Hacker News Reader

Comment by KaoruAoiShiho | Hacker News Reader

Yep this is something that only AI can solve. Same situation applies to education, sales, HR. Human powered bureaucracies and systems suck.

View on HN

AI would be deployed to behave like the median doctor (at best - or maybe the lowest-common-denominator) to avoid blowing up costs with 99%-likely-to-turn-up-nothing hunts for super-rare conditions.

Today you can try to cajole your human doctor into listening more, or ordering more tests, or considering things you heard online or from acquaintances. AI will be guided to take that into account less because a doctor being more sympathetic and bypassing "standard practice" is an expense caused by humanity that the machine can be trained to avoid.

Today you can go across town and try your luck with another doctor. If it's all AI, you'll just repeat your story to the same basic model and get the same basic dismissal.

The problem arose from trying to make people behave like machines in order to save money. Making a machine behave like a machine ain't gonna help.

You need to shift the goal from "saving money" to "helping people." AI doesn't do that.

The costs are so low you can easily inference a bit longer. The idea that a computer would be as lazy as a human is not even close to reality.

Not the cost of running the ai, the cost of potential tests and medical treatment if they do find something, I think?

Nah those make money not cost money (for providers).

You can do an obscene amount of inference for a fraction of the cost of an average doctor's appointment.

Any AI will most certainly reflect the biases of the bureaucracies responsible for their creation.

Nah AI can easily be programmed to be much more patient and investigate edge cases and figure out personalized solutions thoroughly and provide bespoke service. This problem would be solved, though of course there are other issues with biases of the bureaucracies.

If you can do that easily, you will have no shortage of investors. But it’s not easy - getting the data alone is a huge problem.

I think OpenAI has plenty of investors...

https://x.com/deedydas/status/1933370776264323164

OpenAI doesn't lack investment capital. What they still don't have is a good source of high quality clinical data. And this isn't just a matter of buying access to deidentified patient charts from some large health system. Most clinical data quality is kind of crap so using it directly for model training produces garbage output. You need an extensive cleansing and normalization pipeline designed by human clinicians who understand the data at a deep level.

Absolutely true, but is there a system that works perfectly that I can use now that has all that that isn't AI?

In the absence of such a thing OpenAI is already quite good, some theoretical perfect shouldn't be trotted out as a counter if it doesn't actually exist.

There is a massive difference between being able to discuss symptoms and actually being diagnosed. Diagnosing edge cases is precisely the kind of thing that a generative AI needs massive amounts of data on. Where is this magical data source that violates doctor patient confidentiality? And all the doctors just contribute to it and everything is fine? Come on…

That’s true but they have a massive shortage of clinical data. Doctor patient confidentiality is a thing.

> Yep this is something that only AI can solve.

How? I'd expect them to already have standardized lists of the most useful next thing to investigate given what's already known, and a modern "AI" would actually be worse at that than some sort of solver engine with a database of costs/risks (for tests) and conditional probabilities.

Maybe if they're still using (digitized versions of) paper flowcharts things could be improved, but the most powerful tech should be old-school stuff rather than modern "AI".

No normal person would actually be able to use a specialized solver database, the woman in this story would already be dead before the guy figures out that such a thing exists and manage to make an account. https://x.com/deedydas/status/1933370776264323164

LLMs already work fantastically with pretty good UX.

Does "here are the most useful things to find out next" really need that complex of a user experience?

>Yep this is something that only AI can solve.

[citation needed]

Reply on news.ycombinator.com