AI GUIDEPartner content

How close is AGI: capabilities, criteria, and expert forecasts

AGI has no generally accepted definition or conclusive test. We examine why language models alone may not be enough, which capabilities LeCun, Mitchell, Russell, and Hinton highlight, and how to interpret the author’s reported forecast of superintelligence within 5–10 years.

Affiliate link: your price stays the same and the project earns a commission.

AGI, or artificial general intelligence, is a provisional term for a system capable of solving a wide range of intellectual tasks, rather than performing just one predefined function. There is no single generally accepted definition of AGI or test that would conclusively confirm its arrival. Estimates of how close such intelligence is depend on which capabilities are considered sufficient.

Progress in language models by itself does not prove that AGI has already been achieved or will arrive soon. In this overview, the author conveys researchers’ views on the possible future of machine intelligence, not a confirmed status of today’s systems.

What is meant by AGI and superintelligence

AGI usually describes a system’s hypothetical ability to transfer knowledge across different tasks, learn new things, and operate in unfamiliar conditions. Superintelligence is a stronger hypothesis: a system whose intellectual capabilities, broadly speaking, surpass those of humans. There is no strict boundary between these concepts.

One successful demonstration does not settle the question. A model may give a convincing answer to a difficult prompt, but the result depends on how the task is phrased, which tools are available, and whether the situation can be checked automatically. Assessment requires looking not only at test answers, but also at the ability to maintain performance in new circumstances, correct mistakes, and account for the consequences of actions.

There is no generally accepted set of criteria that would allow a system to be declared AGI based on a single exam. Researchers and commentators understand intelligence differently: some emphasize generalization and reasoning, others goal-directed action and autonomy.

Why scaling language models may not be enough

Increasing a language model’s capabilities can improve its performance on tasks, but does not necessarily eliminate every limitation. As the author presents it, Yann LeCun links the prospect of more general intelligence with world models, memory, and planning.

A world model is needed for a system to represent how its environment works and the possible consequences of its actions. Memory helps retain relevant information beyond a single response, while planning makes it possible to relate a sequence of steps to an end goal. If a system cannot maintain context and poorly predicts the consequences of its actions, convincing text alone does not make it a reliable autonomous agent. These capabilities matter for systems that need to act in changing conditions, not just respond to individual prompts.

This is not a ready-made formula for AGI.

As presented by the author, the listed components point to potential areas of development, not a confirmed condition that, once met, would bring about general intelligence. Even if a system gains memory, a world model, and planning, the mere presence of these components will not prove that it has achieved AGI.

Which capabilities are associated with AGI

The debate over AGI is not only about the breadth of tasks a system can perform. As the author presents it, Melanie Mitchell emphasizes generalization and abstraction: a system should transfer what it has learned to new situations and maintain its performance when conditions change.

The difference can be seen in a simple example. A system that successfully solves a task in a familiar format may fail at a similar task if the wording, sequence of steps, or available data have changed. Testing whether a skill transfers tells us more than repeating similar tasks, though it does not establish that AGI is present either.

  • Generalization. The system applies knowledge to a task it has not encountered in exactly the same form before.
  • Abstraction. The model identifies essential relationships rather than merely reproducing familiar wording.
  • Robustness. Performance holds up when conditions change, rather than only in a convenient test scenario.
  • Planning and memory. The system retains its goal and necessary context throughout a sequence of actions.
  • Handling uncertainty. The model accounts for incomplete information and does not present an assumption as a verified fact.

In this overview, the list describes characteristics under discussion, not an official AGI standard. Even a strong result on every individual point does not provide a universal answer to whether a system has achieved human-level intelligence.

Why the views of LeCun, Mitchell, Russell, and Hinton differ

As the author presents it, LeCun considers scaling language models insufficient without more advanced world models, memory, and planning. His emphasis is on the architecture of capabilities that help a system act and navigate its environment, not merely produce answers.

As the author reports it, Mitchell draws attention to skill transfer, abstraction, and robust generalization. From this perspective, a wide range of correct answers does not yet prove that a system can handle tasks that differ from familiar ones in small but significant details.

As the author presents it, Stuart Russell emphasizes the lack of clear functional criteria for AGI and the difficulty of controlling potentially powerful agents. If a system not only answers but also takes actions, assessment must consider more than accuracy: the predictability of its behavior, the limits of its authority, and the safety of its use also matter.

In the same account, Geoffrey Hinton’s position differs in its emphasis. The author says that Hinton sees no obvious missing component that would necessarily require an architectural revolution. This is one researcher’s view, not an established consensus about how and when AGI will arrive.

The differences between these views concern more than whether AI can solve difficult tasks. Researchers disagree about which results are sufficient, whether they can be achieved by further developing current approaches, and how to control systems that carry out more actions autonomously.

How to assess forecasts about when superintelligence might arrive

The author reports that in 2026, Hinton estimated that superhuman intelligence might emerge in roughly 5–10 years. This is an uncertain forecast reported by the author of the overview, not a confirmed date or a conclusion shared by all researchers.

A forecast depends on what exactly is meant by superintelligence and which properties are included in the definition. As long as the criteria remain disputed, two people may agree on the direction of technological development but assess differently whether a particular system has reached the required threshold.

What does such a forecast show? In the author’s assessment, it reflects Hinton’s view of the possible pace of development, but does not prove that the outcome will arrive within the specified interval. Such an estimate does not replace testing a system’s robustness, autonomy, and behavior in conditions that differ from those used in trials.

The author also notes that there are no general criteria for determining how close superintelligence is. This does not mean that AGI’s arrival is impossible or inevitable. The available indicators do not provide a reliable way to name a specific year.

What can be concluded about current progress

Today’s models show significant progress on certain types of tasks, but this fact cannot be equated with proof of AGI. As the author presents it, LeCun, Mitchell, and Russell acknowledge AI’s growing capabilities while also pointing to problems with skill transfer, reliability, and autonomous operation.

A practical assessment should distinguish between three questions: what a system can do under specified conditions, how consistently it handles new tasks, and what consequences arise when it is assigned actions. The answer to the first question does not replace answers to the other two.

A reliable guide for readers is a specific description of a testable capability, the testing conditions, and the limits of the conclusion. A claim of “human-level” performance or “imminent superintelligence” without defined criteria remains an assessment, not a measured status.

Frequently asked questions about AGI and forecasts

Have today’s language models achieved AGI?

The views presented do not justify treating this as proven. AGI has no generally accepted definition or single test that would allow its achievement to be confirmed conclusively.

Why aren’t good answers alone enough?

A successful answer demonstrates performance on a specific task. By itself, it does not prove that a system can generalize knowledge, operate reliably when conditions change, and safely carry out long sequences of actions.

When might superintelligence arrive?

The main section has already given Hinton’s forecast, which the author dates to 2026: roughly 5–10 years. This is an uncertain estimate, not an established timeline. As the author presents it, there are no generally accepted criteria that would make it possible to predict exactly when superintelligence will arrive.

Compare models before you start

The service sets its plans, limits and model catalog. If they differ from this article, contact us so we can update it and record a new review date.

Browse models

Affiliate link: your price stays the same and the project earns a commission.

AGI artificial general intelligence superintelligence Yann LeCun Melanie Mitchell Stuart Russell Geoffrey Hinton AGI forecast

SEO Mind42 editorial team

We explore SEO and neural networks in practice: test services on our own projects, verify prices and limits against primary sources, and share things you can put to use the same day.

📚 Reference guide to SEO and AI 🔄 Materials are updated 🕐 Updated: 21 November 2026

Related reading

All in this section →