Goodfire just launched Silico, a mechanistic interpretability tool that lets developers look inside large language models and tweak their behavior at the neuron level, according to MIT Tech Review. CEO Eric Ho told the publication in an exclusive interview that the goal is to drag AI development out of guesswork territory. “We want to remove the trial and error and turn training models into precision engineering,” Ho said.
What stands out here is the pitch. While most frontier labs chase scale, Goodfire is betting that understanding what happens inside the model matters more than throwing more compute at it.
What Silico Actually Does
The tool packages up techniques Goodfire has used internally for a while. Per MIT Tech Review, here’s what’s in the box:
- Model auditing: maps neurons and the pathways between them inside a trained LLM.
- Behavior tweaking: lets you adjust specific behaviors, like reducing hallucinations, without retraining from scratch.
- Agent-driven workflows: AI agents handle much of the interpretability grunt work that used to need human researchers.
- Training-time controls: exposes “knobs and dials” so teams can shape models during training, not just after.
Ho framed the agent piece as the unlock that made Silico shippable. “Agents are now strong enough to do a lot of the interpretability work that we were doing using humans,” he said. “That was kind of the gap that needed to be bridged before this was actually a viable platform that customers could use themselves.”
How It Fits the Landscape
Goodfire isn’t alone in this space. MIT Tech Review notes that Anthropic, OpenAI, and Google DeepMind are all pushing on mechanistic interpretability, and the publication picked it as one of its 10 Breakthrough Technologies of 2026. The difference: those labs mostly use interpretability internally. Goodfire is selling it as a product.
Ho’s framing of the broader industry is pointed. “I think the dominant feeling in every single major frontier lab today is that you just need more scale, more compute, more data, and then you get AGI and nothing else matters,” he told MIT Tech Review. “And we’re saying no, there’s a better way.”
The Pushback
Not everyone is sold on the marketing. Leonard Bereska, a researcher at the University of Amsterdam who works on mechanistic interpretability, told MIT Tech Review that Silico looks useful but the framing is oversold.
“In reality, they are adding precision to the alchemy,” Bereska said. “Calling it engineering makes it sound more principled than it is.”
That’s a fair caveat. Interpretability research is still young, and even the best tools today give partial answers about what’s happening inside a model. Silico narrows the gap. It doesn’t close it.
Why It Matters
The deployment-versus-understanding gap Ho describes is real. Companies are shipping LLMs into legal work, healthcare, customer support, and code generation while nobody can fully explain why a model produces a given output. That’s a problem for debugging, for safety, and for regulators who’ll eventually demand audit trails.
A shipping product that lets teams inspect and adjust model internals, even imperfectly, moves the conversation forward. If Silico delivers on the hallucination-reduction claims at scale, that alone justifies the launch for plenty of buyers.
The bigger question is whether interpretability becomes table stakes for enterprise AI procurement. If buyers start asking “can you explain what this model is doing,” tools like Silico stop being a nice-to-have and start being a checkbox. Goodfire is positioning early.
Full details on Silico’s capabilities and Ho’s roadmap are available in the original MIT Tech Review report.