Hacker News Reader

Comment by scyzoryk_xyz | Hacker News Reader

I can imagine AI being module-based. Like with plug-ins or apps, call it what you will.

So if you’re, say working in a hospital in a specific role, you will be advised by an AI module for that specific system. Or if you are working in an OR with a specific surgical procedure every day, you will have an AI module active all day to augment your knowledge and deliver relevant information where it’s needed.

It’s going to make some types of work immensely satisfying.

View on HN

I work in a field related to surgeon education. Until the results of AI output can be guaranteed to be deterministic and proven to always be correct, the legal ramifications of an AI providing incorrect information or promoting off-label usage is likely insurmountable. There are dozens or hundreds of eyes that see this information before it is approved for use and weeks can be spent arguing over a single word.

If the doctor decides, and the AI is just a tool among many, then why is there a problem?

I am very dissatisfied with current diagnosis analysis, as they are often wrong, as well.

I really don't want an AI make medical decisions for me, but doctors are overworked and if AI can help spot things, they otherwise would have missed, then I really hope those tools won't be blocked for legal reasons.

A kind of second opinion system might be good.

The danger I see is that you become to relient on it. There is a great science fiction book (forgot the name!) where everyone forgot how things worked (and didn’t need to know!) because machines did all the work. It created this dystopia where people where immensely shallow and immature because all things just happened for them.

"There is a great science fiction book (forgot the name!) where everyone forgot how things worked (and didn’t need to know!) because machines did all the work. It created this dystopia where people where immensely shallow and immature because all things just happened for them."

Are you sure, that wasn't just a ordinary newspaper? At least I get that feeling quite often.

Interesting observation :) Yeah, the mechanism is at work in various ways already if you think of it.

ChatGTP says, "The book you are referring to is likely "The Machine Stops" by E.M. Forster. This novella was published in 1909 and is set in a future where people live in individual rooms, connected to a global network that provides them with all their needs. The machines that run the network have become so efficient that people no longer need to learn or do anything for themselves. As a result, they have become isolated, shallow, and dependent on the network. When the network starts to break down, the characters are unable to cope and must face the consequences of their reliance on the machines."

I don’t think it was called that, but maybe I read it as a retelling of sorts. It was a book probably from the sixties or seventies but it’s possible!

Possibly Player Piano?

There's also A Logic Named Joe, but that's more about ChatGPT sans RLHF.

Maybe diagnosis assistance, but I was specifically responding to the idea that the AI may deliver information. The risk that this information may not be correct even 0.01% of the time will be considered unacceptable from a legal and regulatory perspective. A PDF surgical technique can be guaranteed to be accurate.

One of the GPs I had in my life gave me the medical book to diagnose myself and when I said that’s his job he just picked something randomly and gave me some pills (that I obviously didn’t pick up). Humans are generally worse than a dice for mostly everything. Gpt might be right sometimes and at least won’t be wrong out of laziness or malice (maybe; we don’t understand why it does what it does beyond statistical text prediction).

That’s sort of what I mean when it comes to the AI working with plug-ins - so you’d have a specifically designed {insert medical conglomerate name here } AI vetted and maintained around a particular purpose, say idk noticing suspicious tumors during or outputting information on what’s going on where it’s needed in the system. It could be helpful even with very simple stuff.

One of my roles has been writing surgical edu content, and I’ve used OpenAI to output procedure descriptions. I’ve unfortunately found that they were 99% accurate and then jumbled some significant thing..

Based on how things are unfolding in stable diffusion land, we're going to see general purpose base models that are then fine tuned for specific tasks using LoRAs that can be quickly loaded/unloaded as needed.

>LoRAs that can be quickly loaded/unloaded as needed

No need to shuffle them around, you can combine different finetunes in one composition.

Kindof? But then, for high-risk cases like this you'd need to validate their combinatorial effect.

Correct, as long as they become ROI positive/business backed, one module per profession.

In the future, this extends to robots, of course.

Reply on news.ycombinator.com