Reaction
AI R&D compute limits as an alternative to third-party safety assessments
Another post suggests governments could check development compute use by interviewing employees and providing whistleblower protections.
TLDR
One poster speculates that METR may be seeking broader access for third-party assessment work with OpenAI, explicitly saying they have no inside information. They argue that limiting AI R&D to at most 20% of compute would be easier to verify than a company’s claim that it is acting safely. Another poster suggests governments could begin checking development compute use through employee interviews and whistleblower protections.
Combined views
1.5K
2 Sources, first seen ago
