Making Music and Art Through Machine Learning - Doug Eck of Magenta

TL;DR

Doug Eck, lead of Google’s Magenta project, discusses using machine learning to create open‑source tools for music and art. He emphasizes that current models (e.g., LSTMs) produce primitive but musically interesting results, especially when musicians treat them as improvisational partners rather than composition generators.

Key insights

• The Brian Eno quote about new mediums’ “ugly” signatures becoming cherished directly motivates Magenta: the team deliberately explores “broken” or glitchy outputs as creative material.
• Early LSTM networks for music were “very simple” and rarely produce good long‑range structure; they work best as call‑and‑response partners for skilled improvisers rather than as autonomous composers.
• Evaluating generative music models is an unsolved problem – Magenta hasn’t yet built a viral app with human feedback loops to improve quality.