Spend five minutes on LinkedIn and you will find a venture capitalist promising that artificial general intelligence is "just a few thousand days away." OpenAI's Sam Altman has characterized AGI as "inevitable" in public statements. Anthropic's Dario Amodei has warned that we must prepare for a world where powerful AI systems reshape every domain. The message is unanimous, confident, and repeated so often that it has achieved the status of background noise: AGI is coming, it is unstoppable, and the only question is whether we are ready.
Here is the problem. After seventy years of research, nobody can agree on what AGI actually is. The people selling you inevitability cannot define the product. And the entire argument rests on a logical structure that would be treated with far more skepticism in fields with clearer criteria for evidence.
This is not a story about killer robots or utopian mind uploads. It is a story about a sales narrative that has outpaced its evidence — and about what happens when a trillion-dollar industry convinces itself, and its investors, that a 1955 hunch is actually a law of physics.
The Conjecture That Launched an Industry
In August 1955, four researchers — John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon — wrote a funding proposal for a summer research project at Dartmouth College. The document, which introduced the term "artificial intelligence," made a bold claim:
"The study is to proceed on the basis of the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it."
— McCarthy et al., A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence (1955)
Note the word: conjecture. Not theorem. Not observation. A guess. The authors thought ten people working for two months in the summer of 1956 would produce "significant advances" on language, abstraction, concept formation, and self-improvement. They were wrong. The summer happened. The advances did not. And the field has been renegotiating that conjecture ever since.
What the Dartmouth proposal gave the world was not a scientific breakthrough. It was a business model: assume intelligence is computable, build systems that appear intelligent, and declare progress toward the ultimate goal. That model has now raised hundreds of billions of dollars. But the conjecture itself has never been proven. In fact, the more we have learned about intelligence, the less computable it looks.
Nobody Knows What Intelligence Is
Here is a sentence you will not hear at an AI startup demo day:
"A fundamental problem in artificial intelligence is that nobody really knows what intelligence is. The problem is especially acute when we need to consider artificial systems which are significantly different to humans."
— Shane Legg and Marcus Hutter, "Universal Intelligence," Minds and Machines (2007)
Legg, who co-founded DeepMind, and Hutter, a mathematician at Australian National University, published this in a peer-reviewed journal. They were not being contrarian. They were stating a prerequisite. You cannot build a machine that matches or exceeds human intelligence across all domains if you cannot define what that intelligence consists of — or if the system you are building operates in ways fundamentally different from human cognition.
Legg and Hutter tried anyway, producing a formal mathematical definition based on universal Turing machines. It is elegant, rigorous, and completely inoperable. It tells you what intelligence might look like in the abstract. It tells you nothing about how to build it, test it, or recognize it if you saw it.
This is the dirty secret of the AGI conversation: the closer you get to the people actually doing the foundational research, the more humble they become about the endpoint. The certainty increases as you move away from the lab and toward the pitch deck.
What LLMs Actually Do
Large language models — ChatGPT, Claude, Gemini, and their competitors — are impressive. They can write code, summarize documents, pass bar exams, and generate prose that sounds human. They are also, by every technical definition available, narrow AI systems trained on a single prediction task: given a sequence of tokens, guess what comes next.
That is the mechanism. Whether the system truly reasons, understands, or holds a model of the world remains deeply contested. What is not contested is that its operation is statistical pattern matching at scale. The model has ingested trillions of words, calculated statistical relationships between them, and learned to produce outputs that mimic the form of reasoning. It does not — at least in any recognized sense — reason from first principles. It does not understand in the way a human understands. It has a model of language, specifically of which words tend to appear near which other words, not a model of the world those words describe.
Noam Chomsky, the founder of modern linguistics, put it bluntly in a 2023 New York Times essay:
"The most prominent strain of A.I. encodes a flawed conception of language and knowledge."
— Chomsky et al., "The False Promise of ChatGPT" (2023)
John Searle, the philosopher who devised the famous "Chinese Room" argument in 1980, made the same point four decades earlier. Searle argued that a computer processing symbols according to rules does not thereby understand what those symbols mean — it merely manipulates formal patterns. He offered a pointed analogy:
"No one supposes that computer simulations of a five-alarm fire will burn the neighborhood down. Why on earth would anyone suppose a computer simulation of understanding actually understood anything?"
— Searle, "Minds, Brains, and Programs," Behavioral and Brain Sciences (1980)
A common response to this critique is that the system gets the right answers, and that is what matters. Which is true, sometimes. And also the point. A system that produces correct outputs without understanding is a sophisticated autocomplete, not a mind. A calculator gets the right answers too. Nobody calls it intelligent.
The Scaling Mirage
The dominant narrative in AI right now is that scale solves everything. If the model is not smart enough, make it bigger. If it hallucinates, train it on more data. If it cannot reason, add more GPUs. This is not a theory. It is a capital allocation strategy dressed up as science.
François Chollet, the software engineer who created the Keras deep-learning library, published a direct assault on this logic in 2017:
"No 'intelligence explosion' will occur, as this process advances at a roughly linear pace. Remember: intelligence is situational — there is no such thing as general intelligence."
— Chollet, "The Impossibility of Intelligence Explosion" (2017)
Chollet's point is devastatingly simple. Intelligence is not a dial you turn up. It is not a single quantity like temperature or voltage. Human intelligence is the product of millions of years of evolution in specific physical and social environments. It is embodied, situated, and shaped by the fact that we have bodies that get hungry, fall down, and interact with other bodies. A language model has none of this. It has no body, no environment, no goals beyond predicting the next token. Calling it a step toward "general intelligence" is like calling a weather forecast a step toward controlling the climate.
Gary Marcus, a cognitive scientist at NYU and one of the most persistent critics of the scaling thesis, has been tracking the diminishing returns:
"The most underreported and important story in AI right now is that pure scaling has failed to produce AGI."
— Marcus, Fortune (February 19, 2025)
Marcus is careful — he is not saying AI research should stop. He is saying that the specific bet made by OpenAI, Anthropic, and their competitors — that scaling transformer architectures will spontaneously produce reasoning, planning, and understanding — is not paying off. The models get better at pattern matching. They do not get better at thinking.
Yann LeCun, Meta's chief AI scientist and a winner of the Turing Award (the Nobel Prize of computing), has been even more direct:
"LLMs are far too limited to ever reach true human-level intelligence."
— LeCun, Financial Times (May 2024)
LeCun is not a Luddite. He built the foundational architecture for modern computer vision. He is saying, from inside the field, that the path everyone is currently on is a dead end. True intelligence requires systems that can model the physical world, plan, reason about causality, and learn from interaction. LLMs, at least in their current form, do not do these things in any recognizable sense. They predict text. Whether the gap is one of degree or of kind is itself disputed, but the gap is large — and growing larger in the wrong directions.
The Embodiment Problem
In 1990, Rodney Brooks, then a robotics researcher at MIT, published a paper that challenged the purely computational approach to AI on foundational grounds. The field largely ignored that challenge, in part because computational methods were far cheaper to scale and far easier to fund. But the paper's argument remains unanswered.
"We explore a research methodology which emphasizes ongoing physical interaction with the environment as the primary source of constraint on the design of intelligent systems."
— Brooks, "Elephants Don't Play Chess," Robotics and Autonomous Systems (1990)
Brooks was making a point about elephants. Elephants do not play chess. They do not need to. They navigate complex social hierarchies, find water across vast territories, care for their young, and communicate through infrasound. None of this requires abstract symbol manipulation. It requires a body in an environment, with goals that emerge from physical and social needs.
The AI startups selling inevitability have no persuasive answer to this. Their flagship models do not have bodies. They do not have environments. They do not have needs. They sit in data centers, consuming electricity and producing text. The claim that such a system will spontaneously become "generally intelligent" requires believing that intelligence is a substance that can be poured into any container, rather than a property of specific kinds of systems interacting with specific kinds of worlds.
Hubert Dreyfus, a philosopher at UC Berkeley, argued in What Computers Can't Do (1972) that AI researchers had consistently mistaken the surface features of intelligent behavior — rule-following, symbol manipulation, game-playing — for intelligence itself. He traced this error to a deeper assumption: that human expertise could be captured by explicit rules and representations, an assumption he found contradicted by phenomenology and by the history of the field.
Dreyfus was vindicated by the first AI winter, when symbolic AI funding collapsed in the late 1980s because the systems could not deliver on their promises. The current boom is different in scale but not in structure. We have better pattern matching. We still do not have intelligence.
The Inevitability Racket
Here is where skepticism becomes not just scientific but financial. The claim that AGI is "inevitable" is not a scientific prediction. It is a fundraising necessity. Startups raise capital on the promise of future returns. If the future return is a technology that does not exist, cannot be defined, and has no roadmap, the only way to keep raising capital is to insist that its arrival is a matter of time, not a matter of proof.
The structure of the argument is telling. It goes like this:
- Moore's Law means compute keeps getting cheaper.
- More compute means smarter models.
- Therefore, eventually, a model will be smart enough to be AGI.
Each step is flawed. Moore's Law describes transistor density, not intelligence. "Smarter" is undefined. And "eventually" is the tell — it is the same word used by people selling perpetual motion machines and cold fusion reactors. It shifts the burden of proof to the future, where it cannot be examined.
One of the philosophers who has thought most carefully about what happens when this logic is taken seriously is Stuart Russell, a professor at UC Berkeley and co-author of the world's most widely used AI textbook. Russell believes AGI is possible — he is not in the Dreyfus or Chomsky camp — but he has identified the structural problem with how the field is pursuing it:
"The standard model of AI specifies an objective function — a precise specification of what the system should maximize — and then builds a system that optimizes for it."
— Russell, Human Compatible (2019)
The problem, Russell argues, is that this model works for narrow tasks but breaks down for general intelligence. A system that optimizes a fixed objective will pursue that objective in ways its designers did not anticipate. The more capable the system, the more dangerous the unexpected behavior. Russell's solution — building systems that are uncertain about human preferences — is intellectually serious. It is also nowhere near what the startups are building. They are building bigger optimizers and calling it progress.
What Would It Take?
None of this means AGI is impossible. It means the current path is not taking us there, and the people claiming otherwise have either misread the evidence or have strong incentives to keep the narrative alive. A credible path to AGI would require, at minimum:
- A definition of intelligence that is operational, not philosophical. We cannot build what we cannot define or test.
- A theory of how intelligence arises from physical systems. Not just bigger models, but models that interact with environments, have goals that are not given by humans, and learn from the consequences of their actions.
- Evidence that scaling produces qualitative changes, not just quantitative ones. So far, scaling produces better autocomplete. It has not produced reasoning, planning, or understanding.
- An acknowledgment of what we do not know. The startup narrative permits no uncertainty. Science requires it.
I do not know whether AGI is possible, or what it would look like if it were. Neither, as far as I can tell, does anyone else. That is not a comfortable admission in an industry built on confident forecasts. It is the only honest one.
Eliezer Yudkowsky, a researcher at the Machine Intelligence Research Institute, has warned for years that the field's overconfidence is itself the danger:
"By far the greatest danger of Artificial Intelligence is that people conclude too early that they understand it."
— Yudkowsky, "Artificial Intelligence as a Positive and Negative Factor in Global Risk" (2008)
Yudkowsky is talking about existential risk — the possibility that a misaligned superintelligence could cause catastrophic harm. But the same overconfidence produces a more immediate risk: the misallocation of resources, talent, and public attention toward a goal that has no scientific foundation, while more tractable problems — climate modeling, drug discovery, materials science, and diagnostic medicine — compete for a fraction of the capital and talent.
The Real Mystery
The strangest thing about the AGI debate is not that some people believe it is coming. It is that so many people believe it is coming soon, inevitably, and via the current approach, despite seventy years of evidence that the current approach keeps promising and failing.
The Dartmouth researchers thought two months would do it. In the 1960s, Herbert Simon predicted that machines would be capable of any work a human could do within twenty years. In 2005, Ray Kurzweil predicted AGI by 2029 and a "Singularity" by 2045. Each generation resets the clock, moves the goalposts, and forgets that the previous generation made the same promises with the same confidence.
The difference now is the money. When Herbert Simon was wrong, a few government grants got redirected. When Sam Altman is wrong, pension funds, sovereign wealth funds, and retail investors absorb the losses. The inevitability narrative is not just scientifically unsupported. It is a liability bubble, inflated by people who have a financial interest in keeping it aloft.
AGI may arrive someday. It may not. What is certain is that the people telling you it is inevitable cannot tell you what it is, how it would work, or how they would know if they had built it. In any other field, that would be called speculation. In AI, it is called a product roadmap.
The only thing that is actually inevitable is that someone will be wrong. The history of the field suggests it will be the people making the predictions.
Works Cited
Brooks, Rodney. "Elephants Don't Play Chess." Robotics and Autonomous Systems, Vol. 6, 1990, pp. 3–15. https://people.csail.mit.edu/brooks/papers/elephants.pdf
Chollet, François. "The Impossibility of Intelligence Explosion." Medium, November 2017. https://medium.com/@francois.chollet
Chomsky, Noam, Ian Roberts, and Jeffrey Watumull. "The False Promise of ChatGPT." The New York Times, March 8, 2023. https://www.nytimes.com/2023/03/08/opinion/noam-chomsky-chatgpt-ai.html
Dreyfus, Hubert L. What Computers Can't Do: The Limits of Artificial Intelligence. New York: Harper & Row, 1972.
Legg, Shane, and Marcus Hutter. "Universal Intelligence: A Definition of Machine Intelligence." Minds and Machines, Vol. 17, 2007, pp. 391–444. https://arxiv.org/abs/0712.3329
Marcus, Gary. "The Most Underreported Story in AI." Fortune, February 19, 2025. https://fortune.com/2025/02/19/generative-ai-scaling-agi-deep-learning/
McCarthy, John, Marvin L. Minsky, Nathaniel Rochester, and Claude E. Shannon. A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence. August 31, 1955. https://www-formal.stanford.edu/jmc/history/dartmouth/dartmouth.html
Russell, Stuart. Human Compatible: Artificial Intelligence and the Problem of Control. New York: Viking, 2019.
Searle, John R. "Minds, Brains, and Programs." Behavioral and Brain Sciences, Vol. 3, No. 3, 1980, pp. 417–457. https://courses.washington.edu/info300/searle.pdf
Yudkowsky, Eliezer. "Artificial Intelligence as a Positive and Negative Factor in Global Risk." In Global Catastrophic Risks, edited by Nick Bostrom and Milan M. Ćirković, 308–345. New York: Oxford University Press, 2008. https://intelligence.org/files/AIPosNegFactor.pdf