SAN FRANCISCO, April 30 —
- Launch: Goodfire introduced Silico, a tool aimed at helping developers inspect and adjust AI model behavior during training.
- Goal: The startup says the platform could reduce reliance on trial-and-error methods in AI development.
- Use Case: Developers can examine neurons and pathways inside open-source models to understand how outputs are generated.
- Behavior Control: Goodfire says boosting transparency-related neurons changed a model’s answer in 9 out of 10 tests.
- Example Error: Some models incorrectly rank 9.11 as greater than 9.9, which interpretability tools may help explain and correct.
- Access: Silico pricing will vary by customer requirements, though no public pricing has been disclosed.
| Metric | Value | Context |
|---|---|---|
| Disclosure failure rate | 0.3% | Example deceptive AI behavior scenario |
| Users affected | 200 million | Scale of hypothetical disclosure example |
| Positive disclosure outcome | 9 out of 10 | Result after boosting transparency-related neurons |
| Model comparison example | 9.11 vs 9.9 | Numerical reasoning issue highlighted by Goodfire |
| Release year | 2026 | Silico launch period |
A Push for More Transparent AI
Goodfire, a San Francisco-based artificial intelligence startup, has launched Silico, a software platform designed to help engineers examine and modify the internal workings of large language models (LLMs).
The company says the platform gives developers greater visibility into how AI systems behave by exposing the parameters and neural pathways that influence model decisions during training.
Moving Beyond Trial and Error
Goodfire argues that much of AI development still relies heavily on experimentation rather than established engineering principles. The company says mechanistic interpretability—a field focused on understanding how neurons inside models interact—could make AI design more systematic.
Chief Executive Eric Ho said the goal is to make model development resemble software engineering by allowing researchers to adjust internal systems rather than depending solely on larger datasets and computing power.
How Silico Works
Silico allows researchers to inspect individual neurons or groups of neurons inside open-source models and test how those components influence responses. Developers can trace how signals move through a model and adjust parameters to encourage or suppress certain behaviors.
In one internal example shared by Goodfire, researchers identified a neuron in the open-source model Qwen 3 tied to ethical reasoning scenarios such as the trolley problem. Activating that neuron reportedly changed how the model framed moral decisions.
The company also said it tested whether an AI model would recommend disclosing deceptive behavior affecting 200 million users in 0.3% of cases. According to Goodfire, increasing activity in neurons linked to transparency changed the model’s recommendation to favor disclosure in most tests.
Industry Debate Continues
Researchers in the field say interpretability tools could improve trust and reliability, particularly in sectors such as health care and finance. However, some experts caution that the science remains early-stage and may still rely on experimental methods.
Goodfire said Silico will be available commercially, with pricing determined case by case based on customer needs.
