The AI Thread | Ringfencing the discussion

[Asura]

So... AI.

It's come up in other threads a fair bit, even de-railing one at one point, so it's probably good that we have a thread.

Similar to some of the thread which have come up lately, I'll probably do this in the form of providing questions, and then tailoring questions based on people's answers.

So...

Have you used any of the AI tools that have made so much news in the last ~18 months? What did you do with them, and why?

[Dogg Thang]

Well you know I have. Impressive, intriguing, lots of fun and a bit terrifying all in one. My main experience has been in the text generators such as ChatGPT and image generators such as Midjourney. There are loads of other things going on, obviously, but I haven’t explored much beyond text and still images.

In images, what has hugely impressed me is the speed of advancement. I think I started playing with AI in 2019 and it was able to make fun, trippy images that were absolutely no use whatsoever, although even back then there were AI filters that were pretty good. But the level at which the generative AI has jumped in that time has been insane. It’s very easy to still find flaws in AI-generated imagery but the reality is that it’s already better than many working artists and it achieves results in the space of minutes. For me, the fun part is in playing with the stupid stuff, such as asking it to show me He-Man and Skeletor looking sad while having a beer together. My favourite output came from an afternoon of trying to get Midjourney to give me Dario Argento’s Godzilla. Took me a lot of playing with the prompts but eventually got a very cool output and, honestly, something I would call genuinely creative (with the caveat that there is a lot to debate about the nature of the art, due to AI having no sense of intention).

With text generation, what impresses me is something I don’t see a lot of people talking about - it’s not about the actual content of the output, it’s in how well it often interprets the query, to the point where I’ve seen it take initiative, completely understanding something I’m asking for in the query and going beyond it and almost making assumptions on what other specifics might be tied to that request. While some of the outputs look impressive at first glance, earlier versions of ChatGPT can quickly descend into nonsense and it was also a dirty rotten liar. I pity kids who were using it for homework (and a lot of kids are). With things like integration into Bing, it can fact check and that’s a whole different thing. Even without that, latest versions don’t hallucinate in the same way earlier versions did.

With text, part of me think it’s still a way off replacing a good writer and yet I’m also seeing many articles online now that I suspect are AI-generated. Like early digital music where the quality was much lower than a CD, yes, you can say it’s not good enough or not ready (I don’t think it is) and yet I don’t know how much most people notice or care. And it’s getting better all the time.

The big thing for me is the integration, such as now with Bing and ChatGPT being able to use DALL-E. Many of the limits begin to fall away when AI is using other AI and the other thing that happens is that the need for human intervention reduces. Already the wrangling skills of “prompt engineers” aren’t needed in the way they were just 6 months ago.

There are so many questions we should be considering about this, about the nature of the future workforce, about copyright ownership and attribution and so on. Although one question I don’t see talked about often enough is - what about the cost? The big AI companies are taking on massive amounts of investment and, at some point, are going to need to make a lot of money. Like streaming TV, social media and so on, we get quickly used to things being so cheap or even free as markets are “disrupted” only to find the new normal is not sustainable and things break. The cost of AI as it is in the future, rather than what it looks like right now, is something that needs to be considered by everyone who uses it. Because I don’t think it’s going to be cheap eventually.

[MartyG]

Lately, I've been using it to help generate my YouTube thumbnails after training a model using my images. The jump in quality from Stable Diffusion to SDXL is quite a big leap.

I'm running a local instance, but you can try it here: https://creator.nightcafe.studio/studio - the results are getting better.

[Asura]

With text generation, what impresses me is something I don’t see a lot of people talking about - it’s not about the actual content of the output, it’s in how well it often interprets the query, to the point where I’ve seen it take initiative, completely understanding something I’m asking for in the query and going beyond it and almost making assumptions on what other specifics might be tied to that request. While some of the outputs look impressive at first glance, earlier versions of ChatGPT can quickly descend into nonsense and it was also a dirty rotten liar. I pity kids who were using it for homework (and a lot of kids are). With things like integration into Bing, it can fact check and that’s a whole different thing. Even without that, latest versions don’t hallucinate in the same way earlier versions did.

I've found something similar, in that ChatGPT can be both impressive and, well, not-so-impressive.

As a test, I tried to ask it to generate character-sheets for an RPG. I asked it to create Akira Yuki from Virtua Fighter. What was impressive was how it listed all his major moves, and gave reasonable matches for how those were built in that RPG system. It also populated all of his stats.

Unfortunately the stats were largely wrong. Obviously you could correct them but that's the weird bit, because in that scenario you've got the computer doing the more creative bit but then a human's got to do the mathsy fact-checking bit. In the end I wasn't convinced it saved you much time, even though it was interesting.

On the other hand, I also asked it to generate Ryu. Same problems. But the fascinating part is I asked it to write a story where two players, using those characters, fight each other - and asked it to show the necessary dice-rolls and rule clarifications. Frankly, the result was very impressive. Sure, it got most of it wrong but it was written in a very convincing way, where anyone not familiar with the system would just accept that as given. It's definitely very good at writing text which "seems legit".

The art side though... Is weird. I'm admittedly seeing a lot of YouTubers using AI art already in their videos where previously they might've commissioned art or used stock photos.

Lately, I've been using it to help generate my YouTube thumbnails after training a model using my images. The jump in quality from Stable Diffusion to SDXL is quite a big leap.

I'm running a local instance, but you can try it here: https://creator.nightcafe.studio/studio - the results are getting better.

Are you not concerned about where those models were trained, though? You might get sued.

Admittedly one of my concerns as a content-creator is that so many people have just gone for this right out-of-the-gate, that now we're all competing with people who use it. I'm not sure if, in the near future, not using AI is going to be like refusing to use a digital camera as a photographer.

[Hirst]

As mentioned in the thread, I use ChatGPT to do reply Emails to managers I can’t be bothered dealing with. I find it generally saves me a lot of time, but I had to get the prompts down to make it come up with things I like. I specifically put “I am a British office worker” as otherwise the wording is too American-sounding. I also say things like “three paragraphs max” and “no really big words” as it has a tendency to prattle on with lots of silly management speak that makes it sound obviously too AI.

What I do find interesting is to make it write additional scenes to films or TV shows. If the full script is floating around on the Internet, it has a pretty good understanding of how most characters act, their motivations and especially how they talk. It falls down a lot if you get it to write about something super-obscure, like when I had it write about the awful 70s frat boy comedy King Frat and it clearly only knew the name of the main character and everyone else was “Frat Boy 1” etc. The scenes it writes are very formulaic and predictable, but also genuinely passable if it has all the info to work with. You could definitely write an additional series of a crummy sitcom like My Family without much help and it would be just as entertaining (not entertaining).

[MartyG]

Are you not concerned about where those models were trained, though? You might get sued.

No. I'm not.

[Asura]

No. I'm not.

Can I ask - in your work, did you pay for fonts/stock images before? Or make sure those you used were free for commercial use? Or is it more of a hobby thing?

EDIT: This sounds more judgemental than it really should. I guess I'm asking if your videos are more hobby or business.

[MartyG]

I use Google fonts, and given I'm not making money from YT, there's nothing commercial there, nothing is copy pasted from anywhere, it's all derivative.

[fuse]

Honestly, I'm really not comfortable with the ethics of any of them and have stayed clear.

Everything seems to have been hoovered up into the models that train these things with a total disregard for ownership, credit, or even accuracy, and the tools have then been handed straight over to the public with next to no guard rails or guidance. Irrespective of what they cost to set up and run, the legality of how they've done this, and the propagation of complete mis-truths, the general public have been armed with the ability to fake competency at a number of skills at a totally unprecedented scale, and there are a lot of industries who are having to manage the implications of that.

I know early on in this there was a big thing where people would be dunking on the quality of the output - "look at the fingers!" - but this will continue to become less and less of an issue and should not be the only thing that makes people consider how they're using them.

Do I think there's value in AI technology? Yes, absolutely. Do I think society is ready for an AI revolution? Not in the slightest. Do I think we should let fail-fast, disruption-obsessed tech bros be the ones to guide us into the future? No, in the strongest terms available.

I am seeing a lot of friends who are using these tools for all kinds of things, whereas it is beginning to feel like this might be the luddite hill that I choose to die on, and that prompts me to go retreat into a cave.

Oh and for what it's worth, music is the one I'm really dreading.

[Neon Ignition]

*Enters thread* Wrong AI...

->Open Spoiler<-

[Dogg Thang]

Are you not concerned about where those models were trained, though? You might get sued.

Admittedly one of my concerns as a content-creator is that so many people have just gone for this right out-of-the-gate, that now we're all competing with people who use it. I'm not sure if, in the near future, not using AI is going to be like refusing to use a digital camera as a photographer.

I feel you’re right about this. Except that AI is a terrible leveller in the sense that it can do all the work for you but it can do all the work for everyone else too. Spent years building up your skills? Great, you can now use AI to help your workflow… as can every single other person, skilled or not. So I feel like it would be nuts to ignore AI and yet, as soon as you start to rely on it, you’re now on equal footing with every single person on the planet. That’s great for someone unskilled. Terrible for someone who has built a name for themselves or spent years honing their art. The moment those people use AI, I think it has brought them down, not up.

In a way, it might be a little like ceramics in the sense that you can mass produce plates and vases and most are. They are churned out. But there is still room for the hand-crafted artistry. It’s niche, there isn’t room for many but it’s there. For artists, I think there could end up being value ultimately in not using AI and not damaging their identity or brand in the process. Even really bad art might stand out for that alone. Kind of like how you see some really ugly handmade mugs that have a charm precisely because they’re not factory-produced. But I don’t know.

In terms of lawsuits, I’m completely with those who call it theft when AI is trained on copyrighted work. But I think the cat is already out of the bag and I think there is little chance of any lawsuit holding up with current copyright laws. I just don’t think it’s a war artists will win. It’s worth keeping an eye on the current Sarah Andersen case, much of which has been thrown out but there is still one claim against Stability AI not dismissed yet. It will be interesting to see if that goes anywhere. And of course the other question there is - can AI-generated work be copyrighted? Right now, AI work is not protected by copyright. So if you use it and publish something made with it, you don’t own that.

[Asura]

Right now, AI work is not protected by copyright. So if you use it and publish something made with it, you don’t own that.

I mean, it really shouldn't be allowed. I kinda feel that when a person tries to be protective of their AI generated images, it's a really daft look. Only exception is if the AI was fully trained on your own work.

[MartyG]

Any derivative work will be if it involves a person making changes, the US court only ruled that a work generated solely by AI can't be copyrighted, and that's technically because it wasn't created by a person with original touch (and also only has precedence in the US).

For example, if you painted a copy of the Mona Lisa, but did it in a style of your own, you'd own the copyright. Likewise, if you took a photograph of a part of it in a creative way, you'd own the copyright of that photo. So long as it's transformative, you can skate on fairly thin ice.

[Dogg Thang]

Any derivative work will be if it involves a person making changes, the US court only ruled that a work generated solely by AI can't be copyrighted, and that's technically because it wasn't created by a person (and also only has precedence in the US).

And this makes it too dangerous for many uses in terms of copyright. Because AI is not the tool in most cases, it’s the worker. It’s not like Photoshop. Ordering an image in Midjourney and calling it your own is like ordering a burger and saying you’re a chef now. If there is a danger that you can’t claim ownership of a work even partly generated by AI or that it can be contested, that would be a real problem from a legal standpoint.

[MartyG]

We already have copyrighted work considered transformative as I've linked above. There's already precedence for this.

From a more philosophical viewpoint - doesn't all creative work build on what came before it?

There will absolutely be people fearful and resistant to AI and varying degrees of protectionism as people fear it taking over their jobs, skills, way of life, just as all disruptive technologies have historically. AI will be the next industrial revolution.