Plan A: a serious attempt to write down how this could actually go well.

Who wrote it, and what came before

Plan A comes from Thomas Larsen, Romeo Dean, Brendan Halstead, Eli Lifland, Ryan Greenblatt, and Daniel Kokotajlo, the group whose earlier scenario, AI 2027, did more than almost any document to shape how this debate is conducted. That scenario was bleak: it forecast an intelligence explosion ending in either extinction or, as they put it, irreversible concentration of power.

Plan A is the answer to the obvious follow-up question. If that is the default, what should happen instead? Their reply, in one breath: humanity delays superintelligence until 2040, makes AI research public, lets dozens of companies worldwide catch up to the frontier, and deliberately enters a regime of mutually assured compute destruction.

Crucially, they are explicit that this is primarily a recommendation, not a prediction. It is not what they expect. It is what they would like, written as a story so that its weak points have somewhere to show themselves.

The shape of the scenario

The five-year hold at the top of the human range is the heart of it. Plan A is not a permanent stop, and it is not a race. It is a decision to spend the most dangerous stretch of the curve moving slowly, on purpose, and to use that time on alignment, epistemics, and keeping the agreement intact.

The engineering sentence

Buried in the capability-scaling supplement is the line that carries the whole plan:

This is not a slogan, it is a control law, and it is what separates Plan A from both the pause camp and the race camp. Speed is not the variable being optimized or minimized. It is the variable being bounded, by confidence in control. The plan then does the unglamorous work of asking what that bound actually looks like in hardware, treaty text, and dates.

Their term for where the bound bites is max-controllable-AI: the most capable system we remain confident could not cause a catastrophe even if it wanted to. Plan A scales up to roughly that point, stops, and waits there. It is a rare thing in this debate: a stopping rule defined by our own capacity to cope, rather than by a capability milestone or a calendar date.

The part everyone should read: verification

The plan rests on a claim its authors state plainly: this is an agreement without relying on trust. Anyone can propose a slowdown; the reason slowdowns are dismissed is that no one can tell whether the other side has actually stopped. Plan A's verification supplement is a genuine attempt to solve that, and it is the most technically interesting thing on the site.

Inference-only retrofits

The flagship mechanism is disarmingly physical: passive optical taps (essentially a piece of glass splitting optical signals) copy datacentre traffic to independent servers, while the high-bandwidth networking needed for large training runs is physically removed. The building can still serve models; it can no longer quietly train new ones.

Partial recomputation

Workloads are forced into reproducible packets, and a random sample is recomputed by auditors. You don't need to watch everything. You need the chance of being caught to be high enough that cheating is irrational. They put numbers on it: by 2034, spending 1% of compute on recomputation catches a 10,000-GPU rogue deployment within roughly four and a half hours.

Compute accounting, and the awkward numbers

Chips are tracked from fabrication; new production flows only to verified sites or guarded storage. The authors are candid about the residue: their median estimate of untracked "dark compute" is about half a percent of world supply, and they put roughly a 13% chance on a covert project reaching capabilities that could undermine the deal. That honesty is characteristic, and it is why the document is worth trusting more than a cleaner one would be.

Mutually assured compute destruction

The most arresting idea in Plan A is its enforcement mechanism, and it is deliberately borrowed from the nuclear age. Each party maintains the ability to destroy the others' post-deal compute. The logic is stated without ornament: if that capability is firmly in place, no actor will sanely attempt to pull out of the deal.

The details are startling to read in a policy document: guarded clusters with scorched earth mechanisms in case of seizure; datacentres relocated to international waters by 2035 because they are, in their phrase, much less escalatory to destroy if the agreement collapses. Whatever else it is, this is not naive. It is an attempt to make defection structurally unattractive rather than merely prohibited.

What I think it gets right

It builds the brake before the emergency. Running through the verification plan is an insistence that the machinery must exist in advance: a ready-to-go verification solution (that doesn't require much time to activate) may be incredibly important, with inspection hardware stockpiled and dormant until needed. They even price the delay: an unprepared deal could take most AI services offline for six to twelve months, a prepared one two to five. Capability arrives faster than institutions can be assembled, so the institutions have to be assembled early. This is the single most transferable lesson in the document, and it applies far below the level of treaties.

It quantifies how fast you'd find out. Rather than asserting that violations would be detected, Plan A produces curves: probability of detection within an hour, a week, a month, by size of the violation. Turning "we'd notice" into a measured latency is exactly the move this field needs more of, and it is what makes the plan arguable rather than merely admirable.

It applies scrutiny to itself. The authors argue that most AI policy proposals collapse under what they call scenario scrutiny (the discipline of writing down a detailed, plausible world in which your own proposal works) and that almost nobody applies it to their own ideas because doing so surfaces uncomfortable problems. They then apply it to theirs, in public, knowing it invites attack. Eisenhower's line sits at the top of that section: Plans are worthless, but planning is everything.

The questions it leaves open

The plan's own analysis names its greatest weakness, which they call deal decline, the risk that an agreement does not so much collapse as quietly rot: unenforced, badly executed, or renegotiated into something weaker. Much of their strategy is really a race between alignment progress and this slow decay, and their handoff calculator makes the trade explicit, weighing the alignment confidence bought by waiting against the yearly risk the deal degrades. Reasonable people will put very different numbers into that calculator and get very different answers.

Then there is time. Plan A's clock starts with a 2029 agreement, yet the team notes that Daniel Kokotajlo currently expects things to move somewhat faster than the scenario depicts. The plan requires years of preparatory work (chip tracking, verification R&D, diplomatic groundwork) before its first milestone. If capability outruns that preparation, the plan does not fail dramatically; it simply never starts. Which is, of course, the whole problem in miniature: the gap between how fast the technology moves and how fast we can build the means to govern it.

Why it matters if you're not a head of state

It would be easy to file this under geopolitics and move on. I'd suggest otherwise, because the reasoning transfers cleanly downward. Plan A's structure is: decide how fast the thing you're building can change, decide how quickly you could detect and reverse a problem, advance only while the second comfortably outruns the first, and pre-commit (in advance, in writing) what happens when it doesn't. That is a treaty at the scale of nations. It is also a deployment policy, a model-release process, or a Friday-afternoon decision about whether to ship.

This is the same principle I spent a book arguing for, arrived at independently and worked out in far more operational detail at the level of states than I attempted. Seeing it derived from an entirely different starting point (forecasting and arms control rather than control theory and case history) is, to me, the most encouraging thing about it. Good ideas about staying in control seem to converge, whoever starts the walk.

Read it. It is long, technical in places, and unusually honest about its own weak points. That last quality is rarer than the first two.

Sources: AI 2040: Plan A · Verification Plan · Capability Scaling Strategy · Deal Decline, all by the AI Futures Project.

This is an independent commentary on a publicly published document. It is not affiliated with or endorsed by the AI Futures Project; all quoted phrases are drawn from their public work, linked above, and remain the property of their authors.