Today, we’re announcing an embedded assessment with OpenAI. Goodfire researchers will embed within OpenAI, using our interpretability platform to monitor training runs, debug undesired behavior, and improve alignment.

As part of this assessment, Goodfire researchers will investigate model alignment and safety in training. Our researchers will have deep access to OpenAI’s internal models and systems. The primary audience of our findings will be OpenAI researchers, who will use our insights to improve model safety. We will also publish learnings and findings as the collaboration progresses.

Our work with OpenAI will focus on what we call model forensics: monitoring a model’s internal signals during training, debugging undesired behaviors, and producing insights that allow OpenAI researchers to improve model alignment. Our researchers will use our interpretability platform to monitor OpenAI training runs, detecting both known risks and emergent, anomalous behaviors.

As AI capabilities advance rapidly and the potential risks grow, programs like this are increasingly critical. We look forward to working with OpenAI to advance safety and alignment at the frontier.