A.I. for socialising

[Babar]

Sorry, yeah, yet another AI topic, but it didn't seem covered by the ones we currently have, and I was curious, so I thought, why not?

AI has become more prevalent and commonly used by everyone, even though I feel it doesn't really fulfill a needed niche yet. However, I met someone today who is going through something of a separation with their significant other, and they were talking about how they were using ChatGPT to work through it (with themselves, not involving their SO), and that started sending alarm bells through my head.

They were talking about how it (ChatGPT) can take the patterns of behaviour from their partner and give explanations for it that really helped them understand some stuff. I have my own feelings on AI and its efficacy and safety with this sort of use, but obviously I didn't want to burden them with it, so I kept it to myself beyond some general talk about how an experienced therapist would have more detailed knowledge on the topic that could better help them.

I know of another person who, after a long day would relax with a cup of tea and just talk to ChatGPT as a form of relaxing socialising. Another friend (a foreigner, who probably feels a bit isolated) uses AI to have someone to talk to in their own language and learn the local language as well.

I dunno about you, but the idea that people are turning to AI for these use cases, and the fact of what the AI actually is (essentially LLMs with an autocomplete feature) kinda scares me. I read news about people who got isolated and harmed through the cycle of affirmation by these chat programs, and wonder if there is really a benefit here. Then again, I'm getting old, probably turning into that crotchety anti-technology guy I used to laugh at when younger, so I'm curious...do any of you use AI for this sort of "socialising" or something similar? Do you know people who do? What have been your experiences?

[Danvzare]

I think you're right to have the alarm bells ringing in your head. Especially with regards to using it to "understand" patterns of behaviour, considering AI's tendency to completely make stuff up.

Funnily enough, I'm not completely against the idea of using AI to socialize, since I've considered it myself. The key word there, being "considered", since I then remembered that talking to AI is like talking to someone who works at a call center. All you ever get are souless, braindead responses, with them not knowing anything but never admitting to not knowing anything, just so they can always have a response, even if it's completely meaningless. (I can't stand people who talk like that, so why the heck would I seek out a program that talks like that, but worse!)
Then I remember that I can have a more productive conversation with myself, which I then agree with, and then I agree with my agreement, and the conversation kicks off. 

I think my main problem with socializing with AI, is that I can not only see what's behind the curtain, but I understand and almost fully comprehend it as well.
It's a bit hard to be enamored with something when you have a "meh, that's old news" mentality to it. 

[cat]

I have friend (in a stable relationship, lots of friends and social interactions) who drives a lot. Others listen to music or audio books in their car, he enjoys talking to a chat AI instead for fun.

A colleague learns German and is very self-conscious about speaking. She trained talking to an AI and got significantly better, even when talking to humans.

I don't think those cases are problematic (apart from the huge energy consumption that AI causes)

Just yesterday I read an article about people having something similar to romantic relationships with AI chatbots. Now this is in an area that could get unhealthy. The person that was interviewed compared it to being immersed with a book or movie, the question is how deep you dive into it emotionally.

Interesting topic, in any case.

[Danvzare]

I have friend (in a stable relationship, lots of friends and social interactions) who drives a lot. Others listen to music or audio books in their car, he enjoys talking to a chat AI instead for fun.

A colleague learns German and is very self-conscious about speaking. She trained talking to an AI and got significantly better, even when talking to humans.

I don't think those cases are problematic (apart from the huge energy consumption that AI causes)

I agree. Those cases are in my opinion, are the correct way to use AI. 

Also I really don't understand the "huge energy consumption" that people always bring up. I can literally run the software to generate AI content on my potato of a laptop, and have a decent AI image generated within five minutes (I even had to use a workaround to get it to use my CPU, since my GPU doesn't support it, that's how much of a potato my laptop is).
My laptop literally uses more power when I play games on it, than it does using AI to generate content. 

[bicilotti]

I use LLMs to write the odd `for` cycle in languages I am not comfortable in.

They always output believable code, very reasonable code. And, sometimes, the very believable code they output does not work. The more complex your task is, the better you are investing your time in solving the problem yourself and not working with a computer partner who is as obsequious as dumb.

Which leads me to...

They were talking about how it (ChatGPT) can take the patterns of behaviour from their partner and give explanations for it that really helped them understand some stuff. I have my own feelings on AI and its efficacy and safety with this sort of use, but obviously I didn't want to burden them with it, so I kept it to myself beyond some general talk about how an experienced therapist would have more detailed knowledge on the topic that could better help them.

You did the right thing here, pointing them to proper help while not dismissing their feelings.
I have faith most people are reasonable enough to indulge in such LLM usage for brief periods, but alas not everyone and for sure not when they are experiencing intense pain.

Also I really don't understand the "huge energy consumption" that people always bring up

According to this article, by MIT Technology Review, a generating a standard image draws 1,141 joules. Not extremely big, but enough to make me wonder about the economic feasibility of the endeavour.

[Danvzare]

Also I really don't understand the "huge energy consumption" that people always bring up

According to this article, by MIT Technology Review, a generating a standard image draws 1,141 joules. Not extremely big, but enough to make me wonder about the economic feasibility of the endeavour.

How much is that in comparison to running a game on a high-end PC?

My laptop struggles to run Saints Row 2. It can, but I have to lower the graphics really down to get a reasonable framerate.
My battery lasts around the same amount of time running Saints Row 2, as it does generating non-stop images. So I can get around 30 images on a single battery charge, probably a little less. (I'd say that my battery lasts about 3 hours on full blast.)

Honestly, I'd be more concerned with raytracing and denuvo being damaging to the environment when it comes to energy usage. 

My real concern with AI is space. With all of this garbage people are producing, that's being uploaded to everything and everywhere, it's only a matter of time before the AI slop starts to outpace the rate with which we can store it.
Now you might be thinking that storage is cheap, which it is. But even if it's terabytes for pennies, trillions upon trillions of pennies are more than any company can afford. And with high quality pictures and videos entering the mix, I think it's only a matter of time before we start seeing this become the main issue. The flood is here, and it's looking to be unfathomably big.
Of course if this does become an issue, the environmental impact that follows, goes without saying. 

Now don't get me wrong, I can see how generating videos could be bad for the environment due to energy usage. But generating text and images... nah. I just don't see it.

[Galen]

According to this article, by MIT Technology Review, a generating a standard image draws 1,141 joules.

How much is that in comparison to running a game on a high-end PC?

There's a time component to that comparison, but assuming it takes one second for one of these instances to generate an image, that'd be 1-to-1 with 1141 watts. 570 if it's over 2 seconds, and so on. So assuming that the actual time to generate is pretty small, it's analoguous to running a medium to high end gaming system at full whack. But it's also likely that all of these servers are constantly maxed out generating images, since I'd imagine most images are unusual due to bad results. So it might be more sensible, depending on context, to assume it's closer to 10K joules per useable image.

As of December last year, I'm seeing estimations of 34 million images being generated per day (something that could easily have doubled since, but let's assume it's up to something like 50m now), though that's counting the major models and not the contributions of people AI generating furry porn. So we're already at the point where machines are creating over 100x the amount of art people are making by hand now. According to a statistic attributed to Forbes, 71% of social media images are now AI-generated. It's taken AI a year and a half to take as many photos as photographers took in 149 years. If we've not already reached it, we're probably a year or two out at most from the vast majority of all art ever made being AI slop. So no matter how small the individual contribution to climate change, or water usage, or energy supply issues, the end result is still considerable.

Honestly, the maths for how much isn't quite working out for me, since I'm ending up with only a constant 2000 watts to generate each year's worth of slop. Which I'm sure use collectively more power than two microwaves left on constantly. With 100 terawatt hours a year being a figure I've seen for AI as a whole.

Either way, seeing governments float the idea of people taking less showers to save water for data centres, or dropping Paris Climate Agreement goals because now the slop machine needs 10% of the global electricity supply should be irritating.

To bring my response back to the original topic: Yeah, using LLMs to tackle loneliness definitely feels exceptionally unhealthy. A bit like trying to 'treat' a breakup with hardcore pornography. That's not a substitute, and the moment it even starts to feel like one is the moment someone has gone down an extremely dangerous psychological rabbithole. That's going from 'talking to the cat when I'm alone' to 'me and my toaster are madly in love' (or the people who've supposedly married their rice cooker, anime body pillow, pet rock, or apparently the Berlin wall—if some random listicle is to be believed).

Using it to generate 'information' about partner behaviours definitely seems more benign, but there's still obvious dangers in using a text generation machine to do pretend psychotherapy. There's a reason doctors and lawyers avoid giving authoritative answers online (aside from wanting to be paid for their work). Better to look up actual advice written by humans and not just regurgitated in what we hope is an accurate form, and even better to not do that at all and leave actual evaluations up to real-life paid professionals who can actually understand your own personal context instead of potentially interjecting that your ex was abusive for once not responding to a text message, or that you may have repressed trauma because that's a recurring theme in the text it was trained on.

I suppose the next generations of memes around this are going to be particularly surreal: "Men will literally talk to a generative algorithm trained on the pirated sum-total of human knowledge instead of going to therapy"

[Crimson Wizard]

I had a funny thought recently, about Evil Queen asking ChatGPT about "who's the fairest of them all", and it was the AI that made her go kill Snow White.