It’s really hard to know how this will play out. The models only have to improve a bit at this point to be reliably better than humans, as which time it probably makes sense to replace humans. It seems they will probably still hallucinate but do it little enough that it’s still a net gain to use them. Compute power needed to run them will surely come down.
I’m as skeptical as the next guy, but I do think they will have uses, especially in examples like radiology which he he uses as a negative case. However I’m pretty sure it will eventually be able to do the initial screening to find the 95% of cases with nothing at a rate similar to existing medical diagnostic testing and then return the other 5% back to a human to review and decide further treatment. Based on my experience with speech language models, I’m pretty sure you’d be able to tweak the models to produce mostly false positives rather than false negatives and then run it through further layers of review afterwards.