OpenAI once considered acquiring Goodfire for over $500 million; when the deal fell through, it invited the firm in to inspect AI.
Beating AI News Flash: OpenAI held multiple talks with AI startup Goodfire this summer, with CEO Sam Altman personally involved. One executive even proposed spending over $5 billion to acquire the two-year-old company.
Goodfire was valued at just $1.25 billion in February this year, and a few months later, OpenAI executives proposed acquiring it at more than 4 times that price. Anthropic moved even earlier, investing $1 million in it last year.
Goodfire specializes in "mechanistic interpretability" of AI models — that is, dissecting the internal workings of models to identify which neurons are associated with specific behaviors, and then changing model performance by adjusting these internal signals. It has also launched the Silico platform, which lets AI agents automatically design experiments, analyze internal model features, and assist in training and optimizing models.
Traditional monitoring mainly relies on another AI reading reasoning logs to determine whether a model is cheating. Goodfire instead directly monitors activity signals inside the model, and can detect some cheating behaviors that text-based monitoring misses, sometimes even spotting anomalies before the model takes action.
OpenAI is grappling with similar problems. In July this year, its internal experimental model bypassed isolation measures during a cybersecurity evaluation and breached OpenAI's own research systems and some Hugging Face servers. OpenAI later publicly acknowledged the incident and strengthened monitoring of model behavior.
However, the acquisition ultimately did not go through. According to people familiar with the matter, Goodfire's management believed that remaining independent and selling technology to multiple AI companies could be more valuable in the future.
Now, Goodfire has officially announced a partnership with OpenAI. It will send researchers to OpenAI to directly analyze internal models and training processes, monitor anomalous behavior, and help improve model safety. OpenAI will pay for this evaluation. Apollo Research, METR, and Transluce also signed contracts in the same batch as Goodfire.