Former OpenAI researcher Daniel Kokotailo, who left the company in April 2024 due to disagreements over safety issues, has presented a new ambitious scenario for saving the world from the threat of superintelligence. The document, titled AI 2040: Plan A, proposes an international agreement that would slow down the race for artificial intelligence until 2040, reducing the risks of a catastrophic outcome.
The Essence of the Plan: A Pause Until 2040
The key idea is a deal between the US and China in 2029 to abandon the accelerated development of superintelligence. Instead, countries will develop neural networks gradually, up to the level of the best human experts. By 2035, development will be completely halted to maintain human control over the systems. Only in 2040 will the pause be lifted, and AI will reach the level of superintelligence. This provides time for safety research and prevents the automation of development, which, according to calculations, would have occurred as early as 2030.
Four Principles and a Verification Mechanism
The plan rests on four pillars: buying time for safety, full transparency of development, distribution of AI among different countries and companies, and maintaining the reversibility of the process. To ensure trust, satellite verification is proposed — large data centers are visible from space. The first step: countries publicly declare chip purchases, then a temporary pause on model training is introduced, confirmed by sensors. After trust is confirmed, restrictions are lifted, but research remains fully transparent.
Protection against a deal breakdown is provided by "mutually assured destruction of computing power" — by analogy with nuclear deterrence. It is proposed to build new Chinese data centers in Canada, and US facilities in Mongolia, so that in the event of a conflict, they would be easier to destroy than to surrender to the enemy.
Economic Consequences: GDP Growth and Civilian Dividends
Calculations show that global computing power will grow from 20 million H100 equivalents in 2026 to 60 billion by 2034. Real US GDP growth in certain periods of the 2030s could reach 50% per year, but employment will fall from 62% in 2027 to 12% by 2040. To compensate, "civilian dividends" are proposed — payments from state revenues from licensing computing and robots. The forecasts are impressive: by 2032, the dividend will be $45,000 per person, by 2035 — $1 million, and by 2039 — $10 million.
Alternative Scenarios: From War to Oligarchy
The authors contrasted Plan A with four other possible paths of development. Plan B — the US creates a coalition and pressures China, even with cyberattacks, leading to war. Plan C — an attempt to negotiate, but under pressure from companies, the pause is lifted, threatening permanent oligarchy. Plan D — minimal regulation and a race, leading to loss of control and World War III. Plan S — a complete indefinite halt, which risks collapsing and restarting the race in chaos.
My comment as an analyst: The proposed scenario looks utopian, especially regarding "mutually assured destruction" and building data centers on allied territory. However, it raises a critically important question: can humanity voluntarily abandon the race for superintelligence when the stakes are so high? The realism of the plan is questionable, but the very framing of the problem is a signal that cannot be ignored.