Symbolic behavior in AI: some possible lessons for whales?
summary
Argues symbols are conventions whose presence is best judged by behavior, not labels, and lays out six behavioral criteria—receptive, constructive, embedded, malleable, meaningful, graded—and contrasts architectures and training regimes. Shows an embodied 3D agent generalizes perfectly on a compositional color/object task while a static image classifier fails, critiques AlphaZero-style opaque search and current neurosymbolic toy domains, and highlights pragmatics gaps in large language models. Recommends interactive, social experiments (including for whale communication) to test whether learned patterns are truly meaningful.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 1:10The paper
- 4:27The six behaviors
- 18:22Where does symbolic behavior come from
- 21:44Other perspectives on symbols
- 25:38Interaction vs observing
- 33:09Thoughts on the paper
- 34:54Questions
- 38:39Vector embeddings
- 43:06How do you think about this now
- 45:34Social interaction
- 49:19Examples of strictly not symbolic systems
- 51:07Iconicity and symbolic behavior
- 54:11Symbols and narrative structure
related talks
43mAI Systems: Foundations for Artificial Minds
01h 03mGenerative AI and What Is Meaningful Sperm Whale Communication
45mIllusion of Life
19mBeyond agents: AI as a creative partner
27mEmerging Patterns Agentic AI
21mLearning from History: The Past as a Compass for Systems Engineers in the Age of AI
43mThe Evolution of Design Patterns in the Age of AI
23mGhost in the Machine: AI, Spirituality and a New Humanism