The moment an LLM starts hallucinating, every engineer hits the same wall: you can test it, measure it, deploy it—but you cannot touch the gears. The parameters driving behavior stay opaque, even to the teams that trained them. Goodfire's new tool Silico changes that equation.
The San Francisco startup just released what it claims is the first off-the-shelf mechanistic interpretability product: a debugging suite that lets developers peer inside a language model and adjust parameters during training itself. This shifts interpretability from research paper into production toolkit—a transition the field has talked about for years but never shipped.
The core problem Goodfire targets is familiar to anyone who's shipped an LLM application. When a model misbehaves—hallucinates facts, refuses edge cases, exhibits subtle bias—you attack it with more training runs, more red-teaming, more prompt engineering. It's expensive, it's slow, and you never really know if you fixed the root cause or just patched the symptom.
Silico replaces guesswork with inspection. The tool maps which neurons and pathways activate during specific behaviors, then lets developers tweak those circuits directly. Goodfire has already used these techniques internally to reduce hallucination rates in existing models—now it's packaging that capability for external customers. Ho frames the shift as moving from "alchemy to engineering": "We want to remove the trial and error and turn training models into precision engineering. That means exposing the knobs and dials so you can actually use them during the training process."
Technical credibility comes from Goodfire's approach to scale. Mechanistic interpretability has historically required painstaking human analysis—researchers would spend weeks tracing a single behavior through millions of parameters. Silico uses AI agents to automate much of that work. "Agents are now strong enough to do a lot of the interpretability work that we were doing using humans," Ho explains. "That was the gap that needed to be bridged before this was a viable platform customers could use themselves."
The product launches Goodfire into rare company. Mechanistic interpretability has been a priority at frontier labs—Anthropic published extensively on circuit-level analysis, OpenAI and Google DeepMind run dedicated teams—but no one has commercialized it. Silico potentially democratizes capabilities that have been restricted to organizations with thousands of GPUs and dedicated research staff.
Leonard Bereska, a researcher at the University of Amsterdam who works in the space, sees value in the tool but pushes back on the framing. "In reality, they are adding precision to the alchemy," he says. "Calling it engineering makes it sound more mature than it is." That's a fair caution: the field still grapples with fundamental questions about what it means to "understand" a neural network.
Still, the commercial angle changes the stakes. Frontier labs can afford to fund interpretability research indefinitely. Startups, safety organizations, and application developers have been locked out entirely. If Silico delivers on its pitch, the distributed ecosystem finally gets a seat at the table where model behavior gets decided. The question is whether "precision alchemy" is enough to catch the failures that matter before they ship.