How Humans and AI Learn to Live Together by Google DeepMind

​
This talk is based on recent work by Google Deepmind called 'Societal and technological progress as sewing an ever-growing, ever-changing, patchy, and polychrome quilt'.

The first author of the work, Joel Z. Leibo, a senior staff research scientist at Google DeepMind, will share the details of it with the BuzzRobot community.

Current 'one-size-fits-all' AI alignment is based on a misguided search for universal values, ignoring the moral diversity central to human societies. We consider here instead an alternative theory where collective flourishing arises from individuals learning and adhering to local, context-dependent norms of appropriate behavior — a process that can be modeled computationally as predictive pattern completion.

This approach reframes the goal of AI safety: instead of seeking an impossible moral unification, we should design systems that manage conflict and navigate disagreement by adapting to diverse, community-specific standards.

Read the paper
Join the BuzzRobot community on Slack