OpenAI Allows Third-Party Groups to Test Models During Training
OpenAI will now permit external groups to conduct technical safety assessments while its models are being trained and evaluated, rather than only before launch. This comes after Sam Altman's previous announcement offering evaluators access to desks, badges, laptops, and publication rights. However, the latest update from OpenAI does not name any partners or set specific access terms.
According to Bloomberg, OpenAI is in talks with groups such as METR and Redwood Research for these evaluations. The company emphasizes independence mechanisms, scientific rigor, security practices, and clear responsibilities as its key priorities.
This move contrasts with Anthropic's approach, which named Accenture as its evaluation partner and secured a significant financial commitment from them. In Europe, compliance assessments are conducted by bodies with no commercial interest in the models they evaluate, following standardized protocols and regulations like Article 55.
Dario Amodei has requested an antitrust waiver from Washington to enable rivals to coordinate on safety measures, reflecting the ongoing discussions around model safety and regulation.