Ask three people who build frontier models when AGI arrives and you’ll get three dates that don’t overlap. One says 2027. One says the mid-2030s. One says it isn’t a coherent question, because “AGI” keeps meaning whatever we haven’t built yet. None of them are being sloppy. They’re disagreeing about which part of the problem is hard.
That’s the most useful frame for the AGI timeline: it isn’t a countdown, it’s a set of bets about bottlenecks. Once you know which bottleneck someone thinks dominates, you can usually predict their date before they say it out loud.
Five Forecasts, Five Different Decades
Start with the raw spread, because it’s wider than most coverage admits.
A 2022 survey of nearly 2,800 researchers who publish on AI put the median estimate for “high-level machine intelligence” around 2060. The forecasting community on Metaculus has sat far earlier, with community medians drifting through the early 2030s in recent years. Demis Hassabis has repeatedly offered something close to a 50% chance by 2028. Yann LeCun has argued that the current approach needs at least one more conceptual breakthrough before the question is even meaningful, which pushes his estimate out by a decade or more.
Same field, same published evidence, answers spanning roughly 30 years. Anyone presenting a single date as settled is either selling something or hasn’t looked at the distribution.
The Three Clocks People Keep Confusing
Most timeline arguments dissolve once you separate three things that all get called “AGI arriving.”
1. The capability clock
When a system crosses whatever threshold you set: autonomy over long tasks, novel research contributions, reliable multi-step planning. This is what benchmarks try to measure, and it’s what labs mean when they say they have AGI internally.
2. The deployment clock
When that capability is cheap, safe enough, and available to ordinary users and businesses. This historically lags capability by years. A model that can write production code in a demo and a model that a mid-sized company trusts with its codebase are different products, separated by procurement, liability, and a hundred small reliability questions.
3. The consequence clock
When the economy and institutions visibly reorganise around it. This lags again, by the length of hiring cycles, training programmes, and legal review. Slack was technically ready in 2014 and still took years to change how offices actually worked.
So when someone says “AGI by 2027,” ask which clock they mean. The answer changes the practical meaning enormously. It’s also why serious forecasters attach probabilities rather than years, which is the core argument behind why the honest answer to when AGI arrives is a range, not a date.
What Actually Moves the Date
Four inputs do almost all the work. Watch them and you can largely ignore the press releases.
Compute
Frontier training runs grew roughly four to five times a year in the early 2020s. GPT-4 was estimated in the low 10^25 FLOP range; the largest announced clusters now target an order of magnitude more. Any credible plan for an accelerated AGI timeline depends on those buildouts continuing, and that depends on power, land, and chip supply at least as much as on ideas.
Data
The public text well is close to tapped. The interesting question is what replaces it: synthetic data with verification, reinforcement learning environments, simulator-generated agent trajectories, or expensive human annotation. Which of those scales cleanly is an open empirical question, not a settled one, and the answer will move the date by years either way.
Algorithmic efficiency
Roughly every year, the compute needed to hit a fixed capability level falls by a factor of about three. That compounds. A 10x improvement over four years is bigger than most people’s intuition allows, and it’s the quiet reason capability surprises keep arriving ahead of schedule.
Money
Hundreds of billions of dollars a year now flow into data centres and chips. Capital doesn’t buy breakthroughs, but it buys the thousands of experiments that produce them, plus the compute to run ablations nobody would have funded in 2019. The shifts that genuinely reset the timeline tend to show up here first, and they’re the ones covered in these AGI industry updates that actually change the timeline.
Why the Milestones Keep Moving
Benchmarks saturate. MMLU went from a frontier challenge to a mostly-solved test in about two years. When that happens, the field doesn’t conclude that AGI has arrived. It concludes that MMLU was the wrong test and writes a new one. Handwriting recognition, machine translation, and chess all went through the same cycle: once a machine could do it, it stopped counting as intelligence.
That isn’t purely goalpost-shifting. Each new benchmark tends to be harder in a real way. But it does mean you should be sceptical when a launch announcement pairs a polished demo with a claim about proximity to AGI. What a model can do in a curated ten-minute video and what it can do across a messy Tuesday afternoon of real work are different claims, and what to watch beyond the demo reel is usually the boring stuff: failure rates, cost per completed task, and behaviour when the input is ambiguous.
Reading a Roadmap Without Getting Fooled
Every major lab now publishes some version of a plan: safety frameworks, capability thresholds, tiers of autonomy. These documents are useful, mostly as evidence of what a company is worried about rather than as schedules. A roadmap saying “we expect level 3 in 2026” is a statement of intent, not a prediction, and the incentives run toward optimism.
The tell is specificity. A roadmap that names the evaluation, the failure mode, and the conditions under which it would pause is doing real work. One that lists capability tiers next to calendar dates is closer to marketing. There’s a short checklist for reading AGI roadmap updates without getting fooled that’s worth keeping open the next time one drops.
The Competition Angle Matters More Than You’d Think
Timelines aren’t only technical. If one lab believes a competitor is six months behind, that changes its appetite for shipping something under-tested. Racing dynamics compress schedules and push safety evaluation from blocking to parallel. Whether the race is real or partly rhetorical is a separate question, but the perception of one is enough to shift behaviour, and behaviour is an input to the AGI timeline.
What the People Building It Actually Say
Long-form interviews are more informative than keynote stages, because researchers will usually name their crux. Listen for whether someone thinks the blocker is scaling, sample efficiency, continual learning, or something about world models. Those map onto very different dates. Much of the apparent disagreement between labs is less about ego than about which experiment each person would need to see before changing their mind. AGI expert interviews are a decent place to hear those bets stated plainly.
A Practical List of Signals Worth Watching
If you’d rather track the timeline yourself than wait for an announcement, these carry real information:
- Cost per unit of capability. Price for a fixed benchmark score, tracked over time. It’s objective, and it drives deployment speed more than any headline result.
- Task horizon length. The length of task an agent completes reliably at, say, 80% success. In several measurement sets it has been doubling roughly every seven months, which is a more useful line to follow than any single demo.
- Energy and chip commitments. Multi-year power purchase agreements and fab bookings tell you what labs expect to be running in 2028.
- Hiring patterns. When a lab hires reliability engineers and eval specialists instead of pure research staff, it thinks it is close to deployment.
- Published failures. An agent that runs unattended for a week and quietly corrupts a database says more about readiness than another benchmark gain.
Setting Your Own Working Date
For planning purposes you don’t need the true date. You need a defensible range and a decision that holds up across it.
A workable approach: take the researcher survey median as your slow case, the aggressive lab estimates as your fast case, and treat everything between as plausible enough to matter. Then ask what changes if the fast case is true. If a system in 2028 can do most of a junior analyst’s work reliably, what would you want in place by 2026? Cleaner documentation. Habits for evaluating output rather than admiring it. A clear sense of which parts of your job are verification rather than generation, because verification is the part that stays valuable longest.
The one thing worth avoiding is anchoring on a headline year. The AGI timeline has moved in both directions over the past three years, and the people closest to the work consistently decline to name a single number. Treat the question the way they do: a distribution you keep updating, not a date you sit and wait for.