New proposal outlines compute caps and taxes to slow AI progress
A detailed plan using layered compute restrictions and deployment taxes to safely decelerate AI development.
A new LessWrong post by Felipe Calero-Forero tackles the under-discussed question of how to slow AI progress if it becomes necessary—e.g., due to evidence of catastrophic risk, mass unemployment, or a software intelligence explosion. The author critiques common proposals like token taxes, datacenter moratoriums, and training run pauses as too blunt or harmful. Instead, the recommended approach consists of two layers. First, for catastrophic risks and rapid capability jumps, a three-tier compute restriction: a hard cap on total R&D compute at a threshold, a progressive tax on compute below that threshold, and a cap on individual training run compute as a backstop. This layered system is designed to move slowly through critical capability windows while avoiding evasion. Second, for mass unemployment, a capability-gated tax on inference (deployment) would be tied to the intensity of AI system use, slowing workforce displacement. The author emphasizes that these instruments should be dynamic, updatable with new evidence, and paired with appropriate trigger mechanisms. The post acknowledges that international cooperation (e.g., with China) may be needed but argues that a widening US lead or shifting Overton window could make unilateral action viable.
- Proposes three-tier compute curbs: hard cap on R&D compute, progressive tax below that, and individual training run caps to prevent evasion.
- Suggests a separate capability-gated tax on AI deployment to slow mass unemployment by metering inference.
- Emphasizes dynamic, conditional instruments that can be tuned as evidence emerges, rather than fixed moratoriums.
Why It Matters
Offers concrete, layered policy levers for governments to safely slow AI if catastrophic risks or social instability emerge.