AGI (Artificial General Intelligence)

← AI Watermarking & Content Detection Mechanistic Interpretability →

The contested idea of AI matching humans across most cognitive work, rather than excelling at one task. Definitions vary so much that serious frameworks now grade levels of generality and autonomy instead of arguing over a single finish line. The term drives lab missions, policy, and billions in capital.

Key terms

Narrow vs general AILevels of AGIAutonomySuperintelligenceCapability thresholds

Learn these first

← Large Language Models (LLMs)

Where you meet it in the real world

Lab charters and safety policies, compute governance debates, how AI progress gets measured

Deep dive · 5 min

Types of AI: Narrow, General, Generative

Narrow AI does one job: translate, drive within mapped conditions, recognize digits, predict the next word. Every AI system deployed anywhere today is narrow, including the most impressive chatbots. A model that writes poetry and code is still doing one learned job, next-token prediction, applied broadly.

General intelligence is a goal, not a product

Artificial general intelligence (AGI) would match humans across the full range of cognitive work: learning new domains from scratch, transferring skills, operating over months-long goals. It does not exist. Serious labs treat it as a research direction with wildly uncertain timelines, and any product marketing that implies its arrival deserves your skepticism. The interesting live debate is how far scaling current methods gets, and it is genuinely unresolved among experts.

Discriminative vs generative

A useful engineering split: discriminative systems output a decision (spam or not, which digit, approve or decline), while generative systems output content (a paragraph, an image, audio). The digit recognizer you can train in this tutorial's 3D explainer is discriminative: 784 numbers in, one of ten labels out. An LLM is generative: it manufactures the next token, thousands of times in a row. Agents are a usage pattern layered on top: a generative model calling tools in a loop toward a goal.

In one glance