ChainThink reports that AI interpretability company Goodfire has opened its research tool Silico.
After researchers define their goals, Silico creates a plan, runs multiple experiments simultaneously, tracks progress, and organizes results—supporting model training and optimization, paper replication, and troubleshooting issues in data and model design.
During the research process, researchers still need to review protocols, adjust directions, and evaluate final outcomes. Goodfire stated that Silico is designed for executing and analyzing experiments within the research workflow, but it does not replace human decision-making.
Goodfire also used Silico to analyze Kimi K3, which has 2.8 trillion total parameters, identifying characteristics within the model that control writing style, and adjusted these features to make Kimi K3 respond in a "pirate tone."
Silico is currently available via the macOS client, with a personal plan costing $1,000 per month. Additional fees apply for using Goodfire’s GPU, and integration with your own server is also supported.