- Who
- OpenAI
- What
- launched its first round of embedded assessments
- Quote
OpenAI will embed six Apollo Research safety evaluators in its organization as part of a pilot, Apollo said Monday.
What happened
OpenAI will host six Apollo Research safety evaluators in a pilot examining whether models try to undermine their own or successor models’ alignment and control mechanisms. Apollo says it will receive relevant training code and datasets, reinforcement-learning rollouts, and access to intermediate checkpoints, but its access to staff and information is limited. It can publish independent conclusions, while OpenAI may limit the evidence available to support them publicly.
Read next
See the full storyOpenAI Launches First Round of Embedded Assessments with Four External Research GroupsReported by 3 sources
The story so far
OpenAI has launched its first round of embedded assessments, giving four external research groups—Apollo Research, METR Evals, Transluce AI, and GoodfireAI—deep access to its internal models and systems to investigate alignment and test safeguards, according to OpenAI and co-founder Greg Brockman. In a pilot, OpenAI will host six Apollo Research safety evaluators to examine whether models covertly work against developers' intentions or undermine their own or successor models' alignment and control mechanisms, OpenAI and Apollo said. Apollo said it will receive relevant training code, datasets, reinforcement-learning rollouts, and intermediate checkpoints, but its access to staff and information is limited to agreed claims, specific models, and time periods. OpenAI also reserves the right to withhold information for technical infeasibility, security risks, or IP protection, and while Apollo can publish independent conclusions, OpenAI may limit the public evidence supporting them. GoodfireAI will embed interpretability researchers to use its platform to analyze OpenAI's model internals, OpenAI said. The company stated that GoodfireAI's work focuses on training run monitoring, debugging, and alignment to help OpenAI researchers identify and make safety improvements.
