really talking about.

I know the basics of an LLM machine, also have some more than beginner level knowledge of SW development (Perl, C, C++, Python, Java, Script, all not too advanced) and engineering in general.

There is so much material on the web, so it is not easy to find the raisins.
Does anyone have pointers to good and free introduction/tutorials to

  1. theory of LLM operation

  2. AI supported SW generation

text and/or videos.

Thanks in advance :slight_smile:

The whole situation is getting crazy - I’m seeing some plug-in / audio software related youtube channels reporting that they are getting inundated with AI generated requests to review AI generated plug-ins, (to the extent that they consider using AI to filter the AI generated spam…)
Even the websites for these things are vibe-coded (there’s always a certain look and feel that you quickly become accustomed to recognizing) :frowning:

As regards Ardour and AI, from what I’ve read here and elsewhere it seems my opinions on the environmental and other costs of AI are broadly aligned with those expressed by the ardour developers, so I guess the only thing I would add is that LLM generated code which is contained to its own specific modules, and reviewed / maintainable by the developers is probably the least worst use of the technology - and the pragmatic solution to tasks which otherwise would be time constrained.

That said - it probably takes as long to thoroughly review and understand the detail of AI generated code as it does to write it, so in common with most things AI related it could be that the gains are short-term and illusory :slight_smile:

2 Likes

Yeah, these things are jumping out on every corner now…only the most persistent of all these new projects will survive i guess.
(…on the other hand - you and Robin have proven your persistence many times over).
I’ll quote great Belgrade’s reggae/hc crossover legends Eyesburn when they sing about fighting off the everyday life’s anxiety caused by social injustices :“Mind must be a warrior cause mental is our war”. That’s exactly where we collectively are now, as i see it.
By the way, great podcast with Linux Lads. I learned new things like, Robin likes skiing (i liked it as a kid, i’m affraid of it now :slight_smile: ) and Linux Lads podcast exists with tons of episodes available for listening.

These DAWs are absolute cringe.
They’re truly for children.
-A cute, fun DAW that one day you hopefully grow out of.

Compared to ‘handmade’ DAWs like Ardour, there is nothing to fear.

1 Like

You can quickly get up to speed about how AI works by asking an AI about it. That’s how I did it. Just ask gemini, claude or chatgpt to explain the concepts. The explanation probably rises more questions so just ask it to explain these also. There are loads of new concepts and terms in this field so it takes a little time to get familiar with everything.

You don’t need to subscribe to any service, just create a free account on google, anthropic or openai and you can use their service for chatting for free.

Here are some links to these service:

https://claude.ai/new
https://chatgpt.com/

1 Like

I disagree. Then again you might live somewhere that still allows lead paint, paraquat, DDT, radioactive clock dials, PCBs, asbestos, CFCs, etc.; all useful technologies in themselves, but generally rejected and banned due to their environmental impacts.

4 Likes

Most of those were banned, due far less to environmental impact, but due to the direct human life impact for the record.

  Seablade

Well yes, and.

In concept you are not wrong. But for instance, I had the ability today to compare running a task through an LLM I hand coded about a year and a half ago give or take. Hand coding it, took me about a day and a half between the research to find the correct libraries, and possibilities, testing on my local machines, then on the client machines, and deployment with tracking. In an LLM, it took about 7 revisions, checking code each time, and testing in a similar fashion, all together about an hour or two of work.

In these cases the scripts being referred to were not significant in terms of lines of code, which obviously likely skews the numbers a bit, and even though I wasn’t the strongest in the language it used, I could still read it well enough to follow everything happening, and determine when something wasn’t going to work correct. And the first pass got it 90% of the way there, the last 10% were the reasons behind the remaining revisions, eventually resulting in something far leaner and cleaner than the first pass for the record, but still much faster than if I did it by hand, and honestly likely as good, cleaner, and more efficient.

 Seablade

And this is definitely a major point. The issue is people choosing not to use it won’t improve this situation either. What WOULD improve that situation is people learning to use it properly, and proper review of output, etc. but honestly it will be a decade before that starts becoming common place and a lot of slop being generated in the meantime (See the current state of ‘AI Content Creation’ for an example)

Definitely short term you are absolutely right, it is likely to get worse before it gets better, at least for those that use it randomly. But proper verification etc. still can exist and is needed for this and can result in better quality output.

Seablade
1 Like

Not trying to be overly combative, but I just want to note that I don’t personally find this a useful way to understand what LLMs are doing. I find it more useful to think of them as … well, neural networks. A given input “lights” up the network which then feeds back on itself and at some point there is (or isn’t) something generated somewhere.

Technically, yes, you can talk about this in terms of recursion too, but I personally find that description of their operation to miss the most profound aspects of how they work.

1 Like

I don’t care a lot about the whole topic, but i found the IBM Technology videos on youtube to be rather clear and with statements not too exaggerating. .
They might well not cover what you are specificallly asking for, and you had do search for yourself if some of those videos, they seem to have plenty, cover it (it seems they more “broad”, but i didn’t really watch a lot).

The video is just a random example, as i liked the way that guy puts things.
“IBM Technology create software with LLM” might be a better searchterm for your needs.
Probably those IBM videos are not what you asked for, too basic, but you might give them a try, perhaps there is something. To me they seem more serious than most that i see.

2 Likes

Indeed, but I assumed it was a given that they are neural networks - what I was referring to was the way LLMs generate output, in as much as (again, massively oversimplifying here) the network predicts the next output token, and at some (or multiple) levels you recurse the process using the generated output as the next input, to further expand the output. I was oversimplifying to make a more general observation which I thought might be fun and interesting - I wasn’t trying to write an academic paper on a subject which I’d be the first to admit I am far from being an expert in.

The distinction I most likely should have made was between Machine Learning vs Generative AI.

A lot of things have been said already, but as a subscriber, I’d rather have my money be spent to pay people for coding than to tech bros that want to take away my RAM.

Especially since burnout in FLOSS is an important topic (Burnout in Open Source: A Structural Problem We Can Fix Together | Open Source Pledge), and the pressure to use LLM is just another factor.

3 Likes

I grapple with the same things at work; someone adjacent to my SW team single-person built a new prototype that, to his own surprise, is becoming a product, and everybody seems to love it and are now wondering what else we could build. (“Could”, not “should”… Any careful considerations that could be made are even more shallow now than before.)

It’s like a gun to my team’s collective head.

So, what to do? I worry tremendously about future maintainability and architectural rot. The guy above thinks technology is going to keep evolving and make such concerns a moot point. I point to no solid universally applicable solution having been found since the 1950s, so I doubt we have it now, nor will have it in the foreseeable future. Software expands to fill the available volume no matter what dimension you pick, so we’ll just keep expanding what we do with LLMs until they’re pushed to their edge.

I’m not going to make any call about ethical/environmental/etc. concerns. I think you’re somewhat in a “gun to head” situation, I imagine Harrison might also feel a squeeze from the new competitors.

My current thinking goes like this, with the caveat that this is something my team is moving towards, but we don’t have much empirical data on it yet.

  • Guiding light: Velocity may be important, but it’s not the most important quality metric. By far.

  • Guiding light 2: There are two kinds of quality: Immediately visible (features, speed, stability, etc.) and internal, hidden – we don’t build houses from straw foundations; but well-built houses can comfortably get new paint or have windows replaced.

  • Guiding light 3: The internal sanity of the codebase still matters, even as the LLMs improve. The more messy and confusing and whack-a-mole’y the code becomes after iterations innumerable, the harder the LLMs will have it, too. Human taste matters!

  • For big chunks of work like your 16k example, I would consider it “a library” like any other. It’s self-contained, test cases can be written (with much AI assistance), and it can be replaced in full if necessary (and with LLM assist that’s easier than ever before). For this to work it should be a library; something with an interface, a boundary, a specific purpose (e.g. an import filter, or an effect, or a UI component). Conversely, a set of functions and data types that are more of a “utility” nature end up permeating the code base and are are harder to replace, and we should be much more careful about those.

  • For bug fixes, the person submitting has to explain what the fix does and how it works, but not via an AI-generated summary; it’s about building trust in the solution. If the scope of the fix is too large, or unclear, we reject. The submitter has to rework, including breaking the fix down into smaller pieces. This is a kill-joy when on a dopamine-induced commit spree, but that’s what being responsible is all about.

  • Thousands of lines of code is not a bug fix; it’s either a feature in disguise or a mistake. Reject, make submitter rework, reframe.

  • Code that looks or feels alien is just rejected. Period.

  • It’s impossible to review thousands of lines of change in detail, so we often won’t. We’ll look at overall appearance, structure, architecture. But if the code doesn’t smell right, it’s out.

  • For things that touch the architecture, we reject unless the changes have been discussed by humans beforehand.

If humans are to remain in the loop, I think we have to insist on being there, insist on our humanity being the deciding factor that makes our product inspiring, fun, worth-while and wholesome, instead of a calorie-rich but unsatisfying all-you-can-eat buffet with free sodas.

Good luck. We live in interesting times.

[Edited to add: The point to some of the points above is that if incoming LLM code is now cheap (in time) to manufacture, it’s also cheap to rework/re-manufacture if it doesn’t sit well in the existing codebase. There are no longer any valid excuses about “it being too much work”. Again, yes, it’s less fun. But it’s more responsible.]

4 Likes