AI Learned to Bow a Violin — and Hit a Copying Limit

Bowing a violin looks easy and isn't. Too much pressure and the note screeches; too little and it vanishes into a whisper. Researchers at Columbia University built a detailed computer model of a real violin string — using measurements from actual strings rather than guessed numbers — and then trained six different AI programs to bow it. The goal is better virtual instruments and, eventually, robots that can play convincingly. Their model nailed the physics: it reproduced the correct "sticking and sliding" behavior of a bow on a string 89.1% of the time, against an ideal of 90%.

Two findings stand out. First, they corrected a famous 1960s rule (called Schelleng's law) about the lightest touch a bow can use before the note dies. The old formula used squared terms; the real one doesn't. That's a small but genuine fix to how we understand one of the oldest instruments on earth.

Second — and this matters far beyond music — the AI with memory beat the AI without it. A "recurrent" network (one that remembers what just happened) handled mid-stroke disturbances better, like a violinist recovering when a train rattles the floor. A feedforward network (no memory) only won when the test started somewhere the instrument barely plays anyway.

But here's the honest catch, and it's the part worth remembering: not one of the AI controllers outperformed the simple lookup table that generated its training answers. When researchers plotted the two against each other, the AI was clearly better where the old rule broke down, and clearly capped by it everywhere the rule worked. In plain terms: AI trained to imitate a teacher learns the teacher's blind spots too. The researchers call this a "supervision ceiling." They also found that two common quality scores mislabeled bad sounds as good ones — a reminder that measuring art is harder than it looks.