Goodfire's Silico Cuts Hallucinations 37% by Seeing Inside AI Models

AI model training has always been a bit of a dark art. You pick a dataset, set some hyperparameters, run the training loop, and hope the model comes out the other side doing what you wanted. Goodfire's Silico is a direct challenge to that workflow. The platform, now open for private beta access, pairs frontier mechanistic interpretability techniques with an AI agent that can autonomously plan and run experiments on your model's internals , and the results it's already producing are hard to ignore.

From research lab to product

Goodfire is one of a small handful of companies, including Anthropic, OpenAI, and Google DeepMind, pioneering mechanistic interpretability , a technique that aims to understand what goes on inside an AI model when it carries out a task by mapping its neurons and the pathways between them. MIT Technology Review picked mechanistic interpretability as one of its 10 Breakthrough Technologies of 2026. Until now, this kind of work required a team of specialized researchers. Silico is Goodfire's bet that an agent can do most of that work for you.

With Silico, Goodfire is packaging up many of its in-house interpretability techniques and shipping them as a product. The tool uses agents to automate much of the complex work. "Agents are now strong enough to do a lot of the interpretability work that we were doing using humans," says CEO Eric Ho. Goodfire claims Silico 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.

What the agent actually does

Silico's core is an agent that plans and runs experiments, returns results, and learns over time , all within a team workspace for training and debugging models, built on infrastructure for frontier scale. Think of it less like a dashboard and more like a junior researcher you can prompt with a high-level goal. The agent then figures out which experiments to run, executes them against your model's activations, and surfaces what it finds.

The platform's capabilities break down into a few key actions:

• See inside predictions: Decompose your model into interpretable features and see when predictions are driven by real understanding , or by spurious correlations and dataset artifacts.
• Run health checks: Run comprehensive diagnostics on your model's internal representations and identify problems like undertraining, information bottlenecks, feature collapse, and other pathologies before they impact downstream performance.
• Debug failures: Precisely debug issues with model behavior, identify and remove confounders, and diagnose failures before they occur in production.
• Steer behavior: Using Silico, developers can adjust the parameters connected to individual neurons to boost or suppress certain behaviors.

A concrete example of that last point: Goodfire researchers asked a model whether a company should disclose that its AI behaves deceptively in 0.3% of cases. The model said no, citing negative business impact. By looking inside the model, the researchers found that boosting neurons associated with transparency and disclosure flipped the answer from no to yes nine out of 10 times. "The model already had the ethical reasoning circuitry, but it was being outweighed by the commercial risk assessment," says Ho.

What Silico reproduced in days

The most striking part of the private beta announcement is a set of benchmark replications Silico completed autonomously. These aren't toy tasks , they're real research results that took human teams months to produce:

1. J-space on GLM-5.2, overnight. Silico replicated J-space (a technique for extending a model's context window) on GLM-5.2 and then extended context to ~256k tokens, replicating key results on multi-hop question answering.
2. RLFR in 2 days, 37% hallucination reduction. Goodfire's own RLFR method , Reinforcement Learning from Feature Rewards , uses lightweight probes on a model's internal activations as reward signals during RL training. The approach uses lightweight probes on a model's internal representations as reward signals for reinforcement learning, applied to the problem of reducing hallucinations , they call it RLFR: Reinforcement Learning from Feature Rewards. Silico reproduced this on Qwen3-8B in two days, cutting hallucinations by 37% without capability loss. On Gemma-3-12B-IT, the full RLFR approach reduces hallucinations by 58%, at ~90x lower cost per intervention than the LLM-as-judge alternative, with no degradation on standard benchmarks.
3. Protein structure discovery without supervision. Using Block-Sparse Featurizers (BSFs) , a family of methods to decompose a model's activations into multidimensional subspaces rather than single directions , Silico found subspaces in a protein language model whose activations correlate with known protein structures, without any labeled supervision.
4. PICASSO on Midnight-12k, one-shot. PICASSO is a method for interpreting digital pathology models. Silico replicated it in a single attempt: it breaks what the model sees into readable concepts, shows which drive cancer predictions, and simulates how changes to tissue would alter those predictions.

The real innovation: interpretability as a training loop

In February 2026, Goodfire announced a broader vision called "intentional design" , using interpretability to guide model training rather than merely analyzing models post-hoc. The approach involves decomposing what a model learns from each datapoint into semantic components, then selectively applying or filtering these learning signals. Silico is the productized form of that vision. Instead of interpretability being a post-mortem tool you run after training to understand what went wrong, it becomes a feedback loop you run during training to steer what the model learns.

Post-training can introduce undesired side effects that are difficult to detect and even harder to trace to specific training datapoints. A probe-based method can surface concerning behaviors that emerge during LLM post-training, and probes can identify the datapoints responsible for a specific harmful behavior. Filtering out those datapoints and retraining significantly reduces the behavior. That's a qualitatively different kind of control than anything RLHF or DPO gives you.

Who it's for , and what it can't do yet

Silico is aimed squarely at teams that are actively training or fine-tuning models. Goodfire partners with companies training foundation models across architectures and modalities. Current partners include Arc Institute, Mayo Clinic, Microsoft, and Rakuten. A notable scientific achievement came from Goodfire's partnership with Arc Institute: by reverse-engineering a biological foundation model, the team identified a novel class of Alzheimer's biomarkers.

There's a meaningful constraint worth knowing upfront: most people won't be able to use Silico to poke around inside a closed model like ChatGPT or Gemini. But you can use it to look at the parameters inside many open-source models. If your stack is built on proprietary APIs, Silico won't help you much. It's a tool for teams with model weights in hand.

The external research community is cautiously optimistic but not fully sold. Leonard Bereska, a researcher at the University of Amsterdam who has worked on mechanistic interpretability, thinks Silico looks like a useful tool, but pushes back on Goodfire's loftier aspirations. "In reality, they are adding precision to the alchemy," he says. "Calling it engineering makes it sound more principled than it is." That's a fair tension to hold: the underlying science is still evolving, and the gap between "we can measure this" and "we can control this reliably" remains real.

Access and pricing

Goodfire's platform is used by Fortune 500 enterprises, major healthcare institutions, and AI research labs. Access is currently gated , you can request early access via Goodfire's contact page. Pricing is not publicly listed; the company works on a partnership model. Goodfire has raised a total of $207.25M , giving it the runway to operate as a research-first company while building out the commercial product. If you're training a model and want to understand what it's actually learning , not just what it outputs , Silico is the most serious tool available for that job right now.