Chief executives Dario Amodei and Sam Altman are pushing for slower AI development to avert disastrous consequences, but internal tensions are emerging over implementation. Amodei published a detailed essay calling for independent model evaluations, global coordination between labs and governments, and legal protection from US competition rules. He proposed external testers should have 'employee-like access' including badges, laptops, and internal tools.
While many employees agree slowing down is necessary, staff fear practical implementation could jeopardize their work and compromise security. A particular issue is embedding outside evaluators within labs, which some describe as a source of 'conflict'. OpenAI stated it wants a more pragmatic approach and has already paused certain frontier training, while Anthropic has not put its research on hold.
Implementing increased third-party access while balancing security and data protection poses significant challenges, especially given AI labs' strict security measures and intense industry competition.
Source: Financial Times Companies · Summarized by HeadlinesBriefing