What is Artificial General Intelligence (AGI)?

What is artificial general intelligence (AGI)?

Artificial general intelligence (AGI) is a hypothetical (as in, its not here yet) type of AI that can understand, learn and perform any intellectual task a human can, across domains, without being retrained for each one. As of October 2026, no AI system has been confirmed as AGI, and the companies building it don’t agree on how to define it, which causes a bit of public conflict.

That generic definition is the part almost everyone accepts. The fight starts the moment you try to measure it, as there is no real, systematic way to measure the ‘intelligence’ of an AI system.

Before ChatGPT and LLMs, AI was considered “narrow” in a specific sense: a machine learning model can write code, pass a bar exam or win a math olympiad medal, yet still fail at things a child handles easily, like remembering what you told it last week or understanding why a glass falls off a table.

AGI would close that gap: one system that generalizes, adapts to new problems and keeps learning, and many think that with LLMs we’re heading in this direction, as these systems actually generalize and learn ’emergent behaviours’ that go further than their training goal, which is predicting the next word in a sentence.

Over the last few years, the people building AI have defined AGI in at least four very different ways:

Lens	AGI means…	Who leans on it
Cognitive	Matches a capable adult across reasoning, memory, perception and learning	Researchers, academic definitions
Economic	Outperforms humans at most economically valuable work	OpenAI’s charter
Financial / contractual	Generates a profit threshold written into a contract	Microsoft and OpenAI (2023)
Macroeconomic	Visibly accelerates world GDP growth	Satya Nadella

This article walks through how that conversation evolved, from a $100 billion contract clause to CEOs calling the term itself “sloppy”.

AGI vs. narrow AI vs. superintelligence

AGI sits between the AI we use today and the superintelligence some labs now say they’re aiming for.

Type	What it can do	Exists today?
Narrow AI	Excels at specific tasks (chat, code, image recognition) but can’t reliably transfer skills or keep learning	Yes
Artificial general intelligence (AGI)	Matches human ability across most intellectual tasks and adapts to new ones	No (contested)
Artificial superintelligence (ASI)	Far exceeds the best humans in virtually every domain	No

AGI timeline: key statements and milestones

Date	Who	What happened
Sep 2026	Altman, Amodei	Bloomberg reports leaders have abandoned a shared AGI definition
Mar 2026	Yann LeCun	AMI Labs raises $1.03B to build world models instead of LLMs
Jan 2026	Amodei, Hassabis	Davos: 1–2 years vs. 5–10 years to human-level AI
Oct 2025	Microsoft, OpenAI	New deal: an independent expert panel must verify any AGI claim
Oct 2025	Hendrycks, Bengio et al.	“A Definition of AGI” scores GPT-4 at 27%, GPT-5 at 57%
Aug 2025	Sam Altman	Calls AGI “not a super useful term”
Feb 2025	Satya Nadella	Real AGI benchmark = world GDP growing ~10% a year
Jan 2025	Sam Altman	“Confident we know how to build AGI”; predicts it during Trump’s term
Dec 2024	The Information	Reveals the $100B profit definition in the Microsoft–OpenAI contract
2018	OpenAI	Charter defines AGI as outperforming humans at most economically valuable work

Where the term came from

The phrase “artificial general intelligence” was popularized in the early 2000s by researchers like Ben Goertzel and Shane Legg (later a DeepMind co-founder). They wanted a label that separated the original dream of AI, a thinking machine, from the narrow systems the field was actually shipping. If you want the longer history, our list of the best artificial intelligence books covers it.

For two decades AGI stayed mostly a research ambition. That changed when OpenAI, founded in 2015, made it the company’s explicit mission. Its charter defined AGI as highly autonomous systems that outperform humans at most economically valuable work. Notice the shift: intelligence is defined by work, not by cognition. That economic framing set the stage for everything that followed.

The $100 billion definition

In December 2024, The Information reported that Microsoft and OpenAI had signed an agreement in 2023 stating that OpenAI only reaches AGI when it builds AI systems able to generate at least $100 billion in profits (TechCrunch).

Why would a scientific milestone live in a contract? Because money depended on it. Microsoft had invested roughly $13 billion, and the original deal said OpenAI could stop sharing new technology with Microsoft once AGI was achieved.

A vague definition meant OpenAI could, in theory, “declare AGI” early and cut its biggest partner off. A profit threshold made the trigger concrete (Sherwood News).

At the time, OpenAI was losing billions a year and told investors it wouldn’t be profitable until 2029. By this definition, AGI was many years away, which suited Microsoft just fine. Critics pointed out the obvious problem: a profit figure says nothing about whether a system can reason, learn or understand.

Companies are investing their money like never before in a technology, and expecting inmense returns out of it, which can only be obtained with AGI.

What is Artificial General Intelligence (AGI)?

2025: confidence, then a retreat from the word

Sam Altman goes all in. In a January 2025 blog post, Altman wrote that OpenAI was confident it knew how to build AGI as traditionally understood, and that the company was turning its attention to superintelligence.

In a Bloomberg interview the same month he predicted AGI would probably arrive during Donald Trump’s presidential term, while already admitting the term had become sloppy (TIME summary via LongtermWiki).

Satya Nadella pushes back, from the other side of the deal. In February 2025, on Dwarkesh Patel’s podcast, the Microsoft CEO dismissed companies self-declaring AGI milestones as “nonsensical benchmark hacking.” His alternative test: the world economy growing at around 10% a year, compared with roughly 2% in developed economies today (Benzinga). For Nadella, if AI is truly general, it should show up in GDP the way the Industrial Revolution did.

Then the vibe shifted. By August 2025, around the GPT-5 launch, Altman told CNBC that AGI was “not a super useful term” because everyone defines it differently (Fortune). He began talking about levels of capability rather than a yes/no finish line. Critics like Max Tegmark read this less as humility and more as a way to avoid regulatory scrutiny while still racing toward more powerful systems.

The skeptic: Yann LeCun

Yann LeCun, one of the Godfathers of Deep Learning and most popular minds on AI, doesn’t say human-level AI is impossible. He says today’s path won’t get us there. The Turing Award winner and former Chief AI Scientist at Meta has argued for years that large language models are a dead end for real intelligence, because they learn from text rather than from the physical world.

He also dislikes the term AGI itself, arguing that even human intelligence is highly specialized, and prefers “human-level AI” or “advanced machine intelligence.”

His case: an LLM predicts the next token. It has no persistent memory, no model of how the world works, and no ability to plan by predicting the consequences of its actions. (If you want to understand how generative AI and LLMs work under the hood, start there before judging the debate)

At a 2026 lecture at Brown, he described LLMs as helpless in the physical world and put up a slide telling students interested in human-level AI not to work on LLMs. He defines intelligence as the ability to solve new problems you’ve never been trained on, which actually makes sense: intelligence as the ability to Learn quick.

He gives many examples, like a human person on average needing 20 hours to learn how to drive, vs the more than 4000 hours of a Tesla car needed to drive autonomously.

LeCun left Meta in late 2025 and co-founded AMI Labs, which in March 2026 raised $1.03 billion at a $3.5 billion pre-money valuation to build “world models” trained on video and sensor data, using his JEPA architecture (VKTR). The company is headquartered in Paris and positioned as a European alternative to US and Chinese labs. It has no product yet.

The optimists: Amodei and Hassabis

Dario Amodei (Anthropic) mostly avoids the word AGI and talks about “powerful AI” instead: systems as capable as Nobel Prize winners across most fields. In Anthropic’s March 2025 submission to the White House, the company said it expected such systems in late 2026 or early 2027.

Demis Hassabis (Google DeepMind) sits in the middle. He talks about a roughly 50% chance by 2030 and warns about “jagged intelligence”: models that win math medals but fail tasks a 12-year-old can do.

The two shared a stage at Davos in January 2026. Amodei held to one to two years for systems that outperform humans across most cognitive work, driven by AI automating software engineering and then AI research itself. Hassabis stuck to five to ten years, arguing that verifiable domains like coding and math will fall first, while science and the ability to ask genuinely new questions will take longer (Online Summit). Amodei said he’d actually prefer Hassabis’s slower timeline.

2025–2026: from contract clause to expert panel

The $100 billion clause got replaced. In October 2025, alongside OpenAI’s restructuring into a for-profit company, Microsoft and OpenAI signed a new agreement. If OpenAI declares AGI, an independent expert panel now has to verify it. Microsoft’s rights to OpenAI models extend through 2032, including models after AGI, and Microsoft is free to pursue AGI on its own (Microsoft). Who sits on that panel, and what test they’ll apply, hasn’t been made public.

Researchers tried to make it measurable. Also in October 2025, a group including Dan Hendrycks, Yoshua Bengio and Gary Marcus published “A Definition of AGI”: an AI that matches or exceeds the cognitive versatility and proficiency of a well-educated adult. They score models across ten cognitive domains drawn from human psychometrics. GPT-4 scored 27% and GPT-5 57%, with 0% on long-term memory storage (arXiv). Fast progress, big gap. Measuring models honestly is hard even for ordinary ML, as anyone who has seen cross-validation lie to them knows.

Where it stands now (October 2026). In September 2026, Bloomberg reported that the industry’s leaders have drifted away from any shared definition: Altman calls AGI a sloppy term and talks about superintelligence in “a few thousand days,” while Amodei barely uses the word (Startup Fortune). Nobody has declared AGI, and nobody agrees on a test that would settle it.

What this means if you’re learning ML

When someone says “AGI is two years away” or “AGI will never happen,” first ask which definition they’re using. A profit threshold, a GDP growth rate, a psychometric score and “beats Nobel laureates” are four different finish lines, and the people choosing them often have money riding on the answer.

The more useful question for a learner is what today’s systems still can’t do: persistent memory, learning continuously after training, reasoning about the physical world, and asking new questions instead of answering known ones. Those gaps are exactly where the research, and the next generation of ML jobs, will be.

Want to build the foundations to work on them? Start with our free 90-Day Machine Learning Roadmap, pick from our curated machine learning courses, or go deeper with the ML Starter Kit: 172 pages from your first line of Python to a deployed model.

Frequently asked questions about AGI

What is artificial general intelligence in simple terms? AGI is AI that can do almost any thinking task a person can, and learn new ones, instead of being good at just one thing.

Has AGI been achieved? No. As of October 2026, no lab has declared AGI, and under the 2025 Microsoft–OpenAI agreement any such claim would need to be verified by an independent expert panel.

Is ChatGPT AGI? No. Models like GPT-5 are strong at language, coding and math, but researchers scored GPT-5 at 57% on a human-cognition AGI scale, with 0% on long-term memory.

When will AGI arrive? Predictions range from 2026–2027 (Dario Amodei) to around 2030 (Demis Hassabis) to “not with current LLMs” (Yann LeCun).

What was the $100 billion AGI definition? A 2023 contract clause stating OpenAI would reach AGI only when its systems could generate about $100 billion in profits. It was replaced in October 2025.

What is the difference between AGI and superintelligence? AGI matches human-level intelligence across domains; superintelligence would far exceed the best humans in nearly every field.

If you want to build AGI, and you don’t know where to start from, check out our Free 90-Day Machine Learning Roadmap organizes those concepts into a step-by-step path—from Python and ML fundamentals to model evaluation, projects and deployment.

→ Follow the Free 90-Day Machine Learning Roadmap

Sources

Subscribe to our awesome newsletter to get the best content on your journey to learn Machine Learning, including some exclusive free goodies!

HOW IS MACHINE LEARNING

What Is Artificial General Intelligence (AGI)? Definitions, Predictions and the $100 Billion Debate