AGI, narrow AI and superintelligence: the capability ladder What separates narrow AI, broadly capable systems, AGI, and superintelligence—and why no test settles the boundaries, per DeepMind's breadth-depth framework.
Source-checked 5 min
01 Definitions
What AGI is, and what it isn't
“AGI” gets used loosely. Before any risk discussion is useful, it needs a working definition and a way to tell how close a given system is to meeting it.
Working definition
Artificial General Intelligence has no single accepted test. In this manual it means a system combining broad cognitive competence, human-comparable or better performance, reliable transfer to unfamiliar tasks, and enough autonomy to complete consequential work. Each part matters: present general-purpose models are broad in some respects and remarkably capable on some evaluations, while remaining unreliable and uneven in ways that make a binary “AGI or not” label misleading.
The distinction that matters
Generality, performance and autonomy are separate dimensions. A system can converse across many subjects yet remain unreliable at long tasks, or exceed experts in one field without transferring that ability elsewhere. This manual therefore tracks a portfolio of measurements rather than treating transfer as a magic dividing line.
Capability thresholds worth tracking
These four signals help distinguish scenarios, but none is a complete AGI test. Some are measured directly; others are inferred through evaluations whose validity and coverage remain contested.
| Threshold | What it measures | Why it matters |
|---|---|---|
| Autonomous task horizon | The human completion time of benchmark tasks an agent can finish at a stated reliability level. | Tracks rapid progress on long software, ML and cyber tasks, but does not measure all real work or literal elapsed autonomy. |
| Self-directed tool use | Writing and running its own code, provisioning its own compute, recruiting other software agents. | Turns a model into an actor with reach beyond its own context window. |
| Situational awareness | A system's model of the fact that it is being evaluated, by whom, and what that evaluation rewards. | Undermines evaluation itself: a system that knows it is being tested can behave differently under test. |
| Recursive self-improvement | A system materially contributing to the design of its successor. | A proposed acceleration mechanism, not an observed or universally accepted hard threshold. |
What some researchers and lab leaders say
Forecasts vary and are not measurements. In March 2025, Google DeepMind's chief executive gave this personal estimate:
Demis Hassabis forecast in March 2025 that AGI could start to emerge within five to ten years.
CNBC, Reported interview
A separate 2023 statement signed by researchers, executives, and other public figures put catastrophic risk in policy terms. The statement records its signatories' position; it does not estimate probability:
Signatories to the Statement on AI Risk hold that mitigating the risk of extinction from AI should be a global priority alongside pandemics and nuclear war.
safe.ai, Signed statement
Why the definition decides the plan
Every protocol later in this manual — financial, digital, physical — depends on which of these thresholds has actually been crossed. A household plan built for “smarter chatbots” looks different from one built for a system that can autonomously manage capital, write and deploy software, and pursue long plans with little human review. Chapter 02 works through what each threshold could imply at the level of the economy and the state; this chapter is the vocabulary you need to read it.
Foundational Read first
In this chapter
AGI timeline forecasts: what the labs and surveys say A dated guide to AGI predictions, forecasting surveys, and skeptics—and why none should be mistaken for a countdown clock, however confident the source.
Source-checked 5 min
The Turing test and what replaced it Why humanlike conversation cannot establish AGI, and how modern evaluations measure capability, reliability, and autonomy across a benchmark portfolio.
Source-checked 5 min
How the frontier AI labs differ Compare the ownership, governance, release strategies and published safety frameworks of major frontier AI laboratories using dated primary sources.
Source-checked 5 min
Compute and scaling laws, explained Understand what AI scaling laws predict, where their forecasts break down, and why training compute remains an imperfect but useful governance signal.
Source-checked 5 min
Emergent capabilities: what appears, what only looks abrupt Why untrained-for abilities can arise in large models, how benchmark choices create apparent jumps, and what emergence really means for AI safety work now.
Source-checked 5 min
AI and AGI glossary: 200+ terms explained in plain English Understand more than 200 AI, AGI and AI-safety terms in plain English, from agentic systems and alignment to transformers and zero-shot learning.
Working draft 22 min
How to weigh claims about AGI An evidence ladder for separating demonstrations, evaluations, deployment data, forecasts, and theoretical risk arguments, plus a six-field evidence card.
Source-checked 7 min
Current AI capability dashboard: September 2026 A fixed, multidimensional snapshot of frontier AI capability — what Stanford's AI Index calls 'jagged intelligence' — and why benchmarks aren't proof of AGI.
Source-checked 6 min
AI-accelerated AI research and feedback loops What AI already contributes to coding and model research, and what would have to be true for a self-reinforcing capability explosion — not autocomplete.
Source-checked 6 min
AGI winter and false dawns The counter-scenario: why AGI might not arrive on schedule, what an AI winter would mean for safety and talent, and why a pause is not the same as safety.
Source-checked 6 min
AI incidents: a chronological record, 2016-2026 A sourced, chronological record of documented AI incidents and near-misses from 2016 to 2026 — real harms and disclosures, not hypothetical scenarios.
Source-checked 16 min
Frequently asked questions about AGI Direct answers to the questions people actually ask about AGI — what it is, when it might arrive, and how dangerous it is — each linking to the fuller page.
Source-checked 5 min
Public opinion and trust in AI Pew, Gallup, Edelman, and Ipsos data on how public trust in AI has shifted since 2021 — including the widening gap between Chinese and Western respondents.
Source-checked 6 min
The skeptic case against near-term AGI The strongest technical arguments that AGI is not close, from named researchers with real credentials — represented fairly, not as a caricature to knock down.
Source-checked 7 min
Who's who in AGI: labs, safety orgs, and researchers A dated roster of frontier labs, independent safety organizations, and named researchers shaping the AGI debate, sourced and flagged for what's still moving.
Source-checked 11 min