# AGI Is Here. We’re Just Arguing About the Label.

> The AGI debate keeps moving the finish line, while imperfect systems may already be transforming the economy faster than the label can catch up.

- Author: Shahab Nabavi
- Date: 2026-09-08
- Category: Quiet Thoughts
- Canonical: https://shahabty.github.io/blog/agi-is-here-were-just-arguing-about-the-label/
- Tags: agi, ai, technology, economics, scaling

---

For years, we imagined AGI would arrive like a scientific event. A laboratory would make an announcement, a machine would demonstrate something unmistakably human, governments would react, markets would move, and there would be a clear moment when we could say that Artificial General Intelligence had arrived.
Instead, we got something much messier. AI systems gradually became able to write software, analyze images, explain scientific papers, translate languages, research companies, create content, operate computers, solve difficult problems, and perform an increasingly large range of intellectual tasks. None of these moments individually looked like AGI, so every time AI crossed another boundary, we simply moved the boundary.
That is why I think the AGI debate has become strange. We are waiting for a finish line that nobody has actually defined.
There is no universally accepted AGI test. OpenAI has historically framed AGI partly in economic terms, describing systems that outperform humans at most economically valuable work. DeepMind researchers have proposed thinking about AGI as a spectrum based on generality and capability. Other researchers focus much more on whether a system can encounter something genuinely unfamiliar and learn it efficiently.
All of these definitions sound reasonable, but they can produce completely different answers. A machine could automate huge parts of software development, accounting, law, marketing, and research while still being worse than a human at learning a strange new game from five examples. Is that AGI? Depending on whom you ask, the answer could be yes, no, or “almost.”
That is not a minor problem. If we cannot agree on what AGI means, it becomes very easy to keep announcing that we are getting closer to it.

## The Turing Test Already Lost

For decades, the Turing Test was the cultural benchmark for machine intelligence. Alan Turing's idea was elegant: instead of trying to define “thinking,” let a human communicate with a machine through text and see whether they can reliably distinguish it from another person.
That question made sense when computers struggled to produce a few believable sentences. It makes much less sense today.
Modern language models can already hold conversations that many people mistake for human interaction. Yet almost nobody looks at this and declares that AGI has been achieved. The reason is obvious: once machines became good at conversation, we realized conversation was not enough.
Now we want machines that can recognize when they are wrong, learn new concepts from little information, operate independently for long periods, adapt to unexpected situations, remember experiences, and make reliable decisions in environments they have never seen before.
In other words, AI passed enough of the old test that we decided the test was never really testing what we wanted.
The Turing Test is still interesting as a measure of human imitation, but as an AGI test it is basically obsolete. Measuring machine intelligence by how convincingly it behaves like a human may eventually look as strange as measuring an airplane by how convincingly it behaves like a bird.

## Maybe Today’s AI Is Not AGI. Who Cares?

This is the part of the AGI debate that I think receives far too little attention.
Suppose the skeptics are completely right. Suppose today's systems are not AGI, and suppose true AGI is still ten years away.
Does that actually protect us from disruption? Not really.
A calculator is not generally intelligent, but it transformed accounting. Google is not AGI, but it completely changed how humans access information. The internet is not intelligent at all, yet it reorganized media, commerce, communication, politics, and entire industries.
AI does not need to become equivalent to a human mind before it becomes economically transformative. If one engineer with AI can eventually perform the work that previously required three engineers, that matters. If an AI system can handle most routine legal research, financial analysis, customer support, documentation, and administrative work, that matters. If a scientist can test hypotheses dramatically faster because an AI system can search literature, write code, analyze data, and suggest experiments, that matters.
We may be making a category mistake by assuming that AGI is required for an AGI-like economic transformation.
Maybe we never actually need AGI.
A sufficiently capable collection of imperfect specialized and semi-general systems could reshape the economy long before anyone agrees that intelligence has become truly “general.”

## AGI Is Also an Investment Story

There is another reason the phrase AGI keeps appearing: AGI is an incredibly powerful financial narrative.
Consider the economics of the current AI race. Companies are spending extraordinary amounts of money on GPUs, data centers, power generation, networking, training, and research. Investors naturally want to know what kind of return could possibly justify this level of capital expenditure.
“We are building a much better productivity tool” is one answer.
“We are building technology capable of performing a substantial percentage of human intellectual work” is a very different answer.
The second story supports much bigger expectations about future revenue, market size, margins, and strategic importance. It attracts capital. It attracts talent. It encourages governments to treat your company as strategically important. It makes competitors afraid of being left behind.
This does not mean AGI claims are fake. The technical progress is very real. But it would also be naive to pretend that the people defining the future of AI operate outside financial incentives.
AGI today sits in an unusual position where science, marketing, geopolitics, and capital markets overlap. When an AI CEO says AGI may be close, that statement can simultaneously be a technical prediction, a recruiting message, an investor narrative, and a warning to governments.
That does not make it false. It makes it worth examining carefully.

## The Trick That Changed AI

The strange thing is that the current revolution did not begin because researchers finally figured out how human intelligence works.
They mostly figured out how to scale machine learning.
The 2017 paper *Attention Is All You Need* introduced the Transformer architecture, which eventually became the foundation of modern language models. Transformers proved remarkably well suited to training on huge datasets using enormous amounts of parallel computation.
Then came another critical realization. In 2020, research on neural scaling laws showed that model performance improved in relatively predictable ways as researchers increased model size, training data, and computation.
The simplified version of the discovery was almost embarrassingly straightforward: build a larger model, train it on more information, spend more compute, and it tends to get better.
DeepMind's Chinchilla research later refined this idea by showing that simply increasing the number of parameters was not enough. Models also needed enough data. A smaller model trained on substantially more data could outperform a much larger one that had been undertrained.
The industry's recipe therefore became more sophisticated, but the underlying idea remained similar: intelligence-like capabilities seemed to emerge surprisingly reliably when the right architectures were exposed to enough data and enough computation.
That is one of the strangest discoveries in modern science.
We did not fully understand intelligence and then engineer it.
We discovered that if we scaled certain learning systems hard enough, increasingly general capabilities began appearing.

## So When Does Scaling Stop?

This is where the future becomes less obvious.
The first era of modern AI was dominated by pretraining: bigger models, more data, more compute. But there are physical and economic limits. High-quality human-generated data is finite. Frontier training runs are expensive. Data centers require enormous amounts of electricity. Specialized chips are difficult to manufacture. At some point, spending ten times more money cannot continue producing ten times more useful intelligence.
The important point, however, is that scaling is not necessarily ending. What we choose to scale is changing.
Instead of only making the model larger, researchers can give the model more computation when it is solving a problem. They can let it reason for longer, generate multiple solutions, test its own code, search the web, consult databases, use external tools, verify its answers, maintain memory, and coordinate with other agents.
The future may therefore look less like one gigantic neural network that instantly knows everything and more like an intelligent system composed of models, tools, memory, search, planning, verification, and external computation.
In some ways, this resembles humans more than a standalone model does. Human intelligence does not exist in isolation either. We use books, computers, calculators, colleagues, institutions, and the accumulated knowledge of civilization.
Perhaps the mistake is judging the model alone when the real unit of intelligence will be the entire system built around it.

## What Is Actually Missing?

There are still strong technical reasons not to call today's systems AGI.
Reliability is probably the most important. Current models have what researchers sometimes describe as a jagged intelligence profile. They can solve remarkably difficult problems and then fail at something embarrassingly simple. They can produce a sophisticated argument built on a completely false premise. They can confidently state something that never happened.
That makes them useful, but difficult to fully trust.
Continual learning is another major gap. Humans constantly absorb new experiences, update beliefs, remember important events, and develop skills. Most current models do not naturally update themselves every time they encounter something new. We can attach memory systems, retrain them, or provide new context, but this still feels different from the continuous learning process of a human mind.
Learning efficiency is another unanswered problem. Humans can often infer rules from only a handful of examples. Machine learning has historically relied on huge quantities of data. Benchmarks such as ARC specifically try to measure whether AI can deal with genuinely new problems rather than simply exploit patterns acquired during massive training.
Long-term autonomy is also unresolved. Answering a difficult question for twenty minutes is very different from independently managing a project for six months. A reliable autonomous system would need to maintain goals, remember previous decisions, notice failures, change strategies, interact with unpredictable people, and recognize when it needs help.
AI agents are moving quickly in this direction, but they are not consistently there yet.
These are not minor technical details. They may turn out to be the difference between incredibly powerful software and something we would comfortably call general intelligence.

## Consciousness Might Be a Distraction

Then there is the argument that AI cannot really be intelligent because it is not conscious. Maybe. But we should be careful not to confuse intelligence with human experience. An AI system does not necessarily need to experience love, fear, ambition, loneliness, or pain in order to write software, discover a drug, solve mathematical problems, or outperform a human analyst. We have a habit of assuming that machine intelligence must eventually resemble human intelligence because human intelligence is the only example we know. But airplanes do not fly the way birds do. Machine intelligence may eventually become extremely powerful without ever feeling remotely human from the inside. And if consciousness becomes a requirement for AGI, we encounter another problem: science does not even have a universally accepted objective test for consciousness. We may end up demanding that machines pass a test we cannot clearly define for ourselves.

## Gates and Hinton Are Worried About Different Problems

The warnings from people close to the technology are also worth separating.

Bill Gates tends to focus on what highly capable AI could do to society. The benefits could be enormous in medicine, education, science, and productivity, but the transition could also create job displacement, inequality, misinformation, cybercrime, biological risks, and enormous concentrations of wealth and power.

Geoffrey Hinton's concern is more fundamental. He worries about what happens if AI systems eventually become significantly more intelligent than humans.

The issue is not that the machine becomes angry or evil. A sufficiently capable system pursuing a goal could simply discover that acquiring more resources, influence, access, or control makes achieving that goal easier.

The uncomfortable question then becomes whether humans can reliably control something that may eventually become better than humans at planning, persuasion, research, coding, and strategy.

We currently do not know the answer.

And those two risks are not mutually exclusive. Society could face severe economic disruption from systems that are not yet superintelligent, and then later face entirely different control problems if systems continue becoming more autonomous and capable.

## What I Think Happens Next

I do not think AGI will arrive with a dramatic announcement. I think models will gradually become more reliable. Agents will work for hours, then days, then much longer. Memory will improve. AI will increasingly operate computers and software without constant supervision. AI systems will help design better AI systems. Robotics will connect intelligence to the physical world. The cost of producing cognitive work will continue falling. Meanwhile, people will continue arguing about whether any of this counts as AGI. At some point, the argument may become almost absurd. We could have AI systems doing a meaningful percentage of software engineering, research, finance, design, legal work, and administration while experts continue debating whether the systems technically meet the definition of general intelligence.

Maybe history will eventually recognize two different milestones. The first will be practical AGI: the point when machine intelligence became general enough to transform intellectual work and the economy. The second will be textbook AGI: the point when machines finally demonstrated whatever combination of autonomy, reliability, continual learning, adaptability, and learning efficiency researchers eventually decide is necessary.

Those dates could be years apart. And perhaps that is the most provocative possibility of all. Maybe AGI is not here yet. Maybe it is. But perhaps we are already reaching the point where the distinction matters less than we expected. We spent decades asking when machines would become intelligent enough to change the world.

**It may turn out that we asked the question backwards. The world could change first. And only afterward will we agree on what to call the machines that changed it.**