Ran an experiment with my blood panel: converted it to CSV and sent it to 2 doctors and 3 LLMs.

Not only did the LLMs match the doctor conclusions, they had more details and were easier to digest than the doctors.

We are entering a new era of accessible diagnostics.

But how do we query LLMs with sensitive data like health records without compromising privacy?

Anyone working on solutions? Local inference, encrypted computation, zero-knowledge pipelines anything else??

Founder of @rotki 🐦, the portfolio tracker that protects your privacy.

#Berlin. University of Tokyo graduate. Marathon runner. #ethereum developer. #birding

For privacy purposes, it’s easy to download and set up a model for local inference with ollama for example. It’s especially easy as a developer and takes just a few lines of code. Open source models small enough to be run on a laptop (e.g. Gemma 3) have gotten very powerful. This is not the most user-friendly option but has been working well for me.

I have done it but the results were not nearly as good as the big ones (chatgpt, claude, gemini).

Perhaps in a few years time? I surely hope so as that would be the easiest solution