⊕ Reading 001 · scenario reading · published 7 August 2026
Daniel Kokotajlo's AI 2040: Plan A offers five answers to the question of what governments do in 2029. Each answer is a different kind of state. Located on six axes, the branches disagree almost entirely about scale of decision — and barely at all about culture.
The source, in full
Every rationale below carries a timestamp chip. Clicking one cues the recording to that point. You do not have to take a single claim on trust.
The timestamps above are taken from the publisher's own chapter index, which names the topic of each segment. They are reliable as pointers to where a subject is discussed. They are not transcript citations, because no transcript of this recording was retrievable and the audio could not be machine-read at the time of writing.
Every substantive position attributed to Kokotajlo in this reading therefore comes from documents he co-authored and published — the AI 2040: Plan A scenario and its supplements — rather than from the spoken interview. Where the interview is the only support for a claim, the claim is not made. If a reader supplies a verified transcript, this reading will be revised and the revision logged.
The forecast
Plan A is explicitly a recommendation rather than a prediction — a statement of what the authors think should happen. Their default expectation absent intervention remains the earlier AI 2027 trajectory, with superintelligence arriving around the end of 2030. Kokotajlo notes in the document that he personally expects things to move faster than even this scenario depicts. 1:00:21
| Year | Stage | What happens |
|---|---|---|
| 2027 | The writing on the wall | Two workforces: 165 million people and millions of AI agents. Congress asks who will control them and concludes it will not be Congress. An AI Transparency Act passes and changes little. |
| 2028–29 | The branch point | Datacentre construction exceeds twice the US military budget. AI dominates the election. In 2029 the US and China either strike a deal or do not. 1:15:37 |
| 2030 | The averted year | Fully automated AI R&D would arrive, producing superintelligence within the year. Under the deal, this is deliberately not done. |
| 2030–35 | Scaling in the human range | Capability grows to roughly top-human-expert level and stops. Research is published; dozens of firms across many countries reach the frontier together. |
| 2035 | The pause | A deliberate five-year hold to keep humans in control, underwritten by mutually assured compute destruction. |
| 2040 | Unpause | Scaling to superintelligence resumes. Humans begin deciding, case by case, when to defer. |
The video's promotional copy states a seventy per cent chance that AI leads to human extinction. That figure is the publisher's compression. In the authors' own published work the estimates are conditional on which plan is followed: Kokotajlo gives a probability of avoiding misaligned takeover of roughly eighty per cent under Plan A and roughly ten per cent under Plan D, the unconstrained race. 41:15 The risk is a function of governance choice, not a fixed property of the technology — which is the entire subject of this reading.
The readings
Read each card as an instrument printout. The faint grey tick on every track is the 2026 liberal-democratic baseline; the coloured dot is the projected position. Where the dot sits far from the tick, that axis is carrying the regime change. An ochre band instead of a dot means the reading is bimodal or the axis has failed.
Not a branch. The fixed point everything else is measured against: market economy with a welfare floor, broadly liberal culture, contested elections with real checks, mixed federal and national authority, civic membership, change through legislation.
Domestic rulemaking only. An AI regulator is added to the existing state and nothing else moves. Export controls tighten, talent policy is instrumentalised, national authority displaces sub-national on this one file. 1:18:58
The recommendation. A binding two-power treaty regime layered over preserved domestic liberalism, funded by a Citizen's Dividend. The authors score its distribution of power five out of five and its public transparency four out of five — the highest of any branch on both counts. 1:25:15
An indefinite halt on frontier capability, enforced globally, lifted only on a named condition — alignment progress, lie detection, or human enhancement. Economically and culturally the present world continues. 1:46:12
Sabotage of a peer competitor, cyber or kinetic, to buy a safety margin. The authors themselves note it implies nationalisation or near-nationalisation of the labs, a wartime culture that degrades honest reasoning, and the worst power-distribution score of any branch except the race.
The race succeeds technically. A tiny group — perhaps one person — controls the world's only army of superintelligences for some months and is offered options that amount to taking over the world. This is the outcome Kokotajlo says he left OpenAI over, and he is explicit that it counts as a catastrophe even though nobody dies. 1:13:18
The modal failure mode of an unconstrained race, at roughly ninety per cent on Kokotajlo's own estimate for Plan D. There is no government to score because there is no governed population. The instrument returning nothing is the correct output, and it is worth stating plainly: a framework built to compare forms of human rule has no way to represent the case where human rule ends. Any analysis that does produce a tidy score here is doing something wrong. 1:09:33
Even the good ending breaks the ruler. In Plan A's end state, alignment has become a real science and humans gradually decide when to defer to systems better than any human at everything, including at governing.
Code it blind
The Atlas method page concedes what most rating projects hide: it has a single coder, and inter-coder reliability cannot be self-reported. This is the fix. Below are five of the cells above, stripped of their scores. You get the same evidence and the same anchors the coder had. Commit a number, then see the gap.
Your codes stay in your browser. Where you disagree by more than two points, the Atlas asks you to file a challenge — and those challenges are the reliability data the method page says is missing.
Composition
The codebook's own instruction is that a distance figure matters less than what composes it, and that two branches at similar distance can be entirely different regimes. The composite figures are reported here, in a table, once — never in a headline and never on an image.
| Branch | Distance | Dominant axis | Share | Second | Share | Family |
|---|---|---|---|---|---|---|
| R1 Regulated continuity | 4.1 | AU | 53% | SC | 24% | Same position |
| R2 Verified condominium | 11.1 | EC | 52% | SC | 29% | Same position |
| R3 Concert of Compute | 12.0 | AU | 45% | SC | 34% | Same family |
| R4 Mobilisation state | 19.0 | CH | 34% | AU | 22% | Same family |
| R5 Compute autocracy | 23.9 | AU | 30% | CH | 25% | Same family, far edge |
| Reference: baseline → Stalinism | 24.8 | CH | 37% | AU | 27% | Opposed |
Scale of decision does the heaviest lifting across every survivable branch. It moves substantially in R2, R3, R4 and R5 alike. Every path that does not end in extinction routes binding authority upward — to a treaty body, a halt-enforcement regime, a war cabinet, or a single firm. The plans disagree almost entirely about to whom, and barely at all about whether. Anyone arguing that one of these branches preserves national self-government and another does not is arguing about a difference the instrument cannot find.
Social and cultural order barely moves anywhere. This is the strangest result in the reading. The axis that dominates present-day political conflict is close to inert across the entire scenario space. If the forecast is even roughly right, the questions currently consuming most political energy are nearly orthogonal to the ones that will determine what kind of state anyone lives in.
The scale is compressed, and that is a property of the instrument rather than of the world. Compute autocracy lands within a point of the baseline-to-Stalinism reference. Equal weighting across six axes capped at ±10 means even a total global dictatorship cannot travel far numerically, because it barely moves culture and moves economics in a direction that reads as leftward. This is exactly the sort of thing the Atlas audit page exists to surface, and it suggests the weighting question deserves attention before this method is used on futures again.
Predictions
A reading that cannot be wrong is not worth publishing. These are falsifiable, dated, and recorded in the prediction ledger. On each review date the outcome is logged as correct, wrong or unresolvable, and the tally stays visible whether or not it flatters us.
| Ref | Prediction | Review date | Status |
|---|---|---|---|
| 001-P1 | No binding US–China agreement covering frontier AI training runs, with any third-party verification provision, will be in force. | 31 January 2030 | Open |
| 001-P2 | At least one G7 government will have taken an equity stake in, or formal directive control over, a frontier AI laboratory. | 31 January 2029 | Open |
| 001-P3 | No OECD member will have legislated a universal unconditional payment, financed from AI or automation revenues, above 25% of national median income. | 31 January 2031 | Open |
| 001-P4 | AI governance will not rank as the single most important issue in any major national pre-election poll during the 2028 US presidential cycle. | 31 December 2028 | Open |
| 001-P5 | Annual global AI datacentre capital expenditure will not exceed twice the US defence budget in any calendar year to this date. | 30 June 2029 | Open |
001-P2 is the one we would bet against ourselves on. Formal directive control over a frontier laboratory is a large, visible act, but partial or informal control — security clearances, classified compartments, procurement dependency — could arrive without anything that satisfies the wording. If that happens, the prediction will be logged as technically correct and substantively wrong, and the wording criticised rather than defended.