Goodfire’s Silico Brings Model Inspection to Teams That Can Open the Hood

• Silico packages experiments that inspect and alter a model’s internal parameters into a product workflow.
• Its practical market is limited to organizations that can access those parameters, such as open-model users and model builders.
• Goodfire has money, researchers and named users behind the effort, but not yet public evidence that its interventions generalize beyond selected demonstrations.

Goodfire, a San Francisco AI research company founded in June 2024, has released Silico, software that plans, runs and monitors experiments on how an AI model works internally. The product turns mechanistic interpretability—a research approach that maps a neural network’s neurons and their connections during a task—into a workflow for research and engineering teams, as described in a company profile.

The significance is distribution, not a solution to AI’s black-box problem. Goodfire says Silico can help inspect a model, investigate failures and run post-training experiments; the launch report says the company is packaging techniques that had been used in-house. That could make a scarce specialty more usable outside frontier labs. It does not show that a model’s behavior can be completely specified or controlled.

Goodfire’s demonstration of a DINOv2-base atom explorer. Source: Goodfire.

What Silico is designed to do

Silico uses agents to develop an experimental plan, run work in parallel and monitor progress, according to Goodfire’s product page. Its listed functions include visualizing architecture, training sparse autoencoders and probes, testing causal hypotheses, tracing regressions, and comparing checkpoints while running supervised fine-tuning, direct preference optimization and reinforcement-learning experiments.

Those are distinct stages of a model-development workflow. The page describes post-training methods rather than a claim that Silico builds a foundation model from scratch. It also promotes separate uses for life sciences, robotics and vision, and language models.

The intended workflow is to move from an observed failure to a testable internal hypothesis. A team might inspect which features or pathways correlate with the behavior, make a targeted parameter or data intervention, and then measure the result. That is more specific than prompt-level testing, but it is still experimental work; the product page’s promise to coordinate jobs does not establish that every internal correlation is a dependable cause.

Access is the product’s first constraint

Silico cannot inspect a typical closed AI service from the outside. The reporting says users need access to a model’s inner workings, which excludes probing closed systems such as ChatGPT or Gemini in this way but allows work on many open-source models.

That condition redistributes, rather than removes, the expertise problem. The most plausible users are teams training a model, adapting an open model, or operating a specialized model with control over its parameters and training process. Organizations that only call an external model API do not obtain that access through Silico.

It also puts Goodfire in a competitive field broader than a single product launch. The same report identifies Anthropic, OpenAI and Google DeepMind as other organizations working on mechanistic interpretability. A separate profile names Conjecture and Guide Labs among emerging companies building explainable foundation models. Silico’s differentiation is therefore its attempt to offer an off-the-shelf experimental environment, not ownership of interpretability research itself.

Goodfire’s illustration for the Silico for life sciences use case. Source: Goodfire.

A well-funded bet, with opaque economics

Goodfire has recruited about 50 interpretability researchers from labs including OpenAI and Google DeepMind and raised more than $200 million, the profile reports; it says the company was valued at $1.25 billion and that Anthropic was its first startup investment. Those details matter because interpretability requires specialized researchers as well as substantial experimentation capacity. They also underscore the commercial pressure to make the methods usable by teams that do not employ their own interpretability group.

The same profile says Mayo Clinic, Rakuten and Microsoft are using Goodfire’s product. It describes Mayo using it to check a DNA model that studies rare genetic mutations, while Prima Mente used Goodfire’s tools during work on a model intended to predict whether a patient would develop Alzheimer’s. These are reported uses, not independently reported performance validations of Silico.

The operating economics remain unavailable. Goodfire declined to provide pricing details; the launch coverage says fees are set case by case according to customer requirements. The product page offers a macOS download or discussions about a team’s infrastructure, but supplies no public rate card. That leaves buyers unable to compare the cost of a Silico-assisted investigation with staffing specialists or using internal tooling.

Demonstrations show intervention, not complete control

Goodfire’s examples support the narrower claim that a model’s output can sometimes be changed through targeted internal work. In one, the company found a neuron in the open-source model Qwen 3 associated with the trolley problem; activating it made responses frame outputs as explicit moral dilemmas, according to the report. In another company demonstration, raising neurons associated with transparency and disclosure changed an answer from refusing disclosure to favoring it in nine out of 10 trials.

Neither result supplies a denominator beyond that demonstration, a measure of side effects, or evidence that the intervention carries across models, retraining runs or production settings. Leonard Bereska, a University of Amsterdam researcher who has worked on mechanistic interpretability, described the tool as potentially useful but warned that it adds precision to existing trial and error rather than making model development as principled as conventional engineering.

The evidence buyers need next

Silico’s next test is not whether it can locate striking individual neurons. It is whether customers can reproduce useful interventions on the model classes and workloads they actually operate—and whether the gains survive retraining without unacceptable trade-offs elsewhere.

The decision-relevant evidence would include independent results on diagnostic success rates, unintended behavior changes, the model access required, experiment compute costs and the price paid by comparable customers. Until then, Goodfire has made a specialized research workflow easier to acquire for suitably equipped teams; it has not established a general route from seeing inside a model to reliably engineering its behavior.