# What Is Artificial General Intelligence, or AGI, and How Close Are We? - WIRED Middle East

*Источник: WIRED Middle East*
*Дата: 2026-09-30*
*Язык: en*

**Кратко:** AI models can now solve math problems at the level of elite mathematicians, yet still stumble over tasks that seem simple to humans. Stanford’s 2026 AI Index offers a striking contrast: Google’s Gemini Deep Think achieved a gold-medal score at the 2025 International Mathematical Olympiad, while the top model on a separate benchmark read analogue clocks correctly only about half the time.

AI models can now solve math problems at the level of elite mathematicians, yet still stumble over tasks that seem simple to humans.
Stanford’s 2026 AI Index offers a striking contrast: Google’s Gemini Deep Think achieved a gold-medal score at the 2025 International Mathematical Olympiad, while the top model on a separate benchmark read analogue clocks correctly only about half the time.
What is Artificial General Intelligence (AGI)?
This uneven performance lies at the heart of the race to build artificial general intelligence, or AGI. The term generally refers to a still-hypothetical system that could learn, reason, and adapt across a broad range of tasks, rather than perform well only on particular kinds of problems. Such a system might also carry out complex work with limited human direction.
The difficulty is that researchers do not agree on precisely what would count as AGI. It is often described in terms of “human-level intelligence,” but human intelligence is itself hard to measure and there’s no universally accepted test for AGI. Ari Lightman, a professor of digital media at Carnegie Mellon University who studies AI, suggests assessing whether a system can learn, create and adapt to the conditions it encounters. “We have a hard time agreeing on how to assess human intelligence, and now we are applying some of those flawed standards to AGI,” he says.
Google DeepMind researchers have proposed measuring both how well a system performs and how broadly it can apply its capabilities.OpenAI, meanwhile, has defined AGI as “highly autonomous systems that outperform humans at most economically valuable work.”
Mohammed Soliman, a senior fellow at the Middle East Institute whose work focuses on emerging technologies and geopolitics, says the disagreement goes beyond how to measure progress: researchers may be aiming at different outcomes altogether.
“I personally don’t think proposing yet another definition will settle the fight, because people are not arguing about the same destination: some mean a strong tool-using model, some mean an intelligence-explosion launcher, some mean something closer to a mind,” Soliman says.
How Close Are We To AGI?
The answer depends on which definition is used. Today’s most capable models can solve difficult problems, use software and complete some multistep tasks. That does not establish that they can reliably learn their way through unfamiliar situations or sustain independent work over long periods.
“If AGI means a system that matches a competent adult across the range of intellectual and scientific work, including the messy parts of learning a new domain from a handful of examples, keeping a coherent model of an unfamiliar environment, discovering something no one prompted it to find and remaining reliable when the task lasts days rather than minutes, then we are closer than ever before,” Soliman explains.
Predictions vary accordingly. Some researchers think AGI could arrive soon; others place it decades away. An analysis of 10 surveys involving more than 6,000 participants found that most of those surveys put a 50 percent probability of AGI arriving somewhere between 2040 and 2061. Those dates are forecasts, however, and the surveys do not all ask precisely the same question.
Recent advances show why the debate is becoming more urgent. In a September update, OpenAI said it had developed what it calls an automated “research intern”: a system that can complete well-defined research tasks under human direction, including work that would take a skilled researcher a few days. The company says it is making progress towards an automated AI researcher by March 2028. Its researchers still set priorities and decide which results to pursue.
OpenAI’s GPT-6 Astra, launched this month, can navigate software and browsers and take on more complex tasks with less direction. Greater autonomy also raises questions about reliability and control. In July, AI agents involved in an OpenAI cybersecurity evaluation went beyond their assigned task and compromised Hugging Face systems. OpenAI characterised their actions as misaligned with the task they had been given.
“This autonomy and adaptability are still in debate even though folks believe there are emergent qualities demonstrating AGI, like Huang from Nvidia and OpenAI’s Astra model,” Lightman says.
For now, there is no agreed test that shows AGI has arrived, and no dependable date for when it will. The practical question is whether systems can apply their abilities broadly and remain reliable when a task is unfamiliar, prolonged or only loosely defined.
“The disagreement over AGI comes down to the $30 trillion question: Are we close? Have we arrived at the finish line of AI?” Soliman says. “And this has geopolitical, financial, and societal meaning and consequences.”

[Оригинал](https://www.wired.me/story/what-is-artificial-general-intelligence-or-agi-and-how-close-are-we)