Goodfire ships Silico — mechanistic interpretability leaves the frontier labs and enters the enterprise debugger market
Goodfire released Silico, the first off-the-shelf mechanistic interpretability tool aimed at engineering teams rather than research groups. The product lets developers inspect individual neurons in a trained LLM, trace upstream/downstream pathways, run feature-firing experiments, and edit behavior during training — capabilities previously confined to internal tooling at Anthropic, OpenAI, and DeepMind.
The interesting piece is the buyer profile. Silico's pitch is not aimed at the four labs that already have in-house SAE pipelines; it is aimed at the much larger cohort of teams fine-tuning open-weights models or building proprietary adaptations on top of them. Goodfire's framing — "build AI models the way you write software" — explicitly positions interpretability as a debugger workflow rather than a research artifact. That is the productization step the field has been waiting on since Anthropic published Towards Monosemanticity in 2023.
The capability surface is concrete: pick a neuron or feature group inside a trained model, fire test inputs against it, walk the pathway upstream and downstream to see what drives activation and what activation drives. Silico can also filter training data to prevent unwanted features from forming in the first place — a steering vector applied at the dataset layer rather than the activation layer. Goodfire claims it has already used the underlying techniques to reduce hallucinations in deployed LLMs, which is the production receipt enterprise buyers will want before they sign.
Pricing is case-by-case, which is consistent with where the market is — Silico's first customers are likely to be regulated-industry teams (finance, healthcare, defense) and frontier-adjacent labs that don't have the headcount to build their own SAE infrastructure. The $50M Series A Goodfire raised in 2025 funded exactly this commercialization phase, and Silico is the deliverable. The pattern matches what happened with evals tooling between 2023 and 2025: a research practice becomes a product the moment a regulator could plausibly ask whether you used one.
MIT Technology Review — This startup's new mechanistic interpretability tool lets you debug LLMs → · Goodfire — Silico — Build AI models the way you write software → · PR Newswire — Goodfire Raises $50M Series A to Advance AI Interpretability Research →