Former OpenAI researcher Daniel Kokotajlo, who left the corporation in April 2024 due to fundamental disagreements with management regarding neural network safety issues, has presented a massive and meticulously crafted plan to prevent a potential catastrophe associated with the uncontrolled development of artificial intelligence. Together with the AI Futures Project organization, which he founded last year, Kokotajlo published a document titled AI 2040: Plan A, which serves as an optimistic alternative to his own grim AI 2027 forecast that predicted humanity's extinction or global dictatorship with a probability of up to thirty percent. The new scenario is positioned not as an inevitable prediction of the future, but as a practical recommendation containing a specific set of steps capable of significantly reducing existential risks. The key idea of the plan is the conclusion in 2029 of a historic international agreement between the United States and China to abandon the destructive race to create superintelligence, which without proper control would lead to the complete automation of AI development by 2030 and the total loss of human control over the technology.

The architecture of the proposed agreement relies on four fundamental principles: buying time for deep safety research, complete transparency of all AI developments, fair distribution of artificial intelligence among various companies and states, and the mandatory preservation of the reversibility of all technological processes. To ensure mutual trust between geopolitical rivals, the plan uses a unique verification system based on the fact that large computing clusters cannot be hidden from space satellites. Countries publicly declare all purchases of specialized AI chips, after which a temporary pause on new training runs is introduced, the observance of which is confirmed by special sensors installed directly at the data centers. The most radical element of the strategy becomes the concept of "mutually assured destruction of computing power," which the authors directly compare to the logic of nuclear deterrence during the Cold War. According to the developers' vision, new Chinese data centers should be physically built on Canadian territory, while American facilities should be located in Mongolia, that is, in places that are easiest for the opponent to attack in the event of a military conflict, creating a powerful deterrent against breaking the deal.

The economic model of the plan assumes a colossal growth in global computing power from twenty million H100 equivalents in 2026 to sixty billion by 2034, which will allow the real US GDP in certain periods to reach fifty percent annual growth versus the usual three percent. However, inevitable automation will lead to a collapse in US employment from sixty-two percent in 2027 to just twelve percent by 2040. To compensate for massive job losses, experts propose introducing large-scale "citizen dividends" – regular payments to every adult citizen from revenues that the government will collect from corporations for permits on computations and the use of robots. According to calculations, by 2032 such a dividend will amount to forty-five thousand dollars per person, by 2035 it will grow to one million, and by 2039 it will reach a fantastic ten million dollars. Specialists also contrasted Plan A with four alternative scenarios: coalition pressure on China with the risk of war, attempts at internal regulation leading to oligarchy, minimal regulation with the threat of World War III, and a complete indefinite halt to development, which will inevitably lead to the collapse of the deal in the future.