Goodfire, an AI interpretability lab founded in 2024, has made its Silico platform generally available, offering AI agents that autonomously investigate model behavior using mechanistic interpretability techniques like activation mapping, weight-change tracking, and targeted edits. Users describe what they want to understand in plain language, and Silico plans and executes experiments to answer questions like why a model hallucinates. Goodfire also launched a $1 million grant program for academic and nonprofit interpretability researchers. Notable use cases include Prima Mente's discovery of a new Alzheimer's biomarker by reverse-engineering its Pleiades model, and Reciprocal Research using Silico to accelerate AI cognition studies.
Questions this post answers
What is Goodfire's Silico platform and what does it do?
Silico is a mechanistic interpretability platform from Goodfire, an AI lab founded in 2024, that was made generally available to the public. Users describe in plain language what they want to investigate about an AI model, such as why it hallucinates, and Silico autonomously builds an experimental plan and dispatches AI agents in parallel to run interpretability techniques like activation mapping, weight-change tracking, and targeted edits. daily.dev surfaces developments like this for engineers building trust into AI systems they ship.
How was a new Alzheimer's biomarker discovered using AI interpretability tools?
Researchers at Prima Mente used Goodfire's interpretability tools to reverse-engineer their Pleiades epigenetic foundation model, which detected Alzheimer's disease from blood samples without an explainable method. The analysis revealed the model was relying on DNA fragment-length patterns, a signal not previously used by humans to detect the disease, marking what Goodfire describes as the first natural science finding discovered purely by reverse-engineering a foundation model. Teams applying interpretability techniques to real research problems can follow cases like this on daily.dev.
What is mechanistic interpretability in AI models?
Mechanistic interpretability is an approach to understanding what happens inside an AI model during a task by analyzing its weights, activations, and attention patterns, and mapping neurons and the pathways between them. Techniques include matching activation patterns to human-understandable concepts, comparing weights before and after training runs, and altering specific weights or activations to observe effects on output. Developers evaluating interpretability approaches for their own models can track this space on daily.dev.