Most software described as AI music is trained on recordings. EVY is not. It runs a random recurrent network inside the plug-in, with no dataset and no network connection. The notes you hold are the only musical material it receives.
What random means here
The connections in the network are not learned from examples. They are random at the start, and the network's behaviour comes from the way those connections feed back into one another as notes arrive. A few dozen neurons are enough. The memory is short by design: it holds the last phrase, not a catalogue.
This is the research idea that became EVY. Random networks have been used to make computation cheaper and simpler; here the same property makes a musical instrument that behaves differently each time you play it. EVY was developed within Inria Startup Studio, and Allendia exploits it under licence from Inria.
Why it matters
The practical consequence is that nothing is borrowed from anyone's recordings. There is no training corpus to license, no generated audio to clear, and no upload of your playing. The instrument runs offline.
It is also small. A few dozen neurons can run inside the plug-in on a laptop, which keeps the cost of a session low and the response immediate. You can hold a chord in a live set without waiting for a server.
What it does not mean
Random is not the same as uncontrolled. The controls on the front panel shape the network: capacity sets how much it can remember, inertia resists change, echo favours recent notes, feedback changes how strongly the last choice influences the next. Notes limits the vocabulary to the chord you are holding.
You supply the harmony and the destination. EVY proposes the line. You decide whether it stays. There is no prompt field.