It's Machine Learning, yo

[Mongrel]

How it started, etc.

May 25th: Eating Disorder Helpline Fires Staff, Transitions to Chatbot After Unionization

Today: Eating Disorder Helpline Disables Chatbot for 'Harmful' Responses After Firing Human Staff

It's not like there was no precedent for this either (already!) (CW: Suicide):
⇳ Click to Expand Summary

March 30th: 'He Would Still Be Here': Man Dies by Suicide After Talking with AI Chatbot, Widow Says

A Belgian man recently died by suicide after chatting with an AI chatbot on an app called Chai, Belgian outlet La Libre reported.

The incident raises the issue of how businesses and governments can better regulate and mitigate the risks of AI, especially when it comes to mental health. The app’s chatbot encouraged the user to kill himself, according to statements by the man's widow and chat logs she supplied to the outlet. When Motherboard tried the app, which runs on a bespoke AI language model based on an open-source GPT-4 alternative that was fine-tuned by Chai, it provided us with different methods of suicide with very little prompting.

[hngkong]

What was that IBM quote from ages ago, "A computer can never be held accountable, therefore a computer should always be put in charge of people's lives"? Something like that I think.

[Yoji]

And here I was more fearful of AI poisoning us with generative shit to the point that nothing could be believed anymore, long before Skynet ever truly woke up. I never figured it would actually kill anyone at this point.

ETA: And how rich was that statement by from those tech leaders about mitigating the threats of AI like we should with climate change and pandemics? Because we've done such a crackerjack job with THOSE, right?

[Mongrel]

The problem is the same as it's always been: The assholes who want to use this shit to pad their pockets. Given the trajectory of techbros glomming onto 'AI', I expect a lot more of this shit.

I mean, just look at the example in question. Why was a supposed non-profit helpline so hostile to unions? The union was literally just four employees in this case! FOUR! Some asshole here is making or skimming money they damn well shouldn't be (discerning which charities are in fact, charities, which are corrupt, and which are simply fronts is an old problem too).

This was just another move to cut costs at the expense of basic functions by people who were too greedy to even understand how their org even worked.

Now where else have I heard that story recently...

As for the creator, seems to me that they were far too eager to push their creation without sufficient testing; I think we can safely say that if your software fails catastrophically enough that it makes the national news for doing so less than a week after its deployment, it was not adequately tested. Whether it was professional ego, an attempt to be the next Liz Holmes grifter, or simple negligent incompetence, it doesn't really matter, they also failed in their duty as a physician.

imsurroundedbyassholes.gif

[beatbandito]

ETA: And how rich was that statement by from those tech leaders about mitigating the threats of AI like we should with climate change and pandemics? Because we've done such a crackerjack job with THOSE, right?

The sad fact is the "people in charge" really do seem to think AI is the same level of threat as pollution/climate change. And they're wrong about the severity on both sides.

[Thad]

And of course it's obvious enough that it hardly bears mentioning, but a lot of the techbros, investors, and scammers who are suddenly all-in on AI are the same ones who were all-in on NFTs a year ago. It's another tech fad that's overhyped, poorly-understood, and ruinously wasteful.

[KingRoyal]

Unlike the Blockchain, AI definitely has some real world use cases, but it's mostly in brainstorming and idea testing. This dream that people have that they'll be able to put whole products on the market without human intervention is a pipe dream. Though, that probably will not stop Amazon from getting filled with even more nonsense books.

For example, NVIDIA put out this AI demo presentation


It's such an underwhelming use, requiring a bunch of developers to produce something that a writer could have done in 15 minutes

[Brantly B.]

Yeah I'm trying to recuse myself from the boards a bit here but I reeeeeally can't not bite on this AI vs. NFT comparison, because I've been doing quite a lot of AI stuff lately! And as stated, the huge difference between the two is that NFTs are inherently either useless (merely replacing one controlling entity with a different controlling entity) or worse-than-useless (creating scarcity where none existed before) while AI has pleeeeenty of useful applications, just not the ones that are currently getting trendy.

I'd equate AI much more closely with The Butt, which is appropriate because they're currently each others' most compelling use case. AI is good at some things and, much more importantly, requires a fair amount of detailed work to actually be better at those things than a well-crafted traditional solution. AI and The Butt have a similar problem though in that they appear to be much friendlier to those who don't want to put in the work, which give the impression that they're simply magical and automatic - and then you get hit with either very unwanted output or a gigantic use bill right in the eye. It's the difference between a car that refuses to start if you don't hold down the brake when starting the ignition, and one that'll happily slam into a wall for you.

(It should be noted that the current research thrust of most of these projects that have been released into the wild with seemingly no safety checks is "How do we build these tools without the inevitable Timmies crashing themselves into walls with them?". The issues we're seeing right now aren't unexpected, they're part of the experiment.)

[Mongrel]

The one that's killing me the most right now is the camera that uses 'AI' to take pictures without a lens (but instead has this hilarious rubber spider on the front, so it's not like we're saving space here).

There is absolutely a theoretical use case for all the tech they're putting into that particular experiment, but "superseding regular cameras" is absolutely not one of them.

[beatbandito]

It's such an underwhelming use, requiring a bunch of developers to produce something that a writer could have done in 15 minutes

also related:

[Thad]

Yeah I'm trying to recuse myself from the boards a bit here but I reeeeeally can't not bite on this AI vs. NFT comparison, because I've been doing quite a lot of AI stuff lately! And as stated, the huge difference between the two is that NFTs are inherently either useless (merely replacing one controlling entity with a different controlling entity) or worse-than-useless (creating scarcity where none existed before) while AI has pleeeeenty of useful applications, just not the ones that are currently getting trendy.

Right, fair enough. It's not an inherently useless technology like NFTs, but it has attracted many of the same snake-oil salesmen who were on that fad and are jumping to the latest one. I think we're in the midst of a very similar speculation bubble, even if this time the underlying technology actually has real-world applications.

It's also similar in its massive energy consumption, at least in current implementations.

[KingRoyal]

At this point I'm more worried that AI will be used to grey goo the internet. It's already hard to find stuff in a lot of search engines without running into content that's been harvested and built from other websites. Now people will have the ability to generate cheap product designs and websites and get them published very regularly. It's only a matter of time before Google becomes a see of poorly generated websites

[Brantly B.]

I think we're in the midst of a very similar speculation bubble, even if this time the underlying technology actually has real-world applications.

Oh, clearly.

It's also similar in its massive energy consumption, at least in current implementations.

A lot of that has to do with said Timmies having no idea how to optimize or tune their query engines. There are a lot of wasted cycles going on right now because of misunderstanding (and misinformation) about what's a "good" or "correct" result or how to get there. Even so, it's nowhere near as bad as the infinitely-growing resource demands of the blockchain. Hell, I don't think it's even as bad as spending an equivalent amount of time trying to run Cyberpunk 2077, but I can't find any numbers.

At this point I'm more worried that AI will be used to grey goo the internet. It's already hard to find stuff in a lot of search engines without running into content that's been harvested and built from other websites. Now people will have the ability to generate cheap product designs and websites and get them published very regularly. It's only a matter of time before Google becomes a see of poorly generated websites

That's sure happening but I think we can file that under "stuff that people were already doing but now they're doing it a spicier way." Right now the promise of AI is more dangerous than the AI itself.

I understand that's equivalent to a "guns don't kill people" argument but let's be honest here, a little education would go a long way in both cases.

[Yoji]

Over on DeviantArt, I used to start every session flagging things so the algorithm would stop showing me NSFW stuff (you follow ONE SMUTTY ARTIST and look what happens). Now, I do the same thing for generated stuff, which is somehow even worse. Bikini-chan's foot pics are obvious enough to avoid, but now I see things that make me go "that's... I want to say it's interesting, but something looks off," click through, and find it's generated. It's slow and in small ways, but it feels like the net is being flooded with even more shit.

I can't help but think of bit from Transmetropolitan. Some guy got himself a new set of genitals, made by Uraguayan children working in dangerous conditions for less than a dollar a month. But he's okay with that, because he now has junk that looks, feels, and functions exactly the same, but with a built-in mini disc player.

Like, we have all this tech and we could've done something truly great with it. But... this is what we do with it instead?

[Upthorn]

el problema es el capitalismo

[Brantly B.]

I can't find any numbers.

I'm compelled to dig in here to see if I'm wrong. Information on ChatGPT is dodgy as always but for Stable Diffusion:

The model was trained using 256 Nvidia A100 GPUs on Amazon Web Services for a total of 150,000 GPU-hours, at a cost of $600,000.

Those GPUs have a 400W TDP so they're roughly equivalent to two PS5s at load. So to create (a version of) Stable Diffusion you need the equivalent of 300,000 people running their PS5s for an hour.

Releasing a Spider-Man game is probably worse for the environment.

[Brantly B.]

Okay, now, GPT-3 took 355 GPU-years* to train, so now I'm taking this energy worry a bit more seriously in the case of LLMs.

1 BTC takes about 33,333 GPU-years*** to mine.

* 710 PS5-years.**
** 1,243,920 kWh, I think.

*** 116,798,832 kWh.

[KingRoyal]

You're only counting the energy cost of the training. Keep in mind, every day people are requesting Midjourney, Stable Diffusion and ChatGPT generate new outputs all the time. The image generation, in particular, needs to kick the requests up to some pretty beefy GPU rigs to get timely outputs, especially for that many requests. That can't be cheap

[Brantly B.]

Cheap, no, but still comparable to running a non-Switch game console at less-than-full load. It's a net win if somebody is generating pics of Bikini-chan's feet when they would otherwise be playing Call of Duty. The problem, besides the fact that those things are being run suboptimally on hardware that was never meant for it, is ubiquity; we have people implementing a power-hungry LLM solver where a simple binary tree search or, like, a volunteer human with a keyboard would do. But that's the problem here in every aspect.

[Mongrel]

Editorial on AI liability, cheeky enough that it just as easily could go in the Onion thread:

So my invention is now a threat to the world. Can you blame me? (No, seriously. Can you?)

Trillions of dollars have been spent researching, creating, developing, refining, training and testing AI, and all we’ve gotten is a high-tech drunken uncle whose loose relationship to facts and the truth becomes clearer the longer he speaks and the more often he uses the word “apparently”?