Goodfire's Silico Cuts Hallucinations 58% at 90x Lower Cost Than LLM-as-Judge. Interpretability Just Became a Builder Tool.

Mechanistic interpretability has been a research curiosity for years. Interesting papers, cool visualizations of neuron activations, very little you could actually ship. Goodfire just changed that.

Their new tool, Silico, maps neurons and pathways inside LLMs and lets developers tweak them to reduce unwanted behaviors. In testing, it cut hallucinations by 58% with roughly 90x lower cost per intervention than the standard LLM-as-judge approach. No benchmark degradation. MIT Technology Review named mechanistic interpretability one of its 10 Breakthrough Technologies of 2026.

Here's why this matters for builders, not just researchers. Right now, if you're shipping an LLM-powered product and your model hallucinates, your options are bad. You can add a second LLM call to check the first one's work (expensive, slow, still fails). You can add retrieval to ground responses (helps but doesn't eliminate the problem). Or you can fine-tune, which is a sledgehammer when you need a scalpel.

Silico offers a fourth option: look inside the model, find the pathways responsible for the hallucination pattern, and adjust them directly. It's the difference between treating symptoms and diagnosing the disease. And at 90x lower cost than LLM-as-judge, the economics actually work for production use.

The tool works on open-source models today. If you're running a product on Llama, Mistral, or any open model and hallucination is a reliability issue, Silico deserves evaluation. For teams on closed models, the principles still apply. The field is moving toward giving developers control over model behavior at the mechanistic level, not just the prompt level.

I don't know if this scales to every hallucination category or every model architecture. The 58% number is promising but I'd want to see it replicated across more domains. Still, this is the first interpretability tool I've seen that ships as a product, not a paper.

Source: MIT Technology Review