The AI 2040 Scenario (Plan A — The Deal)

"Plan A is primarily a recommendation, not a prediction."

That line appears early and governs everything after it. The deal — the 2029 agreement, the transparency regime, the pause — is what the authors want. Everything downstream of it, the employment numbers and growth rates and political fights, is what they predict would follow if the world did as they ask. Two kinds of claim in the same prose. Their reason for the form: most policy proposals fall apart the moment someone writes down a detailed world in which they succeed, and they have run that test on their own proposal in public. Where the story strikes you as too convenient, you are standing where they invited you to push.

"There is one job the AI companies want to automate more than any other—their own."

The 2027 and 2028 chapters are a wake-up sequence: white-collar work churning, Congress holding hearings, an election fought over AI. The engine underneath is this one capability. An AI that can do AI research turns capability into speed of capability, and the whole plan is timed off that moment — in the scenario's default world it arrives in 2030 and produces superintelligence within the year. Hold that loosely. It is the load-bearing empirical bet, and nearly every disagreement with this document traces back to it.

"Fortunately, training AIs requires large numbers of AI chips. Most AI chips are in giant datacenters."

You cannot verify an intention, but you can count buildings. That is the hinge of the 2029 sections: frontier training needs vast quantities of chips, chips come from a handful of fabs, and datacenters are visible from orbit and unmissable on the power grid. So the deal starts with declarations and inspections, then a training halt policed by devices checking that datacenters only run existing models. Watch the pattern — every enforcement move attaches to something physical. By 2034 the same logic yields Mutually Assured Compute Destruction: each power puts its new datacenters where the other could seize them, so breaking the deal means losing your own compute first.

"we'll agree to let each other see all the AI research."

Four principles carry the 2030 chapter — Buy Time, Total Research Transparency, Diffuse AI Broadly, Reversibility — and the second is the radical one. Research becomes public; only inference stays private. Notice how much work that does: rivals and auditors catch what overstretched regulators would miss, nobody can quietly train a hidden loyalty into a model, and the prize for racing to a secret algorithm disappears. Reversibility is the principle most likely to feel backwards. The plan prefers progress from more compute over progress from better algorithms, because a datacenter can be switched off and an idea, once published, can never be recalled.

"Although it's supposed to be a slowdown, it doesn't feel like one."

The 2031–2034 stretch is the part readers underestimate. Growth runs near fifty percent a year, robots and compute are rationed by tradable permits, permit revenue funds a Citizen's Dividend, and the dashboard shows employment sliding from the low sixties toward the thirties as median income climbs into the millions. Two threads reward attention. Safety practice splits into alignment (make the system want what you want) and control (make it unable to do harm even if it doesn't), and through these years the plan leans almost entirely on control. And a footnote in 2034 concedes that these AIs are adversarially misaligned and merely contained. The good ending, at this stage, is not a story about trustworthy machines.

"Ultimately, control only works up to a point, and that point is probably somewhere around top human expert level."

The 2035 pause is derived, not chosen for drama, and the analogy that derives it is the best paragraph in the document: an orphaned eight-year-old who has inherited a business empire must supervise the lawyers and executives running it. When the staff start accusing each other, the child cannot judge who is lying. Containment works while you can still evaluate the thing you are containing. That is why the ceiling lands at top-human-expert level, and why the pause holds even as alignment research starts producing results.

"Scaling beyond human level will require relying on alignment based safety cases—arguments that AIs have robustly internalized the values they are supposed to have."

By 2038 alignment has matured into an actual science: theory that predicts, protocols for training honesty, interpretability tools that check the theory against what a model does internally. That is a higher bar than the phrase usually carries — not better behavior, but understood mechanisms and arguments that survive adversarial review. The 2040 unpause rests on a chain of deference, each generation of AI vouched for by the one before, back to the systems whose safety cases humans could read. The authors leave the nervousness of that moment on the page rather than smoothing it away.

"The scenario's deep-dive boxes, collected here as lettered appendices."

On a first pass you can skip all of them. On a second pass, use them surgically: wherever the story provoked an objection, there is usually an appendix answering that exact objection — why China would sign, why the world doesn't simply stop building AI altogether, why alignment progress still isn't enough to relax control. The alternate-timeline branches (a covert program, a flawed safety case approved, deal dissolution) are the failure modes; reading one shows how much weight the main story carries.