Silico Automates AI Research Experiments

Goodfire’s Silico platform automates AI research, from interpretability and model training to GPU orchestration.

Silico is an agentic platform designed to run research experiments on AI models. Given an objective, the system develops an experimental plan, executes multiple tasks in parallel, monitors their progress, and compiles results that researchers can inspect and build on.

The platform combines these agents with interpretability techniques developed by Goodfire. It can train sparse autoencoders and probes, examine a model’s internal geometry, test causal hypotheses, and connect problematic behavior to specific learned representations.

Silico also supports training through SFT, DPO, and reinforcement learning. It can compare checkpoints, diagnose regressions linked to data or model architecture, and reproduce or modify experiments described in research papers.

Compute orchestration is also part of the service. Silico distributes experiments across GPU infrastructure, monitors long-running training jobs, and can use Goodfire’s resources or connect to a customer’s own cluster. The company says it has applied this approach to interpreting Kimi K3, a model with 2.8 trillion parameters.

Individual access is priced at $1,000 per month, with usage allowances renewed weekly. Goodfire also offers custom plans for organizations, including team management, centralized billing, and a zero-data-retention option. Access currently requires an application, with selected researchers onboarded gradually.