Federated learning, a method in which models are trained across many separate data sources without moving the raw data to a central server, is drawing growing interest as a way to expand the private data available for AI training. In a recent Practical AI podcast episode, Lorenzo Sani, a research scientist at Flower Labs and a PhD researcher at the University of Cambridge, discussed how collaborative training could give developers access to sensitive datasets in fields such as healthcare while keeping that information under the control of its owners.
The conversation, hosted by Chris Benson, also covered how model trajectories, the recorded sequences of a model's actions and reasoning, might be used to improve agentic AI systems that carry out multistep tasks. Sani and Benson examined the case for open-weight models, which can be run and adapted locally rather than only through a vendor's hosted service. They also touched on practical applications for locally hosted AI agents, ranging from healthcare and robotics to simple household automation.
For Australian businesses and professionals, the near-term impact is limited. Federated approaches can reduce the need to pool personal or sensitive records in one place, which may help with obligations under the Privacy Act 1988 and the Australian Privacy Principles, particularly for health and financial data. They do not remove privacy risk on their own, however, and they add engineering complexity and cost. Organisations with siloed sensitive datasets that are considering fine-tuning models may want to look at open-source frameworks in this space, but most businesses will not need to change anything yet.
Key Points
Federated learning trains AI models across distributed data sources without centralising raw private data.
Flower Labs researcher Lorenzo Sani argues collaborative training can expand access to private data for LLM development.
Model trajectories are presented as a possible way to improve agentic AI systems.
Open-weight models are discussed as a route to locally hosted, adaptable AI agents.
Privacy-preserving approaches may ease compliance concerns but still require careful governance and carry added cost.