Ask ten people who build AI systems when AGI arrives and the answers barely overlap. Ben Goertzel points at 2027. Sam Altman talks about “a few thousand days.” Demis Hassabis says five to ten years. Yann LeCun argues the current approach won’t get there at all and that we’re talking decades. When roughly 2,800 AI researchers were surveyed in 2024, the median guess landed at 2047, with a 10% chance by 2029 and a 90% chance somewhere around 2110.
That spread isn’t noise, and it isn’t mainly about who’s smarter. It’s about what question each person is actually answering. Strip a forecast down and you find a stack of assumptions about definitions, hardware, data, money and human behaviour. Change any one of them and the date moves by a decade. The range of published researcher timelines makes more sense once you see which assumptions are doing the heavy lifting.
They’re not forecasting the same thing
“AGI” has no agreed definition, and the definitions in circulation differ enough to shift the date by 20 years on their own. OpenAI’s charter describes highly autonomous systems that outperform humans at most economically valuable work. Hassabis tends to frame it around most cognitive tasks. Others insist on embodiment, continual learning, or the ability to invent new science rather than summarise existing science.
A system that matches a median human freelancer on well-specified tasks is a much lower bar than a system that can run a research programme for two years without supervision. Both get called AGI in headlines. They are not the same target.
The threshold matters just as much. A date for a 50% probability is a very different claim from a date for a 10% probability, and people quote them interchangeably. When someone says 2030, ask which number they mean.
The scaling bet is the single biggest fork
Most of the disagreement traces back to one empirical question: do losses keep falling as you add compute, data and parameters, and do capabilities keep emerging as a result?
The extrapolationist case
Training loss has followed smooth power laws across many orders of magnitude of compute. If that holds, and if capability tracks loss, then the constraint is industrial rather than scientific: build bigger clusters, find more data, keep going. Timelines compress to a handful of years because the binding constraint becomes chip fabs and power plants, not breakthroughs. Epoch AI’s tracking shows training compute for notable models rising roughly fourfold to fivefold a year, which makes a 2030 arrival arithmetically plausible.
The wall case
Sceptics accept the power laws and dispute the inference. They point out that benchmark gains have started to look like better pattern matching rather than new reasoning, that claims of “emergent” abilities partly reflect how we score tests, and that the last 10% of reliability is where most of the economic value sits. On this reading, another 100x of compute buys a nicer chatbot, not a scientist. The parts of the route to AGI that are genuinely well-mapped and the parts that aren’t explain why two informed people can look at identical evidence and land 30 years apart.
Physical constraints get priced in differently
Forecasts that ignore supply chains tend to be optimistic. Forecasts that over-weight them tend to be pessimistic.
- Data. Epoch AI estimates that stock of high-quality public text could be substantially exhausted somewhere between 2026 and 2032. Synthetic data and multimodal sources may fill the gap, or may not.
- Power. Frontier training runs are moving into the gigawatt range. Grid interconnection queues in the US run three to seven years in many regions, which puts a hard ceiling on how fast compute can grow regardless of chip supply.
- Chips and packaging. Advanced packaging capacity, not wafer capacity, has been the recurring bottleneck. Anyone modelling chip supply as elastic will produce a much earlier date.
None of these are scientific unknowns. They’re scheduling problems, which means people who work inside them naturally produce later dates than people who reason from pure capability curves.
Different methods, different answers
Some forecasters reason from the inside: what would it take, step by step, to build this? Others use the outside view and ask how often similar breakthroughs have arrived this fast. The outside view tends to produce wider distributions with fatter tails; the inside view produces confident single dates.
Aggregated markets sit somewhere in between. They’re useful precisely because they force a number and update it in public, though they carry their own distortions in thin, illiquid contracts. If you want to know how much weight to put on a market-implied date, reading prediction market odds carefully is worth the ten minutes.
Incentives quietly shape the number
A lab raising capital benefits from a near date. A researcher whose funding depends on the current paradigm failing benefits from a distant one. Neither group is lying, but both are selecting which uncertainties to emphasise. The 2024 survey median moved roughly 13 years earlier than the 2022 edition, which is a large shift for two years of actual progress, and part of that is a change in who answers the survey.
This is why checking a forecast for its hidden assumptions tells you more than the headline date does. Who wrote it, what definition they used, what they assumed about compute growth, and whether they had a position to defend.
Deployment friction rarely makes it into the model
Capability timelines and adoption timelines are not the same thing, and the gap between them is where a lot of forecasts quietly break. A system that scores 90% on a coding benchmark still needs a human reviewing its pull requests. Legal liability, auditability, procurement cycles and the long tail of messy real-world tasks all slow the translation from demo to deployed product.
History offers a rough guide: electrification took about 40 years from the first practical factories to measurable productivity gains across US manufacturing, mostly because plants had to be redesigned around the new technology. AI diffuses faster than that, but not instantly.
What a single date is actually hiding
When you see “AGI by 2032,” several separate numbers have been collapsed into one:
- The capability bar being targeted, and how it’s measured
- The rate of algorithmic efficiency improvement, often assumed at roughly a doubling every eight months
- Assumed growth in training compute, and whether power is a constraint
- Whether data limits bind or are solved by synthetic and multimodal sources
- The probability level attached (10%, 50%, 90%)
- Whether the forecast describes a model in a lab or a system deployed across an economy
Two forecasts can agree on the date and disagree on every input. Weighing the evidence for and against AGI arriving within a human lifetime is largely an exercise in checking each of these line items rather than picking a camp.
The numbers that update faster than the forecasts
Decade-scale predictions are almost unfalsifiable in the short term, which is why they survive so long. Shorter-horizon measurements are more useful for anyone trying to track the actual trajectory. Watch these instead:
Task-length doubling time. METR’s evaluations of how long a task an agent can complete autonomously have shown a rough doubling every seven months. If that holds, an agent that handles four-hour tasks today handles eight-hour tasks next year. If it stalls, the extrapolationists have a problem.
Price per unit of capability. The cost of hitting a fixed benchmark score has fallen by roughly an order of magnitude a year. Falling prices expand the set of economically viable deployments, which feeds back into investment.
Algorithmic efficiency. Progress here substitutes for compute. If efficiency gains slow while compute growth also slows, the curve flattens faster than most published timelines assume.
Benchmark saturation. When a benchmark stops separating good models from great ones, it stops telling you about capability and starts telling you about test design. Watch which benchmarks people quietly stop quoting.
The honest position is that nobody knows the date, because the date depends on choices not yet made about data, power, capital and regulation. What you can do is identify which assumption a given forecast leans on hardest, then watch that assumption directly. That beats arguing about the year.
If you want to see how those assumptions play out under different conditions, sketching out a few concrete AGI scenarios for the next decade is a better use of an afternoon than picking a single number and defending it.