What AGI Is, Where Current AI Stands, and Why Nobody Can Date Its Arrival

October 5, 2026 · 4 min read

Artificial general intelligence — AGI — is a hypothetical machine that matches or exceeds human performance across virtually all cognitive tasks, not just one narrow job. Current AI systems are far more capable than most people assume, but the distance between them and AGI resists a single number: researchers have proposed describing progress along dimensions such as capability depth, breadth, and deployment autonomy, and even designing benchmarks for those dimensions remains an open problem. As for arrival: no one can predict it reliably, because the field has no agreed test for what "arrival" would even mean.

What AGI actually is

The working definition: a system that can generalize knowledge, transfer skills between domains, and solve novel problems without task-specific reprogramming. Researchers generally expect it to reason under uncertainty, plan, learn, hold common-sense knowledge, communicate in natural language, and integrate all of that toward whatever goal it is given. That expectation has never been pinned down into a single testable specification, which is why the next problem matters.

Each major organization defines the destination differently. OpenAI frames AGI around performing most economically valued work; Google DeepMind has published a five-level performance taxonomy running from "emerging" to "superhuman." When the finish line is drawn differently by each participant, crossing it becomes a matter of interpretation.

A paper by Meredith Morris and colleagues, "Levels of AGI for Operationalizing Progress on the Path to AGI" (presented at ICML 2024), tries to bring order to this. It proposes classifying systems along three axes: depth (how well a system performs), breadth (how many domains it covers), and deployment autonomy (how much control remains with humans). The authors are explicit that this is a proposed framework for comparing progress and assessing risk — not a universal definition, not a certification, and not a calendar forecast. They also flag the central measurement problem: designing benchmarks that could actually quantify AGI-level behavior is itself an open challenge.

How it differs from current AI

Today's frontier models reason across domains, hold long contexts, write working code, analyze documents, and conduct research. When DeepMind published its levels taxonomy in 2023, it placed large language models at the bottom rung — "emerging," comparable to unskilled humans on non-physical tasks, below "competent," which means outperforming half of skilled adults across a wide range of tasks. That 2023 placement is a snapshot, not a verdict: whether it still describes today's systems is precisely the judgment the taxonomy leaves open.

Where current systems sit on the Morris dimensions is where the honest answer gets thin. Depth and breadth can be illustrated with individual benchmarks, but the authors note that quantifying AGI-level capability across those axes is an open benchmarking problem, so any single score understates the picture. Deployment autonomy is different in kind: in the Morris framework it is an axis of deployment rather than a fixed property of a model. The same system can sit near the "tool" end in one configuration and closer to "agent" in another, depending on how much control operators retain. Where a given product lands is a property of that deployment, not a label that transfers across all current systems.

There is genuine dispute about how much of this gap remains. Blaise Agüera y Arcas and Peter Norvig argued in 2023 that frontier models already exhibit significant general intelligence, and whether that argument holds depends on which definition you apply — the problem above, still open.

Can anyone predict its arrival

No — and the reasons run deeper than optimism or pessimism.

First, there is no agreed threshold to predict against. A date attached to a moving target is not a forecast; it is an assumption wearing a number. Second, the credible positions do not even agree on what they are predicting: Agüera y Arcas and Norvig have argued that significant general intelligence is already here, while the organizations drawing the finish line place it in different spots, so a forecast tied to one definition says nothing about another. Third, the disagreement reflects unresolved science — whether scaling alone suffices, whether data and compute walls will bind, whether AI-accelerated AI research changes the pace — not ordinary estimation error. When the underlying mechanisms are unsettled, a precise date is not a measurement; it is a bet dressed as one.

What replaces a date is a set of observable checkpoints. If a system starts passing expert-level tests across multiple domains without retraining, if it begins contributing to its own development in verifiable ways, or if its autonomy in production deployments moves from tool toward collaborator, those are events you can watch for. A vendor claiming proximity to AGI is worth asking three questions: which definition are you using, what independent evidence supports it, and where does the system sit on depth, breadth, and autonomy? Claims that pass only under the claimant's own narrow definition tell you little.

That is also the posture of this article's publisher. SentX, which develops the foundation model Victoria, describes AGI and consciousness on its public site as research aims, not achieved results. The destination is real as a direction of travel; the arrival date is not knowable.