Ask Dario Amodei when AGI arrives and you’ll hear 2026 or 2027. Shane Legg has been saying 2028 for years. Demis Hassabis tends to talk in terms of a decade. Yann LeCun argues the current recipe won’t get there at all. That spread looks like a fight about engineering. Mostly it’s a fight about definitions.
A system that matches a median human across every cognitive task, a system that outperforms most people at most economically valuable work, and a system that can teach itself a new field over a weekend all arrive on different schedules. Those schedules could sit decades apart. It’s worth knowing that the AGI countdown is measuring something narrower than most headlines suggest, because the number on the clock changes depending on which definition you started it with.
Where the published numbers actually land
Strip out the marketing and you get five rough clusters, each tied to a different bar for success.
- 2026–2027. Amodei’s “powerful AI” and the most aggressive lab-internal forecasts. Read the fine print: his own definition is “a country of geniuses in a datacenter,” not a machine that can do anything a person can.
- 2028. Shane Legg’s long-standing estimate, restated many times since 2011. It has the virtue of being falsifiable soon.
- Early 2030s. Where the Metaculus forecasting community has hovered for the past couple of years on its “weak AGI” question, which asks for a system that passes a modest battery of human-level tests.
- 2047. The midpoint from a 2024 survey of 2,778 AI researchers, which also gave roughly a one-in-ten chance of arrival by 2028. Notice how wide that distribution is coming from the people who’d know best.
- Never, or effectively never. A real minority position, not a trolling one, usually built on claims about embodiment, continual learning, or the cost of error-free reasoning.
Put those side by side and the honest reading is a fat-tailed distribution, not a countdown. That’s the core argument in why credible AGI arrival dates span from 2026 to never, and it should shape how much weight you put on any single date you read this month.
The compute case, and the wall it’s aimed at
The optimists aren’t guessing. Training compute for frontier models has grown roughly four to five times a year since 2010, according to Epoch AI’s tracking, and capital expenditure has followed. Meta, Microsoft, Amazon and Google collectively committed well over $200 billion to AI infrastructure in a single year. When you can point at a curve that smooth and cheque books that large, a near-term arrival date stops sounding mystical.
Two constraints push back. The first is data: high-quality human text is finite, and estimates for when the useful pool runs dry cluster around the late 2020s. Synthetic data and reinforcement learning environments are the workaround, and they’re promising, but they’re also less proven than scraping the open web was. The second is power and permitting. A single large training run can pull tens of megawatts, and new substations take years, not quarters, to energise.
Watching those two variables tells you more than any keynote. The industry shifts that genuinely change the AGI timeline tend to be unglamorous: chip yields, grid connections, long-term supply contracts.
Benchmarks saturate, then get replaced
MMLU was the gold standard in 2021. By 2024 frontier models were scoring above 88%, and the test had stopped discriminating. Something similar happened faster with maths and coding contests. Then came ARC-AGI: GPT-4 managed a few percent in early 2023, and in December 2024 OpenAI’s o3 hit 87.5% on the semi-private set, at a reported cost of around $20 per task. That’s a genuine jump and a genuine caveat in the same sentence, because a system that solves a puzzle for $20 a go is not yet a system you’d hand a payroll department.
The goalposts do move, and some of that moving is self-serving. Some of it is also just what progress looks like when the easy tests get eaten first. Distinguishing the two is a skill in itself, which is roughly the thesis of what’s actually moving the needle in AGI research versus what’s noise.
The bottlenecks that never show up in demo videos
Reliability is the quiet killer. An agent that completes each step of a task correctly 95% of the time sounds excellent until you chain twenty steps together: 0.95 to the power of 20 is about 36%. Long-horizon autonomy needs per-step accuracy measured in the 99.9% range, and getting there is a different kind of problem from making a model smarter.
Cost has moved fast in the other direction. Stanford’s AI Index put the price of GPT-3.5-level inference down roughly 280-fold between late 2022 and late 2024. That collapse is a big part of why AI feels suddenly ubiquitous even though capability itself improved more gradually. Energy and latency are the remaining drags, especially for anything that has to reason for minutes rather than seconds.
Signals that genuinely move the date
If you want an early read, watch for capability that transfers rather than capability that impresses. A model that tops a benchmark its trainers didn’t design for is news. A model that tops one they did is advertising. The same goes for agents that finish multi-day projects without supervision, and for research自动化 claims that come with reproducible artefacts attached.
Roadmaps are worth reading, just not at face value. Companies publish them to recruit, to raise, and to shape regulation, and the dates inside them shift for reasons that have nothing to do with science. Reading AGI roadmap updates without getting fooled mostly comes down to tracking what slipped and why, not what got announced.
The regulatory drag nobody prices in
Rules are now a live variable rather than a background hum. The EU AI Act’s general-purpose model obligations kicked in during 2025, US export controls keep reshaping which labs can train what, and a patchwork of state and national laws is forming faster than most forecasts assumed. Compliance work doesn’t stop capability research, but it does slow deployment, and deployment is where the economic evidence for AGI claims would actually come from. AGI policy news has been outpacing the models themselves for two years straight.
How to carry a timeline around without getting whiplash
Hold a range, not a date. Something like “meaningfully transformative systems are plausible in the late 2020s and likely by the mid-2030s, with a real chance the whole framing is wrong” survives contact with new information far better than a countdown timer.
Then attach conditions to it. Write down, in plain language, what would have to be true for you to move your estimate up or down by three years. Examples that earn their keep:
- An agent completing a multi-week project with less than 1% human intervention.
- Training runs continuing to scale after the human text pool is exhausted.
- Frontier inference cost flattening instead of falling.
- A capability jump on a benchmark the lab didn’t help design.
Check those conditions quarterly rather than daily. The daily news cycle rewards drama, and the AGI development timeline is a slow-moving variable wearing a fast-moving costume. People who revise their view every week end up anchored to whoever shouted last.