Jeff Crume maps intelligence across perception, thinking, learning and action in this 13-minute IBM Technology explainer.
Intelligence has several dimensions
- Perception includes recognizing patterns, language, objects, spatial relationships and social context.
- Thinking includes inference, causal reasoning, problem-solving, prediction, abstraction and creativity.
- Learning includes memory, feedback, adaptation, transferring knowledge and recognizing limitations.
- Action includes planning, pursuing goals, navigation, coordination and communication.
Three labels, three capability profiles
- ANI, artificial narrow intelligence: capabilities bounded by particular tasks or training, even when performance exceeds humans in some areas.
- AGI, artificial general intelligence: broad, adaptable capabilities roughly matching humans across domains in Crume’s framing.
- ASI, artificial superintelligence: capabilities substantially exceeding humans across cognitive domains.
- His sketch compares a jagged narrow system with broad human-level and broadly superhuman profiles.
Strong performance in one area leaves other weaknesses
- Crume classifies current systems, including large language models, as narrow despite their breadth.
- He points to factual errors, uneven reasoning and unreliable physical-world understanding.
- Caveat: this is a conceptual taxonomy; the talk supplies no agreed AGI test or measured threshold between categories.
Match capability with control
- Crume calls for technical safeguards, security, testing, governance and human oversight.
- His car metaphor puts humans at the controls rather than locked in the trunk of a capable autonomous system.
- His proposed kill switch and ethical framework are governance recommendations, with no implementation or effectiveness test shown.
“Capability minus control equals risk.”