'World Models,' an Old Idea in AI, Mount a Comeback
8 days ago
- #AI
- #AGI
- #World Models
- World models in AI are internal representations of the environment that help AI systems make predictions and decisions.
- The concept of world models dates back to 1943, introduced by psychologist Kenneth Craik, linking cognition with computation.
- Early AI systems like SHRDLU used simple world models, but these couldn't scale to complex environments.
- Deep learning revived interest in world models, with large language models (LLMs) showing emergent capabilities attributed to them.
- Current generative AIs learn 'bags of heuristics'—disconnected rules that approximate responses but lack a consistent world model.
- Robust world models could improve AI reliability, reasoning, and interpretability, but their development remains uncertain.
- Major AI labs like Google DeepMind, OpenAI, and Meta are exploring different approaches to creating effective world models.