Ask five AI researchers when AGI arrives and you’ll get five answers spread across three decades. Ray Kurzweil has held to 2029 for years and shows no sign of moving. Dario Amodei has described a “powerful AI” arriving around 2026 or 2027. Demis Hassabis talks about roughly a decade, with real uncertainty on either side. Geoffrey Hinton puts it somewhere between five and twenty years. Yann LeCun thinks large language models will never get there and we’re looking at decades of additional research.
That spread isn’t laziness or evasion. It’s what happens when smart people disagree about which bottleneck breaks first.
Why no one can hand you a clean date
Part of the problem is definitional. If AGI means “matches a median human on most economically valuable cognitive work,” some researchers will argue it’s already here in narrow slices and a decade away in general. If AGI means “learns any new skill about as fast as a person can,” almost nobody outside a marketing department thinks we’re close. The definition you pick moves the date by ten years or more before anyone has said anything substantive.
The other problem is that progress has been lumpy. Long stretches pass where benchmarks saturate and everyone declares a plateau, then a step change resets expectations overnight. GPT-3 landed in 2020. ChatGPT in late 2022. Reasoning models in 2024. Each jump arrived faster than most working researchers predicted, and each one made the next prediction harder rather than easier.
The four bottlenecks that actually set the timeline
Compute and power
Frontier training runs have grown by roughly 4x to 5x per year in effective compute, paid for by clusters that now run into the billions of dollars. A facility holding 100,000 GPUs draws somewhere near 150 megawatts once you count networking and cooling. Chip supply stopped being the binding constraint a while back. Electricity, transmission queues and the four to six years it takes to bring new generation online are the slow variables now, which is why labs keep signing power deals with nuclear and gas operators.
Data
Researchers at Epoch AI have estimated that the stock of high-quality public text could be substantially consumed somewhere between 2026 and 2032. Synthetic data, licensed archives and video push that wall further out. What nobody has shown cleanly is that synthetic data teaches genuinely new reasoning rather than quietly reinforcing patterns a model already had.
Sample efficiency
A teenager learns to drive competently in about twenty hours of practice. A driving model needs millions of miles. That ratio is the heart of the technical argument, and it’s the reason LeCun keeps pushing architectures that build internal world models instead of predicting the next token.
Reliability under composition
This one gets ignored and it shouldn’t. A system that succeeds on a step 97% of the time sounds excellent until you chain twenty steps together. Run the arithmetic and you get about 54% end-to-end success. That’s why agent demos look like magic on a stage and stall in production. Getting from 97% to 99.9% per step is a grind measured in years, not months.
What the builders themselves say, versus what the marketing says
Public forecasts from lab leaders tend to be more conservative than the framing around product launches, and considerably more conservative than social media commentary about them. When you read the papers, the safety evaluations and the limitations sections, the picture that emerges is usually “impressive narrow capability plus known failure modes,” not “general intelligence next quarter.” Reading what the people building these systems actually say in interviews is a better use of an hour than refreshing a leaderboard.
Roadmaps are direction, not schedule
Every major lab now publishes some version of a capability roadmap with numbered levels and tidy progression. Every conference produces a slide with an arrow pointing up and to the right. These documents are genuinely useful for understanding what a company is optimising for. They are close to useless as calendars, because the levels are usually defined after the fact, once researchers know what the model can do.
If you follow this stuff closely, it pays to know how to read roadmap updates without getting fooled, and to develop an eye for the stagecraft in conference announcements rather than the numbers on the slide. A demo built for a keynote has been rehearsed. A benchmark result published with the harness and prompts has not.
Signals that would genuinely move your estimate
Most things reported as AGI progress are noise. A short list of things that aren’t:
- Multi-hour autonomous work with error recovery. Not a five-minute demo. A system that gets stuck, notices, backs up and tries a different approach over the course of a working day.
- Learning from its own experience. A model that improves at a task through practice without a fresh round of human-labelled data or a human-written fine-tune.
- Transfer without retraining. A skill picked up in one domain showing up in a structurally different one, unprompted.
- Cost curves falling faster than capability rises. Cheap and slightly worse usually beats expensive and slightly better, and it changes who can build what.
Things that don’t move your estimate: another benchmark record, another viral video, another round of funding. It’s worth learning to separate what matters at a model launch beyond the demo reel, and to filter safety claims through which safety signals are real and which are positioning. The safety discourse in particular swings between complacency and doom with very little in between, and neither extreme is well calibrated.
How to plan when the date is genuinely unknown
You don’t need a forecast to make good decisions. You need tripwires.
Pick outcomes specific enough that you’d notice them and check them on a schedule. For example: if by mid-2027 a commercially available system can take a week of junior developer tickets and ship working, reviewed code with minimal supervision, that changes hiring plans, pricing and how you scope projects. If by 2027 agents can handle a full customer support queue end to end at 90% resolution with escalation for the rest, that changes headcount math in a call centre. Write the trigger down now, while you’re calm, and decide in advance what you’d do if it fires.
The second move is building for optionality rather than a specific scenario. Systems that keep human review cheap to insert, data that isn’t locked to one vendor’s format, workflows documented well enough that a model can be dropped in later. None of that requires knowing whether AGI lands in 2029 or 2045, and all of it pays off if it lands sooner than you expected.
The third is patience with the forecast itself. The honest expectation, given the published ranges from people with real information, is a median somewhere in the early-to-mid 2030s with a fat tail on both ends. That’s a boring answer. It’s also the one most likely to survive contact with reality, and it leaves you roughly a decade to get ready. Which, conveniently, is about the length of time it takes to rebuild a power grid, retrain a workforce or ship a product that holds up when the models underneath it keep changing.