Magenta

Robin Sloan describes using machine learning to generate fictional music for the audiobook of his novel Sourdough, explaining how SampleRNN learned from Croatian klapa singing. He discusses how artists can treat machine learning as a creative instrument, especially when people can guide and curate its output. The conversation also explores robots in kitchens, the interplay of technology and older crafts, and the fragility of digital projects as platforms change. Sloan reflects on social media, the history of blogs, and writing hopeful fiction about technology without ignoring its costs.

Doug Eck describes Magenta’s work on open source machine learning tools for music and art, including sound generation with NSynth and sketch generation with SketchRNN. He explains why current models can produce interesting, imperfect results, and how artists can use those results as material to explore rather than as finished work. The talk covers challenges in evaluating generated media, improving tools for musicians, and using reinforcement learning rewards to steer generative models. Eck argues that machine learning is best treated as part of a creative toolkit, with people shaping how it is used.

Tim Hwang describes how his Google public policy role connects AI research with governments, regulators, and social impacts. The talk explores fairness and privacy, adversarial vulnerabilities, automation, and the challenge of making machine learning understandable and useful in products. Hwang discusses opportunities in creative tools, small niche applications, cloud machine learning, and on-device learning, while emphasizing that good interfaces and domain knowledge matter alongside research. He closes with recommendations for learning about machine learning and its history, including past cycles of AI hype.