Silico (3 min read)

Goodfire has released an early preview of Silico. It is a model design platform for designing AI models with the precision of code, looking inside them, and deliberately finding failures.

The company said its interpretability research has led to achievements such as discovering Alzheimer's biomarkers, correcting self-hallucination in language models, and diagnosing performance bottlenecks in robotics models. Silico packages these cutting-edge interpretability techniques so researchers and engineers can use them right away.

The core features are as follows.

• Look inside models: Decompose into interpretable features to distinguish real understanding from spurious correlations.
• Health diagnostics: Catch undertraining, information bottleneck, and feature collapse early.
• Debug failures: Remove confounding factors and find problems before deployment.
• Steer behavior: Use internal features to build stronger predictors and steer generated outputs.
• Generalize with less data: Adjust training distribution, objective, and architecture to generalize more broadly with the same data or less data.

Silico includes model neuroscientist, an autonomous agent that plans and runs experiments in parallel, and also provides a collaborative environment for organizing research threads and reproducing and extending papers. Early access is open now, and MIT Tech Review also covered the CEO and co-founder and the significance of this product.