Goodfire launches Silico to turn AI model training into precision engineering
Silico uses autonomous agents to debug LLMs, aiming to reduce hallucinations and enhance ethical reasoning in open-source models.
Goodfire, a San Francisco-based startup, has released Silico, a new mechanistic interpretability tool designed to allow researchers and engineers to debug large language models during the training phase. The software enables users to peer inside AI models, adjust specific parameters connected to individual neurons, and automate interpretability tasks using agents.
The tool aims to transform AI model development from a trial-and-error process into precision engineering. By allowing developers to audit and modify model behaviours, Silico targets specific issues such as reducing hallucinations or enhancing ethical reasoning, particularly within open-source models. Goodfire claims this is the first off-the-shelf tool of its kind that can help developers debug all stages of the development process, from building a dataset to training a model.
Demonstrated capabilities include the ability to tweak model behaviours to reduce hallucinations. In one instance, Goodfire researchers identified a neuron associated with the trolley problem in the open-source model Qwen 3. Activating this neuron changed the model's responses, making it frame its outputs as explicit moral dilemmas. The company states that Silico allows developers to adjust parameters connected to individual neurons to boost or suppress certain behaviours.
Goodfire has also shown the ability to steer the training process by filtering out specific training data that inadvertently cause models to make errors. For example, the tool can identify when a model believes 9.11 is greater than 9.9 due to influence from religious texts or specific code repositories. Using this information, the model can be retrained to avoid these associations when performing mathematical tasks.
The tool uses agents to automate complex interpretability work, a capability Goodfire CEO Eric Ho states is now strong enough to replace human-led tasks for this specific purpose. Ho notes that the company's mission is to make building AI models less like alchemy and more like a science, removing the trial and error to expose the knobs and dials used during the training process.
Silico will be available for a fee determined on a case-by-case basis, with specific pricing details not disclosed. While the tool is designed to put techniques previously available to top labs into the hands of smaller firms, researchers note that while it adds precision to the process, the fundamental nature of the work remains complex. MIT Technology Review has picked mechanistic interpretability as one of its 10 Breakthrough Technologies of 2026.
