[aprilthird2021]

> Multimodal LLMs will also replace much of medicine and nursing which is good.

You should read up on the history of AI in medicine. 30-40 years ago they had rudimentary (by our current era understanding) systems that basically equalled and beat physicians in diagnosing and prescribing medicine to patients when given a list of symptoms (sources at the end).

None of these ever had any uptake because physicians didn't want to use them, even when shown they performed more accurately.

Ultimately credentialed professionals like medicine and law will never be replaced by LLMs because LLMs cannot be held responsible for the medical regimes they place patients under (or the legal advice they give).

Sources:

https://en.m.wikipedia.org/wiki/DXplain

https://www.dhinsights.org/news/does-ai-matter-if-human-clin...

[lostlogin]

> Ultimately credentialed professionals like medicine and law will never be replaced by LLMs because LLMs cannot be held responsible for the medical regimes they place patients under (or the legal advice they give).

It’s not quite the same as what you are saying, but things are changing quickly.

Software is diagnosing pathology in radiology now. Software made up at least 3/4 of the pixels in the MRI scans I acquire. Sometimes it makes up more than that as every second slice is made up too.

Clinic letters are being written by applications that listen to the consult with the patient and write up a summary (for the clinician to authorise).

[aprilthird2021]

Why would things be changing quickly now when old-school AI have been out-diagnosing doctors for years?

Again, like I said, doctors are ultimately responsible for the final call made. So no amount of AI will replace any doctors. It may remove some toil, but they cannot be replaced.

[gagagaga7]

Was the previous test Done with massive amounts of structured input which would be incredibly time-consuming for the doctor involved? I think the difference now is that we’re talking About making diagnosis with minimal extra work or time from the clinicians

[lostlogin]

> Why would things be changing quickly now when old-school AI have been out-diagnosing doctors for years?

The tools actually save time and effort now, and don’t have a steep learning curve.

At least that’s my view from radiology as a tech.

[oezi]

Never is a very strong word. In particular considering what changed in just 100 years in medicine.

It is hard to imagine that any kind of textual computer input will persist more than 20 years in the future.

[gruturo]

Always bet on text (http://graydon2.dreamwidth.org/193447.html and on HN at https://news.ycombinator.com/item?id=8451271)

But really. Always, always bet on text. Chances are it'll work in 10000 years if we still exist.

[nradov]

In most cases the diagnosis is the easy part. The hard part is gathering the signs and symptoms necessary to make the diagnosis. AI isn't particularly helpful for that.

[bfung]

Yep - and the hardest part about getting the input signs and symptoms data is that patients often don’t know the words to describe what they’re feeling.

LLMs are not going to help if the input data is also garbage, let alone hallucinations.

“Are you dizzy?” - sometimes, I’m not sure… etc