New Platform Peers Inside AI’s Black Box

Goodfire, an AI lab founded in 2024 and based in San Francisco, has made its Silico platform generally available to the public, according to IEEE Spectrum. The company also announced a grant program offering US $1 million in free Silico usage for academic and nonprofit interpretability researchers.

Silico is built around mechanistic interpretability, which aims to understand what happens inside an AI model during a task by interpreting the model's weights, activations, and attention patterns and mapping its neurons and the pathways between them. The platform combines a range of interpretability tools and adds a layer of AI agents. Users describe in plain language what they want to investigate, such as finding out when and why a model is hallucinating; the platform then builds an experimental plan and sends agents to perform tasks in parallel.

Goodfire cofounder and CEO Eric Ho said treating models as black boxes is a choice rather than an inevitability. He described Silico as akin to a microscope for peering inside a model to understand which parts are responsible for what behavior, and said parts can be edited directly.

In one application, U.K.-based AI company Prima Mente worked with Goodfire to understand its Pleiades epigenetic foundation model, which detected Alzheimer's disease from blood samples but without a known reason. Ho said the team reverse-engineered Pleiades and found it was using DNA fragment-length patterns for its predictions, a signal humans had not used to detect Alzheimer's before. Ho added that, as far as Goodfire knows, it is the first significant natural-sciences finding discovered purely by reverse-engineering a foundation model.

Cameron Berg, founder and director of the New York nonprofit Reciprocal Research, said Silico helped him operationalize and execute his research agenda faster than expected, and that he feels like a principal investigator whose research scientists and engineers are AI systems. Berg sees general access to such tools leading to greater trust in AI's ability to conduct research tasks, which he said will accelerate the scientific process.

Ho said not understanding the most consequential technology of our time is a mistake, particularly given emergent behavior from increasingly capable AI agents, and that understanding how models think would allow intentional design of safer, more reliable behavior rather than retroactive correction.