The intelligence threshold
General AI: When Does the Clock Actually Hit Zero?
Something interesting happened between 2023 and 2026. The people who build AI systems — not science fiction writers, not futurists on podcasts, but the CEOs, chief scientists, and co-founders of the most well-funded AI labs in history — started giving very specific timelines for when Artificial General Intelligence would arrive. Not "someday." Not "within our lifetimes." Years. Sometimes months.
Meanwhile, a parallel group of equally serious researchers — including Turing Award winners and the architects of the neural networks underlying all modern AI — said those timelines were fantasy. That current systems, however impressive, are not on the path to general intelligence at all. That we are confusing performance on benchmarks with something much harder: actually understanding the world.
Both camps cannot be right. This article maps the positions, explains what AGI actually means, and presents the five strongest arguments for why it might be further away than the people building it would like to believe.
What is General AI — and what it is not
Artificial General Intelligence (AGI) is the ability of an AI system to understand, learn, and apply knowledge across any domain — not just the domains it was trained on — at a level that matches or exceeds a competent adult human. It is the difference between a system that is very good at a fixed range of tasks and one that can figure out what to do in a genuinely novel situation it has never encountered.
Today's AI, including the most capable large language models (LLMs), is something different. It is narrow in a deceptively broad way. A model like GPT-4o or Gemini 1.5 can write code, summarise legal contracts, pass the bar exam, and generate photorealistic images. But it does this by pattern-matching against trillions of examples it was trained on. It does not know anything in the way a human knows things — it predicts what tokens come next with extraordinary skill.
Today's AI (Narrow AI)
- Excels within its training distribution
- Requires vast labeled data or human feedback
- Cannot learn after deployment without retraining
- No persistent memory across sessions
- Cannot set its own goals
- Fails unpredictably on edge cases
- No embodied understanding of the physical world
- Performance degrades outside training domain
General AI (AGI)
- Transfers knowledge across all domains
- Learns continually from minimal examples
- Updates its own world model in real time
- Maintains persistent, coherent memory
- Sets and pursues goals autonomously
- Handles genuinely novel situations
- Understands causality, not just correlation
- Consistent performance regardless of domain
The distinction matters because AGI is not just "a smarter LLM." It represents a qualitative shift in the nature of the system — from a tool that performs tasks to an agent that reasons about tasks and can construct new ones. Whether that shift requires fundamentally different architectures, more compute, or something we have not invented yet is precisely what the debate is about.
The difference between today's AI and AGI is not the difference between a pocket calculator and a scientific calculator. It is the difference between a calculator and a mathematician.
Predicted arrival year: who forecasts what
The entries below are sorted by predicted target year — when each person believes AGI will arrive — not by when they made the statement. Quotes are drawn from on-record interviews, blog posts, congressional testimonies, and public presentations.
firstcondition
pathparadigm
The definition problem
One reason timelines differ so dramatically is that nobody agrees on the definition. Elon Musk says Grok 3 is "close to AGI." Yann LeCun says no current system is remotely close. Both statements can be simultaneously true if they are using different definitions. Sam Altman's OpenAI has an internal document defining AGI as "a highly autonomous system that outperforms humans at most economically valuable work" — a definition that conveniently excludes philosophical requirements like consciousness or embodiment. Most academic researchers use a stricter definition that includes general transfer learning, causal reasoning, and continual learning.
Five reasons serious people think AGI is further away than the headlines suggest
The following arguments are not from AGI deniers. They come from researchers who accept that LLMs are impressive, that progress is real, and that the trajectory is meaningful. Their objection is specifically to the claim that we are one or two architectural iterations away from general intelligence.
We keep moving the goalposts — and that should concern us
In 2012, playing Go at grandmaster level was considered an AGI-level task. In 2016, AlphaGo did it — and within months, experts explained why it was "just" narrow AI. In 2020, writing coherent paragraphs was a tentative AGI test. By 2023, GPT-4 passed bar exams and medical licensing tests. Each time a system passes the test, the test is retired and a harder one is invented. This is not dishonesty — it is a genuine empirical signal. Every time we build a system that does the thing we said required general intelligence, we discover it did not actually require general intelligence. The goalposts move because our intuitions about what requires "real" intelligence were wrong. That may mean AGI is an illusion of definition. Or it may mean we have not yet hit the task that actually requires it.
Current architectures cannot truly reason — they pattern-match at extraordinary scale
When GPT-4 solves a complex maths problem, it is not deriving the solution the way a mathematician would. It is interpolating within a distribution of mathematical reasoning patterns it has seen before. Researchers have repeatedly shown that trivially rephrased versions of problems current AI solves correctly will cause it to fail completely — something no human mathematician would do. This "fragility under distribution shift" is not a bug that can be fixed with more data. It suggests a fundamental architectural issue: transformers trained on next-token prediction do not build internal world models. They build statistical summaries of the training corpus. The difference only becomes visible at the edges of that corpus.
Embodiment is not optional — and we are ignoring it
A two-year-old child knows that if you put an object behind a box, it does not disappear. This sounds trivial, but it requires understanding object permanence, physical causality, and the persistence of the world independent of observation — concepts acquired through years of physical interaction with the environment. LLMs know this fact because it is stated in their training data. But they do not know it the way the child knows it — as a grounded, embodied understanding. This distinction, between knowing a fact about physics and having an intuitive, predictive model of physics that generalises to completely new situations, is precisely what is missing in current AI. Cognitive scientists and developmental psychologists argue this kind of grounded understanding may not be achievable through text alone, regardless of scale.
Continual learning remains an unsolved problem
Humans learn constantly, updating their knowledge and skills without forgetting everything they knew before. Neural networks do not. When a neural network is trained on new data, it tends to catastrophically forget previously learned information unless specifically prevented from doing so — and the techniques for preventing this do not scale to the kind of open-ended, lifelong learning that characterises human cognition. Every major AI model you can name today is essentially frozen at the moment of training. It can be prompted and fine-tuned, but it cannot genuinely learn from new experiences the way a human employee learns on the job. Solving continual learning at scale is not an incremental improvement to current architectures. It is a research problem that has resisted solution for decades.
The energy and hardware wall may be more constraining than it appears
Training GPT-4 consumed an estimated 50 gigawatt-hours of electricity — roughly the annual consumption of 4,500 American homes. The next generation of frontier models is projected to require data centres drawing hundreds of megawatts of continuous power. The scaling hypothesis — the idea that simply adding more compute will eventually yield AGI — requires not just more compute, but orders of magnitude more compute than we can currently build or power. Microsoft, Google, and OpenAI are collectively planning over $500 billion in AI infrastructure investment through 2030. But even at that scale, the compute required to keep pace with the scaling curves that have driven recent progress begins to exceed the output of entire national power grids. At some point, the question is not whether the architecture will scale — it is whether the physical infrastructure to support it can be built at all.
What this means in practice
The honest answer is that we do not know when AGI will arrive, and anyone who says they do is either using a very specific (and possibly self-serving) definition of AGI or expressing a confidence that the evidence does not support.
What we do know:
- The trajectory of AI capability over the last five years has consistently exceeded the expectations of the people building it, including the people making the most optimistic public predictions.
- The transition from narrow to general AI, if it occurs, will likely not be a single dramatic moment. It will be a gradual expansion of capability across more and more domains until the remaining gaps are small enough to be definitionally argued away.
- The economic pressure to declare AGI — from companies that have raised hundreds of billions on the promise of it — creates a significant incentive distortion in how public predictions are framed.
- The researchers least financially exposed to the outcome — academic AI safety researchers, cognitive scientists, philosophers of mind — are, on average, significantly more skeptical of near-term timelines than the researchers most financially exposed.
A benchmark worth watching
ARC-AGI (Abstraction and Reasoning Corpus), created by Francois Chollet at Google, was designed explicitly to be resistant to the pattern-matching approach that has driven LLM progress. It requires genuine reasoning from minimal examples. OpenAI o3 reached ~85% on the original ARC-AGI benchmark in late 2024, matching average human performance. ARC-AGI 2, released in 2025 with significantly harder tasks, has proven far more resistant — frontier models score in the low single digits there, while humans score around 60%. The rate of improvement on this benchmark, which cannot be gamed by training on more data, may be the most honest signal we have about whether real progress toward general reasoning is happening.
The uncomfortable truth
We are almost certainly closer to AGI than we were five years ago. The uncertainty is not about direction — it is about distance. And distance, in this case, is genuinely unknown. It could be five years. It could be fifty. It could require a fundamental paradigm shift that no one has had yet.
What is clear is that the current moment is not one for either complacency or panic. The systems we have today are already economically and socially transformative. They are reshaping every knowledge industry, including the one you work in. Whether or not AGI arrives in 2029 or 2049, the practical intelligence threshold — where AI can perform most knowledge work better and cheaper than most humans — is a more tractable question. And for that threshold, the timeline is measured in years, not decades.
The question is not whether to prepare for a world with general AI. The question is whether the preparation happening now is commensurate with the magnitude of what may be coming.
The people predicting AGI by 2027 and the people predicting AGI by 2047 agree on one thing: it is the most consequential technology transition in human history. The gap in their timelines is twenty years. That gap matters enormously for policy, for investment, and for how organisations should position themselves. It is not a gap that will be resolved by waiting.