AI Safety Advocates Push to Cap Compute for R&D

A group of AI safety researchers and advocates is urging AI labs to cap the share of compute devoted to research and development. They propose committing minimum shares of compute to safety and monitoring, and to customer inference, limiting resources for training stronger models. The idea comes as Sam Altman floated scenarios including a global AI regulator modeled on the IAEA.

• Proposal: minimum compute shares for safety and inference, remainder for R&D
• Capping R&D compute would slow development of more capable models
• Altman previously outlined three scenarios for pacing AI, including a global regulator