Why is this problem important and relevant?

Unlocking distributed private data is crucial because today’s AI models — even highly specialized ones — are constrained by the limitations of their training datasets.

Take, for instance, an Orthopedic AI Model designed to analyze knee X-rays and detect injuries or conditions.

Currently, such models are trained on datasets available locally within a hospital, university, or region. This creates a serious problem: bias. A model trained on patient data from a single country or ethnicity may underperform when diagnosing patients from different demographics.

Ideally, healthcare institutions would want to train AI models on globally diverse datasets to make them more robust and accurate. But today, strict regulations on healthcare data sharing (HIPAA, GDPR, etc.) make it nearly impossible for, say, a hospital in London to collaborate with a research institute in Japan — or for an injured person in Africa to upload an X-ray and immediately receive an accurate, specialized AI-powered diagnosis.

This bottleneck isn’t limited to healthcare. This could be extrapolated to banking, cyber security, energy and consumer goods.

Personalized AI Models are another frontier where unlocking private data matters. Imagine a future parallel to Apple Intelligence, but open and customizable — where users select a lightweight open-source AI model trained solely on their own device.

Such a model could learn your daily routines, financial health, personal goals, and aspirations — helping you make better decisions (like stopping doomscrolling) rather than reacting only to your immediate queries.

Today’s AI models are reactive because of small context windows and a lack of deep, longitudinal personal data. A personalized, evolving model — trained continuously on-device — could enable true predictive and proactive AI assistance.

In both healthcare and personal AI, the inability to safely access and train on distributed private data is holding back the next major leap in AI’s evolution.

To enable this future, two major constraints must be overcome:

1. Distributed AI training — the ability to train models across decentralized, siloed datasets without moving the data.

2. Trustless coordination among multiple stakeholders — creating mechanisms that allow institutions and individuals to collaborate on model training without needing to trust each other or expose sensitive data.