On the future of music: an interview with composer Robert Thomas (part 2)

Dom Aversano

June 27, 2023

This is part 2 of an interview with composer Robert Thomas. The first part you can read here.

Q. I associate you with Pure Data. Is it still your primary tool? If so to what extent do you think tools shape one's work?

I use Pure Data a lot because it's a universally deployable tool, and you can make installations and all sorts of bespoke things with it. It can be used in apps, game engines, or the web. Also, as it's open source, it doesn't have any proprietary licences associated with it.

Everything I work with is either my own library, which I licence to creators, or it's open source, so BSD licensed. It's easy to work with it from a business perspective and it's well-supported and incredibly stable. From a creative perspective, I use it because it's real-time. I think to do creative things musically and sonically you have to be working in a real-time environment, not a compiled environment. It's always good to be open to happy accidents, which when you're working in real-time can happen, but when you're not it is less likely. So I don't like compiling when I'm working.

And then, what was the second part of your question?

Q. The extent to which the tool may or may not be shaping your work.

I think in some ways it doesn't shape it at all because PD is just Digital Signal Processing and you can do anything you want with it. When you make a new patch it is really just a white page - it's very open, flexible, and at the same time, terrifyingly, overwhelmingly, even dangerously open-ended.

Therefore I've developed a good 'muscle of restraint' to control exactly what I want to do. You don't want to go down all kinds of undefined meanderings in PD. It’s not the place to do that. It can be interesting to try new things and be open to accidents but there's a balance, you need a strong idea before you start coding because the program is so open.

The constraints I was talking about earlier were not constraints about what can be done with the software, they are about what is possible with the wider technology. Some aspects of personalisation can be very difficult to know, such as with contextual and emotional detection, or biometrics. There are limits to what we can and can't reliably understand, which provide creative constraints and require you to work within a framework that is sometimes relatively simple.

A good example would be when you are working with an accelerometer to understand how someone is moving, or with a GPS to work out how fast someone is going. There is only a certain amount of fidelity you can get from that. There are practical considerations, like if you have the GPS whacked way up in accuracy on a device it's going to drain the battery, and the user is quickly going to get really annoyed with the experience.

So you need to say, 'Well, OK, we're going to make a judgement that it is OK after this amount of movement, or we're going to look at the GPS over this amount of time and decide, when we think they are really moving, which will mean there's going to be a sudden change in the state of the user. What could we do with an accelerometer? We can look at how it is changing over time and try to use step detection if they are walking. There is a lot of work in getting such algorithms accurate, which places a boundary around how you creatively respond. I think that is what shapes what you do creatively.

A lot of the time what I am doing inside of DSP is relatively simple, and I try to make it as elegantly simple as I can, from the perspective of stability and reliability but also CPU and memory usage. The most desirable systems are actually the most simple and elegant. Those are the golden systems.

Another issue with tools which happens in DAWs, but especially inside of programming, is there is satisfaction in solving complex tasks or challenges in clever ways. It's very dangerous thing to get sucked into if it takes you away from creating a good musical experience, and one major problem I see in the space with a lot of projects – and I've been sucked into this on some projects as well – is that you try and create something that's a really clever system, but it sounds crap. Compared to a studio production where someone is working with off-the-shelf tools and a DAW using amazing plugins and rendering it down loads of times with intricate and polished production and writing. That's what we have to compete with. We have to be at the same level as that and better in real-time.

So if a system is super complex and really rewarding as a programming project but sounds crap, that's no good, because music lives or dies on emotional experience. If people don't enjoy it as a musical experience, it doesn't matter how clever it is. I see that as the biggest danger in this space.

Q. That's one risk I see with generative music. Algorithms are generating what is being heard, which is different from a live performer where it comes directly from them. With generative music there is the intermediary of the algorithm, and a risk of things sounding hollow and dehumanised. How does one get that deep emotional experience into the work?

Well, I think that's the art of creating algorithms from an artistic and humanist perspective, which has nothing to do with what is happening in machine-learning music at the moment. It is absolutely the opposite. I find it frustrating that the term generative AI has been co-opted by the machine-learning community because the approach to generative music that Brian Eno, Autechre, and I take is to human craft algorithms. It is the polar opposite of throwing everything into a massive deep-learning network and never knowing what is happening inside it, which is what deep-learning language models do. This space is about carefully crafting algorithms to embody as much of yourself and human expression as possible. That is what I am about.

I've heard Brian Eno talking about Steve Reich's influence on him and how he crafted the music through systems. Reich was very specific about it, which made this very interesting possibility of outputs as a generative system. So it is about the seeds and the rules.

When you're crafting things you need to listen to them for enormous amounts of time to hear all these different states, making sure it has an emotional and artistic impact. I think where things go wrong is when you try to either make generic algorithms that will make generic hip-hop, EDM, or ambient music, or even worse, a rule-based system that can make all kinds of different music. When you are that broad there is never going to be any specific quality to it. The worst is when you give up all control and completely entrust it to a network inside the system, such as deep-learning and large-language models where nobody understands what is inside the system. We are trying to make systems to understand what is inside them! How can that be an artistic endeavour? An artistic endeavour is a process of trying things and learning them. If systems are impenetrable I think it's very challenging to have an artistic interaction with them. I believe there are ways that machine learning will be helpful, but a fully automated, unsupervised, completely autonomous system is not particularly creative.

With your question, I think it is the important thing. We need to incorporate many aspects of what we do into the system; things we all do. When I do workshops with musicians I ask, when you are playing what are you doing? Okay, you're doing these types of patterns rhythmically. Oh, you're doing these kinds of intervals. You're doing these types of phrasings. You're always swinging in this way. You're like, 'I'm not, not all of the time', but when you're doing that musical thing, that idea, what are you doing?

These are the things we need the algorithm to do: to distil down a process which is both the artist and an extension of them. It embodies many aspects of the artist, but it can do things that no artist can ever do: create live music for 1000 people all over the world all at once, which is different for each person. Those are the possibilities I'm interested in.

Here are some links if you are interested to know more about Robert Thomas’s work