2026 - *
I subsequently developed an interactive interface in Max (Cycling ’74) through which I can navigate the invisible landscape contained within the model, both manually and through automated processes based on random algorithms. The controls allow exploration of eight latent dimensions of the model — dimensions through which the artificial intelligence internally encoded my sounds after analysis. However, the meaning of these dimensions remains not directly readable to the human user, thereby leaving considerable space for musical experimentation.
​
Depending on parameter combinations within this “black music box,” it is possible to recreate spectral characteristics clearly recognizable from the original acousmatic composition. However, when parameters are modified at speeds or temporal scales significantly different from the articulation of the original materials — such as amplitude envelopes or morphological evolution — new materials emerge in the form of potentially infinite textures resembling either a generic waveform synthesizer or heavily artifacted time-stretching processes.
By using “Big Hungry Parrot” as an instrument in live performance — as I did during the “AI Impro Camp” at jazzahead! 2026 in Bremen — I can play with:
​
-
the speed of parameter changes, determining the emergence of recognizable (“sample-like”) or unfamiliar (“resynthesized”) identities;
-
the number of dynamically controlled parameters, either randomized or fixed;
-
the scaling of parameter ranges, determining fidelity to the original spectra and the degree of distortion, or the emergence of sonic identities foreign to the original composition.
Throughout every aspect of this process — from training to sonic result and control strategies — I perceive strong connections with the character and context of the original composition Atom, in which I projected my sonic imagination of the birth of matter in the universe. For this reason, I consider it essential to expose the audience to the original piece, or fragments of it, before or during the real-time synthesis process, thereby sharing the point of departure for the transformation.
Because the concepts of tensors and latent spaces require a high level of mathematical abstraction, I am also developing a spatial representation of “Big Hungry Parrot” in order to make the sonic process perceptible on a sensory and emotional level, reconnecting it to the concreteness of spatial experience.
​
I began with the virtual visualization of a three-dimensional form generated from the 369 points used to represent the organization of sonic characteristics within the latent space of the model. This agglomeration resembles a complex constellation that is difficult to decipher visually.
I therefore decided to rotate it continuously along all three axes according to the changing speed of the eight latent dimensions. These values are in turn converted into three-dimensional coordinates used to move a light source through the virtual model, revealing only partial regions of the structure at any given moment.
The presentation of this visualization within physical space remains an open question in the project. At present, I imagine reconstructing a reduced spatial configuration of the constellation using 8–16 resonant objects — such as loudspeakers or spherical sculptures activated by sound exciters — distributed throughout the space. Just as light illuminates portions of the virtual structure, sound moves through the three-dimensional environment using ambisonic techniques and activates resonant objects encountered along its virtual trajectories.
​
I realize that sharing these sonic, visual, and spatial aspects with the audience is fundamental for opening a space of reflection on artificial intelligence as a process, beyond merely admiring its visible results.
At this point, I would like to express my concerns regarding the sustainability of artificial intelligence within the field of electroacoustic experimental music.
As mentioned earlier, training a neural network model requires repetition so that the model can progressively refine its internal representation of sonic material. Despite understanding the necessity of this process, my enthusiasm for experimentation collided with the following warning when I first read the tutorial for training RAVE on custom sounds:
“Warning: The duration of a full RAVE training is difficult to predict exactly, as it depends on the chosen configuration, the data, and your machine. Usually, the first training phase lasts about three or four days, and the second phase may take from four days to three weeks.”
​
Three weeks? In other words, 504 hours during which my computer would need to remain continuously powered? I began wondering how much energy such a process actually consumes. I therefore decided to conduct an experiment and publicly share the results in order to stimulate collective reflection.
A power meter connected to my laptop charger measured approximately 9.1 kWh over about 24 hours. What does this correspond to? By consulting Claude.ai I learned that a small apartment equipped with common household (2-person flat) appliances may consume approximately 7.8 kWh in a single day (only electricity).
​
One electronic device: 24 hours, approximately 9.1 kWh
vs
One entire apartment: 24 hours, approximately 7.8 kWh
Considering the current ecological crisis and our shared responsibility within it, and considering that these technologies are available to virtually anyone who wishes to use them — while unsuccessful artistic results can simply be discarded and retrained at further ecological cost — the disproportion between resource consumption and artistic purpose appears deeply questionable.
​
Having undertaken this path despite understanding its implications, I cannot avoid either self-criticism or the public sharing of my doubts and experiences. Otherwise, this energetic sacrifice would have been entirely meaningless. For this reason, I intend to continue using this already-born “Big Hungry Parrot” as it currently exists, without further optimization or retraining. I hope in this way to raise awareness about individual responsibility in the use of this technology.
Although I deeply value experimentation, I cannot ignore the environmental consequences of my artistic choices on the community surrounding me.
These reflections will become part of the artwork itself, for example through a short accompanying text presented alongside the installation or performance, uniting aesthetic experience and critical reflection within a shared space.
​
I would like to conclude with the words of Joseph Weizenbaum, whose writings accompanied me during this process of critical awakening and whose thoughts I hope may inspire reflection and change in those who encounter them.
​
​
“There have been many discussions about ‘computers and human thought.’ The conclusion that forces itself upon me is that the relevant problems here are neither technical nor mathematical in nature, but ethical. They cannot be solved by asking questions that begin with ‘can.’ The limits in the application of computers can ultimately only be expressed in statements containing the word ‘should.’ The most important fundamental insight arising from this is that, at present, we know of no way to make computers wise, and that therefore we should not assign computers tasks whose solution requires wisdom.”
Joseph Weizenbaum, Computer Power and Human Reason: From Judgment to Calculation (1976)
Big Hungry Parrot
“Big Hungry Parrot” is both a sonic, visual, and spatial exploration of the concept of latent space in neural audio synthesis and a critical reflection on individual responsibility in the use of artificial intelligence technologies.
The project originates from the consideration that, just as a musical instrument can be modeled by training a neural network on recordings capturing different playing techniques, an entire acousmatic composition can conceptually and technically undergo the same process. If a composition is regarded as a collection of sound materials independent of their temporal organization, compositional transformations, and superimpositions, artificial intelligence can encode their overall timbral characteristics and subsequently decode them, generating real-time sound synthesis based on new combinations of these characteristics.
By imposing a new temporality onto this abstract image of sound materials — both at the macrostructural level (which sonic identities follow one another?) and at the microstructural level (which spectral characteristics evolve within a single sound?) — it becomes possible to obtain a continuous flow of shifting sonic perspectives on source materials that remain permanently fixed on a medium.
​
As a symbolic starting point for my own exploration of this technique, I decided to use the dataset of my first acousmatic composition, Atom (2009), to train the RAVE neural network model developed by IRCAM. After a total of 28 hours of training on my high-end gaming laptop — consuming approximately 12 kWh during that time — I obtained an instrument capable of representing and generating the sonic characteristics of Atom. The original composition is played before and during the activation of the model in order to share a clear reference point for the reconstruction process with the audience. I will return later in this text to the ethical and ecological implications of this process.