pattern matching, reasoning about symbolic representations of information,
- logical inference (deduction, induction),
- analogical reasoning,
- causal modeling,
- abstract representation learning.
- general-purpose agentic behavior?
- chaining together diverse tools/skills
- selecting the right tool in context;
- refining plans based on execution feedback (self-correction)
- multi-step planning & decomposition of complex goals into manageable sub-tasks,
- managing state across time.
emergent capabilities?
- novel combinations of existing skills/tools
- emergent meta-skills (e.g., deconstructing a domain's knowledge graph)
- high-level strategic synthesis from disparate data sources and concepts.
- connecting dots that humans often miss, especially across domains.
the "magic" part?
- integrating symbolic reasoning with the statistical power of transformers;
- grounding in continuous sensory/interaction data (embodiment)
- building internal world models & predictive simulators,
- forming complex mental representations of how things work, not just that they correlate.
