Artificial General Intelligence

Defining AGI

Artificial General Intelligence refers to AI systems that can perform any intellectual task that a human can, with comparable flexibility and efficiency. Unlike narrow AI, which excels at specific tasks like playing chess or recognizing images, AGI would transfer learning across domains, reason about novel situations, and adapt to tasks it wasn't explicitly trained for.

The term emphasizes generality: not just being smart at one thing, but being capable across the full range of human cognitive abilities.

The Narrowing Gap

Frontier AI systems have closed many of the gaps once thought to separate them from AGI:

Common sense reasoning: Benchmarks designed to be easy for humans and hard for machines keep falling. ARC-AGI-2, built specifically to resist AI when it launched in 2025, went from single-digit frontier scores to near-perfect results in under two years.

Transfer learning: Modern LLMs are fundamentally transfer learners, they generalize across domains with zero-shot and few-shot prompting, performing tasks they were never explicitly trained for.

Causal understanding: Frontier models demonstrate growing capacity for causal reasoning, though deep causal understanding in novel physical domains remains a frontier challenge.

Open-ended learning: Retrieval and in-context learning let models absorb new information on the fly. True continual learning, updating skills permanently without forgetting old ones, is one of the last major gaps and one of the most actively researched.

Generality involves more than answering isolated questions. A system must also carry knowledge across tasks, sustain a plan, recover from mistakes, and recognize when it needs new information. Machine-checked mathematics offers a useful test of sustained reasoning because intermediate results can be verified. Physical-world competence and continual learning test different dimensions, so no single achievement establishes AGI.

Measuring AGI

There's no consensus on exactly when AI becomes "general." DeepMind proposed a framework with levels from narrow AI through superhuman AGI. Others suggest AGI arrives when an AI can do any economically valuable work, or when it can improve itself.

Some researchers argue we'll cross the threshold gradually, with AI becoming "generally capable" in more and more domains until the distinction becomes moot. Others expect a sharper transition when key insights unlock general reasoning.

Timelines Collapse

Forecasts have moved sharply earlier. In a survey of 2,778 AI researchers, the aggregate forecast for machines outperforming humans at every task arrived 13 years sooner than in the same survey one year earlier. The people building the systems are more aggressive still: Anthropic's Dario Amodei has said powerful AI could arrive as early as 2026 or 2027, and Google DeepMind's Demis Hassabis has pointed to around 2030.

Part of the remaining disagreement is definitional. Google DeepMind's Levels of AGI framework separates breadth and depth of capability from autonomy, which is why systems can already be superhuman in some domains while still short of general in others.

Why It Matters

AGI is the threshold where machines can do the work of discovery itself: research, engineering, and invention at the speed of software, including the work of improving AI. That makes it the ignition point for the intelligence explosion and the moment the Singularity stops being a forecast.

Autonomy is a separate design choice. A broadly capable system can still operate under human direction, and the gap between "almost AGI" and "definitely AGI" is the window for making sure it does.