Kevin Bretz

I work on real-time interactive systems, machines that have to perceive, decide, and act within a constantly evolving context. I also have a passion for reinforcement learning, as well as all things audio. These come together in my current focus, the musical setting, where I want to contribute towards AI that plays with people rather than instead of them.

Research interests

What intrigues me the most is the live scenario, where a model cannot wait for future information or reasoning before committing to an action, much like we rely on our intuition during certain tasks. I envision an agent jamming alongside a performer: holding the groove, providing accompaniment, communicating nonverbally, anticipating changes, reacting to the unexpected, and incorporating its own improvisations. And because such an agent would create the way we do, live and from intuition, it would also be a way to explore and push the boundaries of computational creativity, for example via reinforcement learning and multi-agent self-play. I would like to spend my PhD working towards this.

The tools I want to build

Most of the attention in (generative) AI goes towards automation, and not without reason. Prompt-based and agentic workflows provide enormous efficiency gains for countless applications across the board, creating a huge financial incentive for further development. Yet I argue that, for some applications, a human in the loop remains indispensable. Nowhere does this become more apparent than in the arts, where the process matters at least as much as the outcome. Real creation is a back and forth, a culmination of countless individual decisions, each fueled with intent and shaped by factors outside of our conscious control. Only a small fraction of that can ever be conveyed within a prompt. So I want to build the other kind of tool, the kind you play with rather than prompt, one that complements your process and extends your palette rather than standing in for them.

Collaborate, Don't Automate.

I came to this from the technical end, a builder more than a player. My affinity for audio was forged during my media technology degree, where I became enamored with audio and music technology, both in software and hardware form. It is also where I was first introduced to neural networks. Fast forward to my master’s in AI, and I cannot help but notice a disconnect between the sheer amount of attention that AI art is getting and how the general public, and artists in particular, feel about it. To me, this divide stems from the current emphasis on one-shot, fully automated systems, and closing it means building tools that put agency back in the hands of the artist. If not for their sake, then for the sake of the art itself. Without the human, can we even call it art?

Artists are not going anywhere - but neither is AI. I want to make a career out of building tools that provide artists with new opportunities, instead of taking them away.

My projects cover the ground this vision stands on, with work ranging from reinforcement learning to generative audio, and some additions I am equally proud of. Personal projects are in development and will be featured after publication. The journal is where I think it through in writing. If your group works on anything close to this, I would love to hear from you.