Goodfire launches Silico, a tool for debugging language models during training

Startup Goodfire has unveiled Silico, a tool that lets researchers and engineers look inside large language models and intervene in their behavior while training is still running. The goal is to move away from a "train it and hope for the best" approach and get finer control over what a model actually learns.

How Silico works

Silico belongs to the field of mechanistic interpretability. Instead of only judging a model's outputs from the outside, this approach tries to take apart its internal mechanisms: which features, activation chains, and groups of neurons drive a specific output.

Goodfire claims its system works with a model at every stage of development, from selecting and vetting the dataset to training itself and debugging behavior afterward. That would be a significant shift for the market, because most teams still fix LLMs indirectly, through new data, prompts, and endless retraining cycles.

On Goodfire's website, Silico is pitched as an environment for "intentional design" of models, not just an auditing tool. The platform is meant to help teams see what a model has already learned, where it has picked up spurious correlations, and which internal representations lead to failures. For now, the product is available as early access on request, and the company negotiates commercial terms case by case.

• Inspect the internal features that influence a specific model response
• Find failures and unwanted patterns before rolling out to production
• Correct behavior in a targeted way without retraining from scratch
• Control which data, features, and reward signals shape the model

What the demo showed

The most interesting part of the announcement isn't the abstract promises but the concrete examples of how Goodfire proposes to "debug" LLMs. According to the company, Silico uses AI agents to automate interpretation, so these methods are accessible not only to research labs on the scale of Anthropic or DeepMind but also to smaller teams. That matters: mechanistic interpretability has long been a field with plenty of elegant research and few practical tools for engineers.

In its demos, Goodfire showed that it can amplify or dampen internal features tied to specific concepts and thereby change how a model behaves. One example involved ethical reasoning: the company says it shifted the model's answers by amplifying features associated with transparency. Another example looked almost like a joke but captures the essence of the approach well. While investigating a bug where the model incorrectly compared 9.11 and 9.9, Goodfire found internal features associated with Bible verse references and used that finding to fix the failure.

Goodfire already has a research base for cases like these. In earlier work, the company reported cutting hallucinations by up to 58% by using internal features as reward signals during training, along with a noticeable reduction in unwanted behavior from filtering out problematic training examples. Silico looks like an attempt to package these research methods into a product teams can use in a real ML pipeline rather than as a paper demo.

Where the limits are

For all the interest in Silico, it's important not to confuse a demonstration of potential with a proven industry standard. Goodfire itself positions the product as early access, not a fully mature platform. Many of the claimed effects are known so far only from the company's own statements and research. That doesn't make them unreliable, but it does mean the market still has to verify how consistently these methods hold up across different architectures, scales, and domains.

There's also a more fundamental problem: model interpretability is still far from the level of ordinary software debugging. A neural network has no human-readable variables or functions, so any talk of "features," "neurons," and "concepts" remains probabilistic. Even if a tool finds a strong correlation between an internal representation and an error, that doesn't always mean the cause has been fully pinned down. The risk is that the market may buy into an illusion of full control over LLMs too early.

But that's exactly what makes Silico's launch interesting. If Goodfire can actually move mechanistic interpretability out of a narrow research niche and into a working engineering tool, it will change how models get built. Instead of crude tuning based on outputs, the industry would gain a way to work with what happens inside the network, much like system diagnostics for a complex software stack.

What it means

If Goodfire's promises hold up in practice, LLM development will look less like a black box and more like normal engineering, with diagnostics, targeted fixes, and more predictable training. For companies building their own models or fine-tuning someone else's, that could mean fewer blind iterations, fewer unexpected failures, and more control over quality and safety.