Silico by Goodfire

Silico is a platform for AI research teams that want to run long, complex experiments without constantly monitoring them. Instead of babysitting training jobs or manually coordinating multiple steps, researchers give Silico a high-level research goal. The platform then develops an experimental plan, runs the work in parallel across available compute, monitors progress, and returns results that can be inspected and built upon. It's built on Goodfire's interpretability research, so it's designed for teams doing serious mechanistic interpretability work, not just quick demos.

What makes Silico stand out is its agent-based approach to research. You give it a goal, and it handles the rest — planning experiments, coordinating multiple agents, and keeping track of context across long-running tasks. It comes with state-of-the-art training and interpretability libraries, and it can orchestrate compute across different infrastructure providers. That means researchers can stop worrying about job scheduling and focus on the science. The platform also supports common training methods like SFT, DPO, and RL, and lets you compare checkpoints and test targeted interventions.

Silico is best suited for AI research teams at organizations like Microsoft, Mayo Clinic, and Prime Intellect. It's useful for exploring how a model works (visualizing architecture, training sparse autoencoders, mapping neural geometry), diagnosing model failures (tracing regressions to undertraining or data artifacts), training and improving models, and replicating or extending published research. If your team spends more time managing experiments than actually doing research, Silico is worth a look.