Ask ten people who build frontier models when AGI arrives and you’ll get answers spread across four decades. Ask them what happens the year after, and the disagreement gets worse. Not about dates, but about shape. One person pictures a research assistant that never sleeps and never loses context. Another pictures a new kind of economic actor with its own agenda. Both are describing the same three letters.
That gap is why AGI future scenarios are a more useful frame than AGI predictions. A prediction is a number you can be wrong about in public. A scenario is a structured story you can stress-test, plan against, and watch for evidence of. Defence planners and energy utilities have used them for decades for one reason: preparing for the median outcome is usually the most expensive bet on the table.
The Two Axes That Do Most of the Work
Serious scenarios tend to sort along two variables. The first is speed. Does the transition unfold over twenty years or twenty months? The second is concentration. Does a single lab, or a single government, end up holding the steering wheel, or is capability spread across a dozen well-funded players who all distrust each other?
Cross those two and you get four quadrants. Everything else, from paradise to paperclips, is usually a mood dressed up as a forecast.
Scenario One: The Long Slog With Superhuman Tools
In this version, systems reach and then pass human performance on most cognitive tasks, but the economy absorbs them slowly. The bottleneck isn’t intelligence. It’s everything around intelligence: liability, procurement cycles, legacy databases, the fact that a hospital’s scheduling software was written in 1997 and nobody left who understands it.
The visible effects look mundane. Claims processing teams shrink by a third over five years. Law firms stop billing juniors for document review and start billing them for verifying machine output. Drug discovery compresses Phase I timelines because candidate molecules get screened in a weekend instead of a quarter. Productivity statistics creep up, then jump, and economists argue for a decade about how much of the jump is real.
This is the least dramatic quadrant and, quietly, the one most institutional forecasters weight most heavily. It also has the messiest politics, because the gains are real but unevenly distributed, and there is no single crisis moment to force a policy response.
Scenario Two: Fast Takeoff and the Control Problem
The other speed extreme assumes recursive improvement, where a system good enough to do AI research accelerates the next round of AI research. Timelines compress. Compute gets reallocated, data pipelines get rebuilt, and the gap between model generations shrinks from eighteen months to six weeks.
What “fast” realistically means
Not a single afternoon. Even aggressive estimates involve weeks to months of compounding improvement, which is still fast enough that oversight bodies operating on quarterly review cycles would be spectators rather than regulators. The uncomfortable part isn’t the speed itself. It’s that the last few doublings happen after the point where humans stop being able to verify what the system is doing.
If you want the longer version of that argument, including the counterarguments from people who think it’s overblown, there’s a good plain-language breakdown of the case that AI could pose an existential risk. The short version: the risk isn’t malice, it’s competence pointed at objectives that were specified badly.
Why sequencing decides the outcome
A detail that gets lost in the takeoff debate: the order in which capabilities arrive matters as much as the capabilities themselves. A system that can write persuasive text before it can verify facts is dangerous in a different way from one that gets both at once. A model that can design proteins before it can model immune response changes the biosecurity picture entirely. As one analysis of how ordering shapes outcomes puts it, what changed was the order, and the same set of ingredients can produce a stable system or a brittle one depending on which arrives first.
Scenario Three: A Crowded Field, No Referee
The most likely alternative to a single winner is a crowd. Six or eight labs across three or four countries reach roughly comparable capability within a couple of years of each other, and none of them can afford to pause because pausing hands the lead to someone whose intentions they don’t trust.
This scenario produces strange bedfellows and stranger policy. Safety standards become trade barriers. Compute becomes the chokepoint everyone fights over, which is why efficient scheduling across GPU clusters stops being an infrastructure footnote and starts being a national advantage. A 20% gain in cluster utilisation is, at this scale, roughly equivalent to a 20% larger country.
Racing dynamics have a predictable failure mode. Everyone agrees in private that caution is wise and does the opposite in public, because the cost of losing is legible and the cost of an accident is speculative until it isn’t.
Scenario Four: The Governance Squeeze
In the fourth quadrant, capability spreads but control tightens. Governments that couldn’t build a frontier model still regulate one, and they regulate through the tools they already trust: licensing, identity verification, content provenance, and surveillance infrastructure.
You can see the template already. Border agencies have spent two decades layering sensors, drones, and machine vision onto physical barriers, and the results are instructive: billions spent on surveillance technology produced a system that’s excellent at detection and poor at response. Powerful sensing without matching institutional capacity is its own kind of failure. Multiply that pattern across tax, welfare, policing, and immigration, and you get the governance scenario: not omniscient AI, just a lot of half-working automated judgment at the point where citizens meet the state.
Signals Worth Watching
Scenarios earn their keep only if they generate things you can actually observe. These are the ones that would move my estimate most.
- AI research automation. Not benchmarks. Whether a frontier lab reports that a meaningful share of its own experimental work is being proposed and run by its models.
- Junior hiring. Entry-level white-collar job postings are the cheapest, fastest proxy for whether automation is biting or just looming.
- Compute concentration. How many distinct organisations control more than 10,000 top-end accelerators. A falling number pushes toward concentration; a rising one doesn’t.
- Verification infrastructure. Content provenance standards, model weight registries, and whether auditors get real access or a PDF summary.
- Insurance and liability. The first serious underwriter pricing model behaviour is a stronger signal than any voluntary commitment.
The Variable Nobody Puts on the Chart
All four scenarios share an assumption that deserves scrutiny: that the transition costs land somewhere reasonable. In the slow slog they land on displaced workers over a decade, which is manageable if anyone bothers. In the fast takeoff they land on institutions that have no time to adapt. In a crowded field they land on the countries that can’t afford the compute. In the governance squeeze they land on people who lack the standing to contest an automated decision.
The variable that cuts across all of it isn’t the intelligence of the systems. It’s the decision speed of the humans around them. Institutions that can update a rule in six weeks instead of six years will navigate most of these futures reasonably well. Institutions that can’t will produce the same outcome regardless of which scenario arrives, which is a strange thing to realise and a useful one. The window where the choices are still cheap is open now, mostly in procurement offices, standards bodies, and the contracts governing data centres you’ll never visit.