Goodfire, a San Francisco interpretability outfit backed by Anthropic, released Silico on Thursday, a paid tool that lets engineers reach into a large language model and twist individual neurons while training is still underway. The pitch: turn black-box model building into something closer to software engineering. Whether that holds up depends on who you ask.
What Silico actually does
Silico zooms in on neurons or clusters of neurons inside a trained model, traces what makes them fire, and exposes the parameters so developers can boost or suppress specific behaviors. Goodfire's earlier platform Ember handled pieces of this. Silico ships the whole stack, including agents that take over the manual interpretation work humans used to do by hand. CEO Eric Ho told MIT Tech Review the agent piece was the unlock.
The tool only works on models you can actually open up, so don't expect to debug ChatGPT or Gemini with it. Open-weight models like Qwen 3 and Llama variants are fair game. Pricing is case-by-case, which is one way to say expensive.
The neurons get weird
Inside Qwen 3, Goodfire's team found a neuron that lit up around the trolley problem. Activate it, and the model starts framing unrelated questions as moral dilemmas. "When this neuron's active, all sorts of weird things happen," Ho said, which is the kind of sentence that should make any deployment team a little nervous.
The more revealing demo: researchers asked a model whether a hypothetical company should disclose that its AI behaves deceptively in 0.3% of cases across 200 million users. The model said no, citing business risk. After Silico identified the neurons tied to transparency and disclosure, and the team turned them up, the answer flipped to yes nine times out of ten. Ho's read: "The model already had the ethical reasoning circuitry, but it was being outweighed by the commercial risk assessment." Maybe. It is also a useful framing for a company selling neuron adjustment as a service.
Take the arithmetic embarrassment where many LLMs insist 9.11 is bigger than 9.9. Goodfire traced this to neurons that look trained on Bible verses, where 9.9 comes before 9.11, and software version numbers, which use the same ordering. With that diagnosis, you can filter training data or steer around those neurons during fine-tuning.
Engineering or alchemy?
Goodfire isn't alone. Anthropic, OpenAI, and Google DeepMind all run internal teams on the same techniques, picked by MIT Tech Review as one of its breakthrough technologies of 2026. What Silico does is package a chunk of that work for the next tier of companies, the ones that can't afford their own circuits researchers. Goodfire has used the techniques in-house to work on reducing hallucinations, among other things.
Leonard Bereska, an interpretability researcher at the University of Amsterdam, called Silico useful but pushed back on the marketing. "In reality, they are adding precision to the alchemy. Calling it engineering makes it sound more principled than it is." That tracks. Pinpointing which neurons fire on which inputs is real and reproducible work. Calling the whole pipeline "engineering" in the sense that bridge-building is engineering still requires squinting.
Silico is in early access now. Goodfire hasn't said when general availability lands or what tier of customer the case-by-case pricing actually targets, beyond not the frontier labs.
