[DiscourseFan]

I don't know if its something that we will be able to measure precisely or at least in aggregate at first since the positives will only come about in the long term. Anecdotally, all my recent trips to the doctor have basically been as follows:

1) Give some (frontier) model my symptoms, circumstances, etc.

2) Model outputs a list of possible causes, some more urgent than others

3) If its not something benign, I go to a doctor

4) Doctor orders blood work, checks me up, and I discuss possible diagnoses with them.

Around 60-70% of the time the model was correct about the illness, the doctor just confirms it with empirical evidence. This means that I am able to pretty effectively triage myself, saving me a considerable amount of time.

Now its already the case that the role of doctors is diminishing in medical practice, and Nurse Practitioners are performing the majority of the same work. Armed with AI tools, Nurse Practitioners are effectively at the same level of skill and knowledge as an actual MD, and we would need far less MDs, then to do medical care, reserving them for specialty cases, expertise where AI tends to fail. So there is both more room for particular and intricate medical research and care for the most stubborn conditions, and more capacity for care in general.

In the long run, this will lead to improved outcomes, greater cognitive and physical health, and lower costs for child rearing. All of this is an economic boon. And all of it will not be apparent in the numbers for decades. I believe this is the case for the majority of the processes that AI is improving.

[dimbletimbers]

I’m sorry but this is a bad take. We’re suffering from a healthcare crisis, globally. We don’t have nearly enough doctors. Even in the US where wait times are relatively low, it sometimes takes months to get an appointment with my PCP, let alone a specialist. And while I do sometimes feel when I see an NP at an urgent care like I could have used an LLM for all the time and attention they paid to my concerns, that’s not a compliment to the LLM.

Besides that 60-70% is not an acceptable hit rate and there’s already early evidence of skill atrophy in AI saturated fields like radiology. We need way more people offering more personalized care, using more judgement and offering more compassion, not a chat interface that says “you’re totally right to push back, the post nasal drip was the smoking gun.”

Maybe some forms of ML, even LLMs, become part of the solution to that shortfall, but we’re a long way from proving that’s a net positive let alone a big leap forward.