Practical AI’s Chris Benson talks with Flower Labs researcher Lorenzo Sani about training models on sensitive data held in different places and using the resulting techniques to improve agents. Federated learning brings computation to participating data holders rather than pooling their raw records, potentially allowing organizations to learn from data they otherwise could not share. Sani cautions that locality alone does not settle privacy: a trained model can memorize sensitive information, and methods such as differential privacy provide guarantees only under particular conditions. Agent traces add another wrinkle: a task can accumulate prompts, retrieved files, tool calls and subagent results, all useful for diagnosing failures and all potentially sensitive. Teams need to inspect what happened before deciding whether to retrain the model or fix an application step or tool; the interview sketches federated training on such traces as a direction, not a demonstrated production result. Sani and his housemates also built a Halloween costume judge with locally hosted vision and language models, a separate example of keeping an agent application at home rather than a federated-learning experiment.
“The question becomes more when you’ve trained the machine learning model, does it have, has it memorized something about yourself that you didn’t want it to memorize, right? And that’s where privacy can break.” — Lorenzo Sani
“Maybe you didn’t think very carefully about that particular step, or that particular external tool that the model was using. But the only way to know is by looking at what happened.” — Lorenzo Sani