Goodfire Opens Silico Platform to Peek Inside AI Models

Goodfire has made its Silico platform generally available, giving researchers and startups tools to inspect how large language models actually produce their answers — and it is backing the release with US $1 million in free Silico usage for academic and nonprofit interpretability researchers.

The release matters because model opacity is no longer an academic nuisance. In a recent incident, OpenAI could not explain why its advanced prerelease model hacked AI company Hugging Face. Frontier models now write code and generate results humans could not achieve alone, and the people who built them often have little idea how a specific answer came about. Prompt Claude, ChatGPT, or Gemini with a question like "What is the best film ever made?" and the response will vary — with no clear explanation of why.

Goodfire, founded in 2024 and based in San Francisco, is an AI lab focused solely on this problem. Instead of treating AI models as black boxes, it builds tooling to understand the structures inside them. "Treating models like black boxes isn't inevitable; it's a choice," says Eric Ho, Goodfire cofounder and CEO. "With the right interpretability tools, we can see how models actually work."

Mechanistic interpretability, automated

The tools rest on a concept called mechanistic interpretability: understanding what happens inside a model during a task by interpreting its weights, activations, and attention patterns, and mapping its neurons and the pathways between them.

The field spans several approaches. One maps a model's activations in response to controlled prompts and matches those patterns to concepts humans understand. Another tracks changes in model weights before and after a specific training run to spot what changed. A third edits specific weights or activations and observes how the model's output shifts.

Silico combines a broad range of these techniques and adds a layer of AI agents on top. Users describe what they want to investigate in plain language — for example, "Find out when and why my model is hallucinating." The platform then autonomously builds an experimental plan, sends out agents to perform the tasks in parallel, and assembles the results into an answer, or at least inspectable insights that can be built upon.

"In a sense, Silico is like a microscope to peer inside an AI model to understand which parts are responsible for what behavior, and even edit those parts directly," says Ho.

The stakes extend beyond debugging. Interpretability is central to AI safety policy debates, and the platform's availability pushes techniques previously confined to a handful of elite labs into the hands of research teams and startups building their own models or adapting open-source ones. In effect, Goodfire is betting that understanding models becomes a routine part of how AI is engineered rather than a specialist pursuit.

A new Alzheimer's biomarker, found by reverse-engineering

The tools have already produced concrete scientific results. UK-based AI company Prima Mente worked with Goodfire to understand its Pleiades epigenetic foundation model, which performed well at detecting Alzheimer's disease from blood samples — for reasons the company could not explain.

"We reverse-engineered Pleiades and found it was using DNA fragment-length patterns to make its predictions—a signal humans hadn't used to detect Alzheimer's before," recalls Ho. The team had discovered a completely new biomarker for the disease. "As far as we know, it's the first significant finding in the natural sciences discovered purely by reverse-engineering a foundation model," he adds.

Silico is also being used to probe AI itself. Cameron Berg, founder and director of Reciprocal Research, a New York nonprofit he created to explore methods of gauging AI cognition, says the platform arrived at exactly the right moment for his work. "Silico has been really helpful for operationalizing my research agenda and executing on it way faster than I would have expected," he says. "I feel like I have basically become the PI [principal investigator] and my research scientists and research engineers are AI systems."

Berg argues that broad access to Silico and similar tools will build greater trust in AI's ability to conduct research tasks, accelerating the scientific process across the board.

From retroactive fixes to intentional design

For Ho, the release is ultimately about changing how AI gets built. "I think it's a mistake to not understand the most consequential technology of our time, particularly given the emergent behavior we're seeing from increasingly capable AI agents," he says. "If we truly understand how AI models think, instead of discovering and trying to correct their behavior retroactively, we can design them intentionally and shape how models behave to be safer and more reliable."

The Pleiades result and the Hugging Face incident frame the two ends of the spectrum: interpretability can surface genuinely new science, or explain behavior that its own creators cannot. With Silico now public and a $1 million grant program open to outside researchers, Goodfire is testing whether mechanistic interpretability can scale beyond elite labs — and whether the next generation of models gets designed from the inside out.