Manifesto
Why this name, and what we are trying to build.
Why this name?
Allendia takes its name from Salvador Allende, the Chilean president overthrown in 1973 by a fascist coup, and from his Cybersyn project. Cybersyn was an attempt at decentralised economic management. It used technology not to concentrate power but to distribute it, against an imperialist blockade.
That history shapes what we do. We are obviously not doing the same thing; the contexts are too different. But it carries an idea we hold close: technology can be a tool for emancipation, not only for control. At Allendia, at our very small scale, we try to put advanced AI in the hands of musicians rather than platforms.
A tool, not a master.
The Chilean example reminds us that technology is never neutral. Or rather, it is always the ground of a political fight. Ours is meant to stand with the people who want to democratise the means of musical production. Not to replace musicians, but to give them new tools, instruments that surprise and inspire, without taking the gesture away from them.
Navigating contradictions
We are anti-capitalist. We reject startup ideology, growth at any cost, fundraising, the hyper-optimisation of work. We want a lasting company, self-financed, with a small team, without multiplying tasks or squeezing people. We want to pay the people we work with properly, and owe nothing to investors who would ask us to grow or die.
That does not solve our central contradiction: founding a company means entering the system we reject. But we want to bring our contradictions into the light and own them fully, so we can navigate with them. We build our practice and make our decisions from that tension, without erasing it. In practice, that means accepting that we will slow down, turning down lucrative opportunities when they take us away from our values, and discussing every dilemma together.
In praise of détournement
We have invented a MIDI generator that looks like no other. Because the thing everyone is doing right now is deep learning: models trained on millions of works protected by copyright, without consent or payment. Those models consume astronomical amounts of energy, and what they produce tends to homogenise content and reduce creative diversity.
Automation is sometimes sold in the name of easing working conditions: more free time to do what you love, like music or drawing. But the first concrete uses are image and music generators that go straight after the people who manage to make a living from those practices, already precarious enough. It is a con, and we refuse to contribute to it.
EVY runs on a recurrent neural network with random connections, entirely untrained. It has never listened to a single second of music. It invents, it improvises, it glitches.
Our inspiration is the history of technological détournement in music: Link Wray stabbing his amp with a pencil and inventing distortion; the hip-hop DJs of the 1970s who started spinning vinyl backwards on their turntables and invented scratching; Cher and her engineers pushing Auto-Tune to the limit on Believe, creating a robotic effect that nobody had yet dared use as an aesthetic. Each time, it was not the intended use that changed music. It was the use it was never meant for, the gesture that forces the tool out of its function. EVY is our attempt to make a tool that invites people to hijack normative AI from the moment it is designed.
Concretely, this changes everything:
- No protected data was used to train EVY, because there was no training. A simple, radical way of not pillaging artists' work.
- EVY uses very little energy. Training is the most expensive part, energetically speaking. No training, no overheated data centre. Even in use, with 2 to 50 neurons instead of the millions in large models, the computation is so light that it runs instantly on any laptop, with negligible consumption.
- EVY does not produce finished works, but MIDI phrases that musicians can re-record, edit, and work into their own composition. The human stays at the centre. AI is an instrument, not an end in itself.
Randomness is political
We chose randomness, against the determinism of large models. Because predictability is the weapon of platforms. The more predictable a system is, the more it can be monetised and normalised. We want surprise, happy accidents, unforeseen events that open gaps.
This approach belongs to a long tradition of serendipity and constrained play. The surrealists' exquisite corpse was not simple entertainment: it was a weapon against the stifling rationalism and productivist logic of a bourgeois order that already carried the threat of fascism. Randomness was a method for making the unexpected happen, for short-circuiting control, for letting truths emerge that reason cannot reach. We take up that gesture, at our scale, with our tools.
EVY exposes around fifteen parameters to guide the network. No prompt here. You turn knobs, you sculpt live, as on a synth. You gradually learn what each parameter does, you gain control, you develop an intimacy with the instrument. That is what we call sculpting randomness. We do not delegate creation. We channel chaos.
How we use LLMs
We use AI every day in our work: to write code, generate documentation, solve technical problems faster. These tools, for all their faults, have one virtue: they put skills that used to be out of reach within reach of a small team with no capital. To build a plug-in like EVY, you need low-level programming, audio, MIDI communication formats. Used well, an LLM saves us months of work. It lets us prototype quickly and test ideas without hiring three more engineers. It means we do not have to raise funds, take on debt, or depend on investors. It is a tool for keeping our independence.
We are nonetheless worried by their colossal impact on the world. From now on, when their performance allows it, we will use Mistral and DeepSeek. Mistral first: a French model whose values are (slightly) closer to ours, and which can be hosted by more responsible providers. DeepSeek next, for its power and very low cost, and because in the current geopolitical context, refusing US hegemony is a political priority for us. We do not want to feed GAFAM, their working conditions, their transphobia at the level of the state, their active support for the genocide in Gaza, their environmental and economic deregulation.
We know that neither Mistral nor DeepSeek is a perfect tool. Far from it. Ideally, we would use (and contribute to) LLMs that are more respectful of personal data, the environment, and working conditions, hosted by non-profit organisations, open weight and open source, and actively working on the alignment problem. If you have ideas, information, or proposals, write to us.
What comes next
In the future, we do not rule out using machine learning for some features. But we commit to:
- Using open-source pre-trained models wherever possible, to limit our ecological impact and avoid financing GAFAM.
- Never using protected data to train new models without explicit agreement and fair payment to artists.
- Never outsourcing data annotation under precarious conditions.
This manifesto is a statement of intent, a compass for our own decisions. It is here so you know where we start and where we are going. It will change as our practice changes, mistakes included. In the meantime, EVY is here, ready to be manhandled. If you have ideas, anger, tracks, or questions, write to us. We answer.