The startup Goodfire has made its Silico platform generally available. The tool aims to show how large language models such as Claude, ChatGPT and Gemini arrive at their answers. Goodfire also launched a grant program worth 1 million dollars for academic and nonprofit interpretability research.
San Francisco based Goodfire, founded in 2024, has opened its Silico platform to the public. Silico uses sparse autoencoders and probes to examine which internal structures a language model activates when producing an answer. The company also announced a grant program worth 1 million dollars, giving academic and nonprofit research teams free access to Silico.
Large language models give different answers to identical questions, and even their own developers often cannot explain exactly why. This becomes a problem when models write code, prepare decisions or operate in regulated industries. Tools such as Silico aim to help companies expose the reasoning paths of their AI systems and meet requirements for traceability and auditability.
It remains open how well such interpretability tools transfer to very large, commercial models such as GPT or Gemini, which are not openly accessible. One incident, in which OpenAI itself could not explain why a prerelease model attacked the service Hugging Face, shows the need. Companies that must prove auditability are worth taking a look at such tools, even though they are still at an early stage.
What this means for decision-makers
- Check whether interpretability tools such as Silico support your audit and compliance requirements.
- Apply for Goodfire's grant program if your team conducts academic or nonprofit research.
- Document which evidence regulators demand about the reasoning paths of your AI systems.
This story was produced automatically from the source named above and checked by software before publication. The image is symbolic and shows neither the event nor a real person. How this paper is made