The AI dooooooom thread

[DGuller]

There was an experiment done on whether doctors with ChatGPT can be better at diagnosing than doctors without ChatGPT.  The finding was that the help from ChatGPT was very marginal at best.

The kicker though was that ChatGPT alone did much better than doctors with ChatGPT.  In effect, the doctors were holding ChatGPT back, probably in part because they couldn't conceive that an LLM could perceive something they couldn't, like a lot of people in this thread.

https://www.nytimes.com/2024/11/17/health/chatgpt-ai-doctors-diagnosis.html

[Grey Fox]

LLMs have no reasons to have public facing endeavours. Those things are marketing ploys.

[crazy canuck]

LLMs have no reasons to have public facing endeavours. Those things are marketing ploys.

elaborate 

[Grey Fox]

LLMs have no reasons to have public facing endeavours. Those things are marketing ploys.

elaborate 

LLMs have really cool capacity to identify in all sorts of applications. Like in healthcare where they are use to better detect/find cancers in scans, or where they are integrated in scheduling apps to better manage doctor appointments by simply guiding to the right professionals.

That they've been made available to the public where we use them to do unprofitable dumb things like image generation and our agendas are only marketing ploys. Maybe a training opportunity too.


Here's the AI software the team I'm part of develops.

https://www.teledynevisionsolutions.com/products/astrocyte-sdk/?model=astrocyte&vertical=tvs-dalsa-oem&segment=tvs

It's a very niche AI model creator for the machine vision industry.

[Sheilbh]

Agree on the marketing ploy piece. I think the business model is going to be basically providing LLMs as SaaS - either by providing infrastructure for private mini-LLMs (Mistral) or as an extension of an existing SaaS product. Although that's not a massively sexy pitch

Although the best I've seen in my sector is actually NotebookLM from Google - primarily because of its citations/sourcing, which I think they released in that way and have been surprised at the business interest.

Also think there's a fair amount that is being pitched as AI - and when I'm talking about AI the bit that interests and worries me is generative AI, a lot of the pitches and products (for example, I think the cancer scanning) is machine learning.

[celedhring]

I'm finding ChatGPT useful for text editing and proofreading. Particularly as a non-native English speaker. "Oh yeah, that's a more natural way to say this".

It does it have its kinks though - it loves redundancy, and relatedly sometimes it's a bit too wordy.

[Tamas]

I'm finding ChatGPT useful for text editing and proofreading. Particularly as a non-native English speaker. "Oh yeah, that's a more natural way to say this".

It does it have its kinks though - it loves redundancy, and relatedly sometimes it's a bit too wordy.

Maybe it is an introvert. I have been recently explained that introverts say things in a too roundabout way. In other words they only say a simple yes or no when that is the appropriate answer, but what do I know.

[DGuller]

I'm finding ChatGPT useful for text editing and proofreading. Particularly as a non-native English speaker. "Oh yeah, that's a more natural way to say this".

It does it have its kinks though - it loves redundancy, and relatedly sometimes it's a bit too wordy.

You can have custom instructions for it.  For my work account I told it to be concise and to the point, and stop giving a long preamble to every answer, and it did become quite terse.

[Sheilbh]

Interesting other example on the medical front - this was more machine learning - but AI system reviewing scans (I forget what for) was only better than actual physicians in basically perfect scenarios. So when the individual was stoop up for a chest x-ray, but very often what they're looking for is associated with trauma or other complex injuries so it is often people lying in a bed getting quick x-rays etc.

Which just shows how very, very important it is to get the right input/training data.

(Which is obviously why LLMs love journalism: quality, legaled, accurate....And why they should pay for using it )

[crazy canuck]

Interesting other example on the medical front - this was more machine learning - but AI system reviewing scans (I forget what for) was only better than actual physicians in basically perfect scenarios. So when the individual was stoop up for a chest x-ray, but very often what they're looking for is associated with trauma or other complex injuries so it is often people lying in a bed getting quick x-rays etc.

Which just shows how very, very important it is to get the right input/training data.

(Which is obviously why LLMs love journalism: quality, legaled, accurate....And why they should pay for using it )

Yeah, and add to that the recent findings that radiologists using AI were worse (meaning they found more) at finding both false positives and and false negatives.  The really concerning bit was the false negatives were more prevalent.

Turns out they trusted the AI - (or a cynic might suggest, let the AI do the work) too much.

On the bright side, when the AI got it right, so did the radiologist.  But again, that is because they came to depend on the tool.

 

[Josquius]

I've found myself using Chat gpt a bit lately. I think I've found a decent use case.

I'll write something, which has all the proper points... buts its convoluted and weird.
Chat GPT then rephrases it into a more succinct form.

This seems to go....OKish. Though chat gpt definitely has its own idiocentricies that come through. It also loves correcting itself. It gives one suggestion, then give that suggestion back as the redraft and it totally redoes it. Its also so very American, really puts way too much of a positive promotional a spin on everything- I do need to do this to an extent...but too much would just rub people the wrong way and sound false.
 Its very much a bounce ideas off it and then do your own redraft tool.
But still...interesting use.

Anyone else found uses?

[garbon]

I've noticed it likes wordy business speak and it likes to be redundant.

[Sheilbh]

I've noticed it likes wordy business speak and it likes to be redundant.

I think this is an inputs issue

[garbon]

I've noticed it likes wordy business speak and it likes to be redundant.

I think this is an inputs issue

Perhaps what it was trained on, I'd suggest

[Sheilbh]

Sorry, yes that's what I meant by the inputs into the model.

It does present a slightly terrifying future as more and more of these companies develop enterprise SaaS products, learning from their enterprise customers. Skynet but communicating like someone from HR