[pj1115]

This is super cool, but maybe it's important to distinguish projects like this from frontier medical AI research (i.e. new models). This however does exactly what lots of health advocates and even pharma companies have been getting at for decades: informed doctor-patient discussion and patient self-advocacy. Organising and interpreting medical records is an awesome use case.

I imagine any potential user of a tool in this state would know what be wary of, but I do worry a little that LLMs have been trained on bad health advice or misinfo, and there will be very limited training data from real patient diagnoses. Even if dangerous data is minuscule in training terms, can good prompting de-risk dangerously bad advice?