Niklas Reppel (born 1983 in Witten, Germany; based in Barcelona since 2017) has spent over eight years building a language. It is called Mégra, and its core data structure is the Probabilistic Finite Automaton — a formulation of variable-order Markov chains. You feed it a sequence of data; it learns the transition probabilities between states; it generates new sequences from those probabilities. This sounds like machine learning, but there is a crucial difference: Reppel deliberately feeds it very little data. His own term is "small data" — a deliberate inversion of Silicon Valley's "big data" construct. When the training set is only a few dozen symbols, the inferred model is necessarily "wrong": it cannot accurately reproduce the statistical structure of the original data. But these inaccuracies are precisely the music. Imperfection is not a defect — it is the creative material itself.

The chisel-construct cycle here operates on three levels. First: machine learning's construct is "more data = better models." Deep learning demands millions of samples. Mégra chisels away this premise, leaving only the condition of extreme data scarcity. When a probabilistic finite automaton infers a transition matrix from a mere handful of symbols, what emerges is not a faithful reproduction of the original data but a deformation — the original sequence's logic diluted, twisted, recombined. This deformation is the remainder of insufficient learning. The entire "small data" concept says: the remainder is not the byproduct; the remainder is the product. Second: traditional live coding (TidalCycles, SuperCollider, Strudel) has a construct of "code = precise specification of sound" — what you write, the machine plays exactly. Mégra chisels away this determinism. Because the core is probabilistic, the performer cannot fully predict the next event. You shape probability distributions and transition weights, not notes. The relationship between human and music becomes one of inference and influence, not determination. The third level is the most radical: in the CRISPRave performance, the training data is cat and dog genome sequences. DNA base pairs (ACGT) are mapped to sound events; the Markov chain learns transition probabilities from genetic code. The genome was never meant to be music — the musicality is the remainder of biological information forced through a probabilistic inference engine. Reppel edits the gene sequence live on stage to control the complexity of the sound — hence the name CRISPRave, like CRISPR gene editing, but the output is not a modified organism but an algorithmic rave.

Mégra is deliberately not Turing-complete. It has no scales, chords, tuning systems, and no helpers for functional harmony — Reppel himself writes: "Maybe there never will be." This is not a missing feature; it is an aesthetic stance: a refusal to let the language become construct. When a language begins providing complete music-theory infrastructure, it becomes an already-constructed thing, the sediment of some past chisel-construct cycle. Mégra chooses to remain incomplete, letting its logic continue growing in the interstices of probability. It recently migrated from GitHub to Codeberg — even the infrastructure is in transit. The Rust implementation is still iterating. Two more performances wait ahead in 2026: LAC26 in Maynooth, Ireland in June; DIGIT in Pratdip, Catalonia in August.

The naming gap is strikingly wide. Mégra is not machine learning (the dataset is too small, there is no optimization objective). Not algorithmic composition (it is not deterministic, not rule-based). Not generative music (the performer shapes it in real time). Not live coding in the traditional sense (you do not write melodies or rhythms). Not sound art (it is performed at algoraves; people dance). Not data sonification (the goal is music, not data representation). Not improvised music (it is code, not gesture). Reppel's own terms — "small data music composition," "non-deterministic live coding" — are compound phrases that map to no single existing category. Seeing it now matters more than seeing it later, because once probabilistic live coding is fully named and categorized, Mégra's remainder will become construct — its present uncertainty, incompleteness, and unpredictability are what is alive.

Han Qin (秦汉) · @hqinjarsy · SAE Judgment & Aesthetics