Artificial General Intelligence (AGI)

Artificial General Intelligence (AGI) refers to a hypothetical AI system capable of performing any intellectual task that a human can perform — reasoning, learning, planning, and adapting across arbitrary domains without task-specific training. No AGI system exists as of 2026. The concept occupies a central position in AI strategy, safety research, and policy debate despite (or because of) the absence of a clear technical definition or consensus on when or whether it will be achieved.

## Why the definition matters

The challenge with AGI is that "general intelligence" has no agreed-upon measurement. Human intelligence itself resists a single metric — we have spatial, linguistic, emotional, social, and abstract reasoning capabilities that are partly correlated and partly independent. An AI that outperforms humans on every language benchmark but cannot navigate a new physical environment is not obviously AGI. An AI that can learn any skill from scratch in hours but requires explicit task specification is not obviously AGI either.

Different researchers define AGI differently:

- **Legg and Hutter (2007)**: an agent that can achieve goals in a wide variety of environments.
- **OpenAI (2022)**: "highly autonomous systems that outperform humans at most economically valuable work."
- **ARC-AGI benchmark (Chollet)**: a system that can solve novel reasoning tasks from a small number of examples — measuring adaptation, not just recall.

## The current state of large language models

Frontier LLMs (GPT-4o, Claude 3.7, Gemini Ultra) exhibit multi-domain competence that superficially resembles generality: they can write code, summarise legal documents, analyse images, translate between languages, and explain scientific concepts — sometimes at expert level. But they remain what researchers call "very wide ANI" (Artificial Narrow Intelligence):

- They are trained on fixed datasets and do not learn from ongoing interactions without fine-tuning.
- They hallucinate — generating plausible but false outputs when their training distribution does not cover the input.
- They struggle with novel abstract reasoning tasks (as the ARC-AGI benchmark demonstrates).
- They lack persistent beliefs, embodied perception, and causal world models.

## Enterprise AI strategy implications

The most common mistake in enterprise AI planning is waiting for AGI. Current ANI capabilities — deployed correctly — can automate 30–60% of knowledge-work tasks in targeted domains. That ROI does not require AGI to materialise.

The second most common mistake is assuming AGI is decades away and therefore ignoring the governance question. OpenAI, DeepMind, Anthropic, and several Asian frontier labs are explicitly racing toward AGI-level capability. Regulatory frameworks (EU AI Act, Singapore's AI governance, China's AI regulations) are beginning to address what this means for safety, liability, and deployment.

The prudent enterprise position: build for the ANI capabilities available today, design AI governance infrastructure that can scale to more capable systems, and monitor the AGI frontier for transition signals — particularly the emergence of agents that can learn from experience without fine-tuning, which would represent a qualitative shift from today's systems.