Definition
Artificial General Intelligence (AGI), also known as Human-Level Machine Intelligence (HLMI), is the ability of a machine to accomplish any intellectual goal at least as well as a human. Nils Nilsson defines HLMI as AI able to perform around 80% of jobs as well or better than humans. While narrow AI excels in specific domains, AGI is characterized by its breadth and ability to learn new skills across diverse environments.
Why It Matters
It represents a transition where machines can perform any intellectual task a human can, with existential stakes for our species. Getting this right could solve all other problems, while failing could lead to human obsolescence.
Core Concepts
- Universal Intelligence: AGI represents the transition from narrow, task-specific ability to general, adaptive ability.
- Learning to Learn: A key requirement for AGI is the ability to acquire new skills without explicit re-programming, mirroring the human capacity for Life 2.0.
- The Threshold of Life 3.0 (Technological Stage): Once a machine achieves AGI, it may quickly cross the tipping point of being able to design better AI, leading to an intelligence explosion.
- The 80% Metric (Nilsson): A pragmatic benchmark for HLMI based on the ability to replace human labor across the majority of the economy.
- Sparks of AGI: Early demonstrations of human-like reasoning and “common sense” in Large Language Models (LLMs) like GPT-4.
- The Unicorn Drawing Test: A benchmark for AGI reasoning where a model is asked to draw a concept (like a unicorn) by arranging component parts (horn, tail, legs) based on an abstract understanding of its “essence” rather than pixel-matching.
- Stack-ability Feats: LLMs demonstrating “common sense” by solving complex physical arrangement problems (e.g., stacking a laptop on eggs and a nail on a bottle cap) in a trillion-dimensional representational space.
- Moravec’s Landscape: AGI is achieved when the “rising sea level” of computer capability reaches the highest peaks of human competence (social interaction, physical dexterity, creativity).
- Expressive World Representation: The theory that the more parameters a model has (now reaching trillions), the richer its ability to mirror and eventually supersede human abstract reasoning.
- The Creativity Monopoly Challenge: AI challenging the quintessentially human capacities for creativity, art, and the manipulation of language, which were once thought immune to computing.