JOURNAL / INTELLIGENCE & EVALUATION

What is AGI?
Define the range. Test the transfer.

Artificial general intelligence is a term for broad, adaptable AI capability. Its meaning depends on what a system can do, how well it does it, what it can learn, and how independently it can act.

Pedestrians move through a city intersection beneath tall buildings.INTELLIGENCE / SCOPE & GENERALITY / SYMPLOKA
ONE TASK IS A SMALL WINDOW

Generality is about what transfers when the task changes.

A strong result in one familiar setting does not reveal how a system will adapt to an unfamiliar one.

01 / A TERM WITHOUT ONE FINISH LINE

AGI is a research question, not a universal certificate.

Artificial general intelligence (AGI) usually refers to AI that can handle a broad range of tasks and adapt beyond a narrow, pre-defined setting. Researchers do not use one universally adopted pass-or-fail test for the term. Definitions depend on what counts as a task, how much prior experience is allowed, and how much independent action the system can take.

A 2024 position paper at the International Conference on Machine Learning proposes a useful way to make claims more precise: describe capability by its breadth and depth, then consider autonomy separately when evaluating deployment and risk. It is a framework for discussion and measurement, not a fieldwide AGI certificate.

Read the ICML paper on levels of AGI ↗, and see how the question differs from the broader definition of AI ↗.

02 / THREE QUESTIONS, KEPT SEPARATE

“General,” “good,” and “autonomous” do not mean the same thing.

Breadth / generalityHow many kinds of tasks can the system handle, including tasks it was not specifically tuned or instructed for?
Depth / performanceHow well does it perform within each task family, and against which human, expert, or machine baseline?
AutonomyHow independently can a deployed system choose steps, use tools, and continue before a person reviews or approves its actions?

A system could perform exceptionally well in a narrow domain without being general. A broadly capable model could also require careful supervision before it takes consequential actions. Report each dimension explicitly instead of collapsing them into one label.

Explore what makes a system an agent ↗ and how alignment relates goals to behavior ↗.

03 / SKILL IS NOT THE SAME AS LEARNING

A high score can reflect experience with the test.

A system may be excellent at a task because it has seen similar examples, received extensive training, or been given tools and scaffolding tailored to that setting. Those results matter, but they do not by themselves tell us how efficiently it can learn a new skill or adapt when the conditions change.

In On the Measure of Intelligence, François Chollet proposes measuring intelligence as skill-acquisition efficiency across a defined scope of tasks, while accounting for prior knowledge, experience, and generalization difficulty. That is one research proposal, not a universally accepted definition. It highlights a central evaluation question: what can a system infer from limited experience on a task it has not already practiced?

Read the paper behind the Abstraction and Reasoning Corpus (ARC) ↗ and the ARC-AGI benchmark repository ↗.

04 / BENCHMARKS SHOW SLICES

Different tests illuminate different parts of general capability.

Humanity’s Last Exam / academic knowledgeThe 2026 Nature paper presents 2,500 expert-authored questions across more than 100 subjects. It measures difficult, closed-ended academic performance. The benchmark’s authors explicitly say that even high accuracy would not by itself demonstrate autonomous research or AGI.Read the Nature paper ↗Inspect the benchmark code ↗
ARC-AGI-3 / adaptation through actionIntroduced in 2026, this benchmark uses unfamiliar, interactive environments. An agent must explore, infer goals, learn how the environment behaves, and plan actions without explicit instructions. It probes a focused slice of adaptive, agentic capability.Read the ARC-AGI-3 paper ↗Explore its open-source toolkit ↗

Neither benchmark is a complete test of intelligence. A score answers questions about a task set, scoring rule, and test conditions. See how to read a benchmark without turning it into a verdict ↗.

05 / WHAT STRONG EVIDENCE NEEDS

Test transfer, learning, and failure—not only polished answers.

  1. Define the scope. Name the task families and the environments a claim covers.
  2. Make novelty meaningful. Separate tasks that are new to the system from examples likely seen during training or prompt design.
  3. Account for prior knowledge and tools. Report the data, instructions, tools, retries, compute, and human support available during a test.
  4. Measure more than task accuracy. Test adaptation, reliability, calibration, recovery from errors, and the resources used to reach a result.
  5. Separate capability from autonomy. Explain which actions the system can take, which require approval, and how it stops safely.
  6. Publish the limits. Show the cases where performance fails and what the benchmark does not measure.

A stronger claim is specific enough to reproduce and narrow enough to challenge. Progress becomes clearer when tests expose both what transfers and what breaks.

06 / WHAT THE LABEL DOES NOT TELL YOU

AGI does not automatically mean superhuman, conscious, or safe.

Generality describes the range of tasks. Superhuman performance describes how well a system does a task relative to a reference. Autonomy describes how independently it can act. These are different properties; a label should not be used as a shortcut for any of them.

Capability tests also do not settle whether a system has subjective experience. That is a separate question. Nor does broad capability guarantee that a system’s goals, actions, or failure modes are safe. Those require their own evidence and safeguards.

See Symploka’s research questions on evaluation and safety ↗ and the practical meaning of AI alignment ↗.

QUICK ANSWERS

Keep the claim attached to its evidence.

Is AGI already here?There is no single accepted pass criterion. Ask which definition, benchmark, and conditions a claim uses.
Is AGI the same as an AI agent?No. An agent is a way a system can select steps and act; agency alone does not establish broad capability.
Does AGI mean superintelligence?No. Breadth and performance are separate dimensions in research frameworks.
Does a benchmark prove AGI?No single benchmark measures every task, learning condition, risk, and form of autonomy the term might include.
RESEARCH & PRIMARY SOURCESMorris et al., Levels of AGI, ICML 2024 ↗Chollet, On the Measure of Intelligence (2019) ↗Phan et al., A benchmark of expert-level academic questions to assess AI capabilities, Nature (2026) ↗ARC Prize Foundation, ARC-AGI-3 (2026) ↗Humanity’s Last Exam evaluation code ↗ARC-AGI open-source toolkit ↗