since I finished my masters Thesis, I have been working a lot more with GPT-3 and I can’t stop to find potential and good ideas. The following video is showing the current development status:
I’m looking to bounce ideas off of others builder as well! Would love to show you the education platform that I am building as well as discuss ways that we can leverage the power of OpenAI to make a massive impact in the world.
What are some good platforms to connect with you on? Perhaps we can schedule a zoom call in the private messages here.
The demo looks good.
I have some feedback and questions.
On the first read, I thought AI-Doctor is a doctor on AI field. So the title was a bit misleading. Not a problem, but if you want to share the video again perhaps you find a better title.
Then I was thinking at the character from the video. I wonder if it’s a good idea to use a human-like AI character which acts as a doctor.
I’m not sure if it’s okay or not. I just thought that as a patient I wouldn’t like to talk with a robot which looks like a human. This is somewhat strange.
What is this ?
The conversation seemed sort of long (many tokens). How did you do it?
I assume you are running this on straight Davinci 2?(just noticed your using Curie fine tune)
I think there would be potential to use Weaviate vector db as a backend, create medical info embeddings with OpenAI, semantically query the db using user query as input, and use Davinci 2 to generate response from the returned results.
This way you can continue to add medical info to the db and your response times will stay around 200ms up to 1 trillion records.
Imagine the disclaimer you will need to provide with this though
Oh, yes a title like “medical AI” would make more sense, but this is just a prototype/project I am working on, so nothing that will see the light of day any time soon.
The face is a visualization from Synthasia.io, but no elemental part of the project, so I can change that to a hovering Siri ball, or just some plain text appearing. Still have to find something that appeals to humans in need of medical assistance.
Hyperparameters like Temperature, TopP ect.
The model is fine tuned on medical dialogues and emapthetic dialogues from two large datasets, thats why the answers tend to get longer.
I am running on Curie, as it is cheaper, and i am a college student with limited ressources.
That sounds like an interesting project to try out. I have found fine tuned-curie to “missbehave” in about half the cases, therefore I am really interested in working on guiding the converstaion using hard medical facts. Pivot the entire model to be a bit more task orriented, without loosing the free style of conversation.
As mentioned before, I do not intent to make money of of this. It is basically a research project, so no real application at the moment. But when this gets past the prototype phase, ill probably write one
Thank you for your feedback, however this is not intended as a public project and the focus is not on making money. I am just working on creating systems for academic research.
This is very cool! One comment I have is that I would be hesitant to release this to the public as GPT-3 has told me that I should jump off a roof in the past when chatting with GPT-3 as a therapy bot. How are you going to control/prevent these types of responses from GPT-3?
It recommended killing infants too. Kind of weird to be honest. That’s what I am working on in my PhD. I love the potential GPT-3 brings to the table, but I want to make it safe.
Which is almost impossible.
Currently I am working on an underlying knowledge web, which contains keywords for the communication to be guided along, and then use the Keywords and GPT-3 only in a function of an NLG to communicate an answer.
Great demo, huge market potential. I got close to Babylon some time ago that do something similar. I have interest in extending this to mental health, which I would argue is more stretched in human resources. I’m developing a physical device so no PC or phone is needed for conversations. Please drop me a message if you would like to explore collaboration.