Show IH: 💊 MedGPT - Get information on medicines, treatments and diagnoses

I think generally the issue isn't with the model training. It's more with how these models work to begin with. Large Language Models operate off the principle of basically figuring out what is most probable to come next in a piece of text. In general, what is likely to come next not always equal to what's true. E.G. If you give it a bunch of context saying X medical fact is true, then the model would be more likely to agree with you since the context makes it seem more probable. OpenAI tries to prevent it from saying false (but probable) facts. But it's not foolproof and false facts can slip through. This is sort of an active area of research for all the AI giants and it's something nobody has solved yet. Even Microsoft, Google, and OpenAI have been unable to fully solve this issue. It's not a trivial task to solve simply because large language models like GPT care about probability not facts as mentioned before.

There's tricks you can do with taking medical information => converting it to embeddings => running an embedding search when GPT is asked a question => feeding the medical information the embeddings point to into the prompt. That will decrease the likelihood of false information. Other solutions like training your own model are less likely to work. Yes, there's been companies who have custom trained LLMs on datasets and achieved better performance in specific areas (e.g. Bloomberg's recent financial LLM comes to mind). But those are companies with massive amounts of resources (even then they haven't fully fixed the false information problem afaik). Any individual is less likely to be able to achieve this because 1) There's not many open source LLMs to base it on that also allow commercial use (LLAMA is not for commercial use) 2) Training is one thing. Training it to not say false information is a whole different beast. As a whole, any LLM you train yourself will probably be lower performance than GPT and be even more likely to say something false IMO. OpenAI put a lot of effort into preventing AI models from saying false information and it's unlikely an individual can match their work. So I would not recommend trying to train a model yourself.

Regardless though, even if you do the embedding trick there will still be a chance the model will say something false for the reasons mentioned above. I would generally shy away from using GPT (or any large language model) for medical usage. I've seen advances made on LLMs regarding medical advice on the news (e.g. Med-Palm-2 is probably the closest thing to a human doctor). But even then, they aren't perfect (plus Med-Plam-2 isn't even fully public yet). In general, if Google, Microsoft, and OpenAI can't solve the false information problem then it's unlikely you will be able to yourself. You can use GPT for a lot of stuff (sales, marketing, etc). But when it comes to medical advice (which is a highly sensitive topic), the characteristics of GPT make it too risky. By using these models for medical advice, you're essentially putting something on the web that might occasionally give bad medical advice to someone and put their health at risk. Not to even mention that this is very risky to you from a legal perspective. Yes, you can have a disclaimer, but regardless from my perspective it's still a risky proposition.

1. Not sure. In order to make it significantly better than chatGPT alone you would need to do something like give it access/ability to scan medical research journals or something like that, which would be a lot more work and impossible to know how much difference it makes until you try it.
2. :checkmark:
3. Interesting, I've never heard that term before and my dad is a physician.