Washington Hones in on AI Safety, What Are AI Doomers Proposing

The concrete proposal from the "AI doomer" camp is not to stop artificial intelligence entirely, but to slow it down deliberately—reaching superintelligence by 2040 instead of 2027 or 2028. The plan, outlined in a document called AI 2040, calls for highly controlled, government-supervised development rather than a ban. As one observer put it, the goal is to "keep the inference flowing" with current models while dramatically slowing frontier training runs.

The Mechanism: Compute Caps and Physical Controls

The central valve for slowing AI capabilities is compute. The AI 2040 proposal focuses on controlling the hardware rather than the software, because chips and data centers are easier to track than algorithms. The first mechanism is a pause on new frontier training runs, enforced by requiring any facility with more than 10,000 H100-equivalent GPUs (roughly $100 million in equipment) to apply for a permit and submit to independent auditing. The proposal calls for countries to declare their AI compute inventories to the international community, similar to nuclear warhead declarations.