AGI vs ASI: where superior intelligence fits

Breadth, performance and permission

Google DeepMind's Levels of AGI framework separates how well a system performs from how broadly it can apply its abilities, and discusses autonomy as a deployment consideration. That is a useful structure for reading the terminology, even when researchers disagree about precise thresholds. Read the framework abstract.

In ordinary discussion, AGI refers to general capability across many cognitive tasks, while ASI refers to a stronger claim of broadly superhuman capability. Neither label, on its own, tells a buyer which tools a product can access or which actions it is permitted to take.

Four terms worth keeping separate

Narrow AI is specialised around a limited task or domain. A system can be better than a person at a particular task without that result demonstrating broad general intelligence.

Artificial general intelligence is about breadth as well as performance. To assess an AGI claim, look for evidence across different tasks and unfamiliar situations, and read the definition the authors actually use.

Artificial superintelligence raises the performance claim further. Ask whose performance is being exceeded and whether the evidence covers the breadth implied by the claim.

Agentic AI uses tools to act in an environment. An assistant with permission to edit files has more operational reach than the same assistant answering questions in a chat window. That additional reach does not by itself establish additional intelligence.

A buying decision should name the job

Suppose a company wants an assistant to reconcile delivery records. It needs evidence about matching records, finding exceptions, explaining uncertainty and working within its permissions. Whether the vendor calls the product an agent, an AI employee or a step toward AGI does not answer those questions.

Write the acceptance criteria before watching the demo. Include incomplete records, duplicate entries and cases that require a person to decide. Otherwise, the buyer may end up evaluating presentation quality rather than the job they needed done.

Why the newspaper covers all of them

The Super Intelligence Times follows the connection between current systems and more ambitious research. We will describe a tool improvement as a tool improvement, and identify a forecast as a forecast. That leaves room to report meaningful progress without treating every release as the same milestone.

When comparing articles, keep the definition attached to the claim. If two sources use different thresholds for AGI, their timelines cannot be compared as if they were forecasting the same event. The next useful question is what evidence would change each author's assessment.