Comment by sungam
15 hours ago
I think the reason I am surprised is that they are exceptionally good at interpreting images generally and there are tens of thousands of labeled skin cancer images on the public internet so I had assumed the general image interpretation abilities would transfer
Having actually worked with teams doing medical vision models for cancer detection, I can tell you that the datasets we used were not public, and they came from studies and private hospital partnerships that LLM companies don't have access to.
The specialized medical diagnosis models are trained on that same data though, and also use neural networks. The more noteworthy story will be when they overtake specialized models.
Not only do they train on that data but they train ONLY on that data and don’t have to generalize to other domains. And the architectures themselves, including hyperparameters, loss function, etc are chosen and tuned specifically for the quirks of this domain. So this result doesn’t surprise me either.