Lorenzo Sani: federated learning fits agentic AI's private data

Sani's argument builds on his experience developing federated learning tools for language model training during his PhD at Cambridge, where the goal was not only privacy but also access to new data sources that would otherwise be unavailable. He notes that federated learning is already being tried in industry, though some companies abandon it when the techniques prove inconvenient for their specific use cases. The conversation situates agentic AI as a step up in complexity: agents can take actions, branch into sub-agents, and interact with external services, producing trajectories that mix model outputs, user prompts, and third-party tool calls. Because those trajectories can contain sensitive personal context, Sani sees federated learning as a natural way to train on them without centralizing the data. He also flags that privacy guarantees from techniques like differential privacy depend on the setting and are not universally applicable.