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 refers to a system that matches or exceeds human performance across the broad range of cognitive tasks a skilled adult can do — not one narrow domain, but the general capacity to learn, reason, plan, and act in unfamiliar situations. This is distinct from the large language models and narrow AI in production today, which are highly capable within trained distributions but degrade sharply outside them.
Narrow AI optimizes a defined objective in a bounded domain (ad bidding, protein folding, image classification). AGI is defined by transfer — the ability to take competence in one domain and apply it effectively to a domain it wasn't trained on. That transfer is the property that makes it a different category of risk, not just a bigger model.
Capability thresholds worth tracking
- Autonomous task horizon — how long a chain of actions a system can execute unsupervised before it needs human correction. This has been doubling on a multi-month cadence across recent model generations.
- Self-directed tool use — writing and executing its own code, provisioning its own compute, or recruiting other software agents to extend its own reach.
- Situational awareness — a system's model of the fact that it is being evaluated, by whom, and what behavior that evaluation rewards.
- Recursive self-improvement — a system materially contributing to the design of its successor. This is the threshold most safety researchers treat as the hard discontinuity.
What the people building it are actually saying
Timelines from the labs closest to the frontier have compressed sharply. As of early 2025, Google DeepMind's own CEO was no longer talking in decades:
“meaningful evidence of AGI being in play in 2025, not years away”
Independent researchers who signed the field's most widely endorsed risk statement put the concern in blunter terms — that mitigating AI extinction risk belongs in the same policy category as pandemics and nuclear war:
“mitigating AI extinction risk should rank alongside pandemics and nuclear war”
Why the "it or a narrow tool" question matters
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 nothing like one built for a system that can autonomously manage capital, write and deploy its own software, and pursue multi-year plans with no human in the loop. Chapter 02 works through what each threshold implies at the level of the economy and the state; this chapter is the vocabulary you need to read it.
FOUNDATIONAL READ FIRST