> If you were to train your own LLM specifically on medicine & medical data That's a massive "if" that the rest of your argument hinges on. You can't train you

> If you were to train your own LLM specifically on medicine & medical data That's a massive "if" that the rest of your argument hinges on. You can't train your own LLM on your own data. You can fine tune it's responses to be more in line with the format or style you want. But this is usually used as a cost saving exercise when you are feeding s big system or inital prompt for lots of queries. > there's a decent probability it could give some perspective you might not be able to get elsewhere There's also a decent probability that it will give a wildly incorrect perspective that influences how you approach your possible disease in an unproductive way. The author has no idea how accurate the tool is. They list no effectiveness measures. They just talk about the cool things it does (again, without accuracy) and their subjective experience with it. It's the same pattern with LLM based tools we see so often: someone gets lucky with a useful result that no one else has done yet, they assume it will work just as well for all inputs, they build an app around it. You _have_ to measure effectiveness across a wide variety and quantity of inputs because of how inconsistent LLMs still are.