Ask ten people who build AI systems whether AGI will arrive before they die and you will get ten different answers, delivered with wildly different levels of confidence. Some say the odds are better than even within fifteen years. Others, including researchers who have worked on the problem for four decades, think the question is barely coherent. The disagreement is not really about technology. It is about what counts as general intelligence, and how much of human competence is actually the kind of thing you can train.
That makes a straight yes-or-no answer dishonest. What you can do is look at the specific arguments on each side, the numbers behind them, and the signals that would tell you which camp is winning.
First, agree on what you are asking
AGI, or artificial general intelligence, usually means a system that can learn and perform any intellectual task a human can, at roughly human level or better, without being purpose-built for that task. The trouble is that definition hides a hundred judgement calls. Does it need to hold down a job for a year? Write a novel that a publisher would actually buy? Teach itself a new field from scratch?
The boundaries are so fuzzy that serious researchers often argue past each other. If you want the long version of that argument, there is a good breakdown of what human-level AI really means and why nobody agrees. For our purposes, the practical version is this: a system that can be dropped into a genuinely unfamiliar problem and make progress on it the way a smart, motivated person would.
The case that it arrives sooner than you expect
The bullish argument rests on one observation that has held for more than a decade: when you scale up compute and data, capability follows. Not linearly, but reliably, and often further than the sceptics predicted.
Compute keeps showing up
The compute used to train frontier models has been doubling roughly every six months. GPT-2 in 2019 was a 1.5 billion parameter model that could barely finish a paragraph. Six years later, training runs cost hundreds of millions of dollars and labs are signing deals for gigawatt-scale power. The bottleneck has moved from chips to electricity and land, and the money is chasing it. Getting thousands of accelerators to work together without wasting half their throughput is now an engineering discipline of its own, and even scheduling GPU clusters effectively has become a research area with real dollars attached. When capital and engineering talent both point at the same wall, walls tend to move.
The track record favours the optimists
In 2015, most people would have laughed at the idea of a machine passing a bar exam, writing working code, or holding a conversation about a photo of your kitchen. Those all happened. Each time a milestone fell, the goalposts shifted to something vaguer: creativity, common sense, long-horizon planning. Goalposts shifting is not proof of progress, but it is a pattern. Forecasting communities have also moved. The Metaculus aggregate for “weak AGI” has drifted earlier as model capabilities accumulated, and surveys of AI researchers put the median estimate for machines outperforming humans on most tasks somewhere in the 2040s to 2060s, with a meaningful minority betting on the 2030s.
The case for doubt
Sceptics do not deny the progress. They argue that scaling has a shape, and the shape is not a straight line to generality.
Sample efficiency is still embarrassing
A child learns what “hot” means from one touch. Frontier models burn through trillions of tokens to approximate the same concept, and still get it wrong in ways a toddler would not. If general intelligence requires learning from a handful of examples in a new domain, current architectures have no obvious mechanism for it. More data does not fix that, because the problem is not the quantity of data. It is what the model does with it.
The last stretch is the long one
A system that is 95% reliable at a task is a demo. A system that is 99.99% reliable is a product. Those are very different engineering problems, and the gap between them is where most of the money and most of the years go. Autonomous driving has been “two years away” for a decade for exactly this reason. Agentic AI is showing the same pattern: impressive on a curated demo, brittle across a five-hour workday, because errors compound. Ten steps at 95% accuracy gives you a 60% chance of finishing cleanly.
Intelligence may not be one thing
Human cognition looks less like a single general-purpose engine and more like a bundle of specialised modules glued together by culture and language. If that is true, then chasing a single scalar “general intelligence” could be like chasing a single number for “fitness”. The concept falls apart under scrutiny, and so does the timeline.
Where open models change the picture
One factor that rarely appears in 2019-era forecasts is how fast open-weight models have closed the gap. Each year the best downloadable model gets closer to the frontier, running a year or two behind rather than a decade. That matters for the AGI question in two ways. It means any capability breakthrough diffuses quickly to thousands of teams, which accelerates follow-on progress. It also means nobody can quietly sit on a general system, which changes the safety and governance maths considerably. The state of open models in summer 2026 is a decent snapshot of just how narrow that gap has become.
Signals worth watching instead of arguing
Timeline debates are mostly unresolvable, but some measurable indicators will tell you which way things are actually going. Watch these rather than the press releases:
- Task horizon length. How long a task, measured in human working hours, a model can complete unaided before it fails. This has been roughly doubling every seven months. If it keeps going, it is the single most informative number in the field.
- Cost per unit of capability. Inference prices have collapsed. When a task that cost a dollar costs a cent, deployment patterns change overnight, and deployment is where real-world competence gets tested.
- Novel research output. Not summarising papers, but producing a result that human experts did not already know. That is the line between impressive retrieval and genuine reasoning.
- Robotics. Anything that has to work in a messy physical world exposes gaps that benchmarks hide. Progress in vision-language-action models is a good proxy for whether the intelligence transfers.
- Economic absorption. Labour markets are slow, but they are honest. Where AI quietly becomes the default tool, as it is starting to in parts of the design stack, you are seeing capability that survives contact with real work.
So, will AGI exist in our lifetime?
The honest answer is that it depends on which lifetime you mean, and that is not a dodge. For someone who is 25 today, the question spans roughly sixty years. That is a very long time in a field that went from toy chatbots to competent coding assistants in under a decade. For someone who is 75, the window is a decade, and the odds are considerably worse. If you are asking on a fifteen-year horizon, most informed estimates land somewhere between “plausible” and “coin flip”. On a forty-year horizon, the sceptical position becomes much harder to defend, simply because compounding progress is unforgiving.
The more useful reframe is that AGI probably will not announce itself with a press release. It will look like a series of systems that are slightly too good at your job, then your colleague’s job, then most jobs. By the time the label fits, the practical consequences will already be in place. Whether or not the term survives the next few decades, the thing it points at is being built in front of us, at a rate that keeps surprising the people doing the building.
Plan for the version where it arrives and the version where it stalls at 90% of human competence. Both are live possibilities, and the near-term work of making systems reliable, auditable and cheap is the same in either case.