Item — HN Reader

I think people are more impressed with the "well-described explanation", which is a (mostly bullshit) interpolation from prior text on the internet. If you just focus on the "feed a narrative and get out a probability of a particular diagnosis" -- the ability to do that with other machine learning models has already existed.

If you give me a sample of 100k ER text records and final diagnoses, I would be pretty confident I can estimate a model that gets you +50% accuracy on my laptop using open source tools.

1 comments

What percent of human doctors diagnoses is also well-decribed bullshit, though?

Totally get the point -- in the ER it is a mix of narrative and tests, so hopefully not so many. But in general practice, probably more BS than any of us would like.

Part of the difficulty with healthcare (that I myself don't have an easy answer to), is that you have very asymmetric error costs. So chest pain, even if heart attack is low probability, you should be risk adverse in many scenarios and still do due diligence to eliminate that as a possibility.

Not saying that smart systems are not possible, just still skeptical that level of complexity is possible on the backbones of LLMs. The smart sounding narratives I am concerned are mostly red herrings though in terms of outcomes we care about, accuracy/making right decisions/etc.