#2756

Ah yes, the “dead mom” haircut.

Explained soon…. that is very ominous. Outside a statment of your mom is dead to us I fail to see ways it can be explained soon, other than her getting killed. Which I don’t think you would do.

I think the reason why she isn’t here maybe due to the fact that her dad explained that Kate’s mom still saw her as a monster even after she had to flee, something her dad couldn’t bear with, more about that in this page of the comic https://www.yoshsaga.com/comic/searching-for-a-lost-daughter/

The mother thing has a far more recent mention thing, as she’s a crazy woman actually doing illegal stuff in her hatred of Chimeras. Kate has been in contact with Jane and knows of her, but this is the first time they have met in person. She was invited though.

Oh! I see, on what page of the story is that mentioned about Kate’s mom? Just curious to know since I missed that

Here it is – https://www.yoshsaga.com/comic/2561/ Pages around it are talking about it all too but this is the one that actually shows her mother.

(Warning – Rant about AI ahead Read at own peril)
Also it’s unfortunate it was some time ago when I first got a chance to play around with backgrounds made using AI img2img, where i would draw a background and then “enhance it” with AI. Most of the time creating garbage – and being disappointed that I wasted time with it or used it in the comics. So then for a while I stopped and someone showed me a new one that claimed different stuff and it wasn’t using stolen art etc etc – and it did produce better results but – to get what i wanted I was basically drawing and detailing backgrounds already and it was just modifying it a bit and throwing clutter and noise in it. So I stopped bothering with that too and just went back to painting and using my cheap background tricks- like comic artists have been doing since before digital art existed… Hey even today most comics don’t even do backgrounds 75% of the time. Now i have a good collection of gradient techniques for simple rooms and a bunch of textures, brushes, and materials I can use to throw together background, and I am better as painting in general. I do wish i could learn how to do rapid modeling for 3D but i was never good at modelling in 3D. I could do sculpty modeling fairly well (That’s a type of puddy like clay stuff) but in 3D i was much better at texture and material work back in the day. I’m way out of practice with it now though. Still, I would love to use some 3D assets for background stuff and creating perspective changes. Funny enough- perspective shifts and things like that that I think make things more dynamic- AI is terrible at them. Because they average out so much information you end up with Generic 80% of the time, terrible 19% of the time, and 1% of the time you end up with something that might look kinda interesting- and then you realize that its more like a reverse image search than it is actually making art, and if you have the exact same model, seed, program, prompt and settings- it will generate the exact same image. It’s not really creating, in a sense, it’s always already been created by this program, but if you did all the exact same things to me- asked me to make the exact same image for you as a commission- Well unless i legit copy and paste the image and say “here ya go-” I can’t do that. I can’t recreate the exact same image on a different day, in a different place, no matter how much my tools and settings and compensation are identical. My settings- don’t work like a computer. If you think that makes me better, thanks for the compliment! If you think it makes me worse- im sorry 🙁

Large Language Models (a subset of AI) are snapshots of accessible culture, the “stolen art” is mostly in the public domain. It isn’t really creative (unless you count their made up “hallucinations) in the way that deep neural nets can be when they make pattern matching ona level neither humans nor other methods do.

But some means to make them so are researched. For example, by assigning different LMMs different roles they can creatively work as a group. For example assigning various expert roles, an integrator and one criticising the result seems to work. For backgrounds in art, perhaps a flat color expert, a shading expert, a consistency expert and a critique? The critique can be prompted to “be creative”.

Torbjörn – Im curious what you mean by “the “stolen art” is mostly in the public domain.” Because Public Domain has 2 meanings, it’s official meaning is things that have no protection via copyright which is only things from like over 70 years ago (in the USA, life of the creator + 70 years is the default) and even some of those are still not free to use- see how Mickey Mouse Steamboat Willie only version only just got released into the public domain, vs what many people think “if it’s available publicly it is “in the public domain” which is completely incorrect but is a common thought. A bunch of my friends art is in these models, and they didn’t give permission for them to be used that way.

And if you are talking specifically the LLMs using texts, this is known to be false as well because modern information can be found in LLMs including information behind paywalls and publishers are suing for their stuff being used without permission.

As for your point on the expert roles – none of those exist and I dont think LLMs can do anything like that- a collaboration of LLMs trying to create their own thing- wont make anything actually new still, it’s not having any unique thoughts or creations, it’s all preordained 100%. You might argue its the same for people, but it can’t be proven in people because we cant account for every single parameter with people, but we can with computers. LLMs, give then exact same parameters will always output the exact same thing.

Sage – To my knowledge, LMMs don’t repeat output. Many large softwares are “creative” in that way, if they are for example compiled at runtime output can depend on computer configuration. But repetitions is not used to measure creativity, and there are some early attempts to do that with software. [“AI writing is improving, but it still can’t match human creativity – Computer program finds that ChatGPT and its ilk remix words well, yet their output remains derivative”, Science, Dec 19 2024]

The group work LLMs exist. [“Multi-Agent Collaboration Mechanisms: A Survey of LLMs” arXiv:2501.06322v1 [cs.AI] 10 Jan 2025] But I don’t think they are developed for this application, hence the suggestion.

Perhaps there is some nuance to the word “text” that I\m not familiar with, but generally it should be written texts in the context of LLMs. Yes, LLMs may include paywalled material as training material if the constructor paid for it. This is a viable criticism as for instance “public domain”, “fair use” et cetera is not global rights and there may be infringement even if used for training and not copying. But there is no outright “trainingright” legislation that I know of [not an expert]. I don’t think court cases have been made yet, if they can be made. We’ll have to wait and see what comes out of it. Meanwhile, today’s LMMs are there and can be used.

So I’m no LLM: I wrote LMMs twice.

Also, on paywalled material sources. I suspect that the data bases and/or web crawlers used as text sources may have reached material that *should* have been paywalled, I see copies on the web all the time. That would be part of the legal processes that should oversee LLM use.

Torbjörn – Just because copyright material is found on places it wasn’t meant to be, or things that were paywalled find their way to other places does not give any legal right to use that material. Fair use does not cover what AI does because it is copying, and its is the owners right to allow or disallow it for any reason as of US law. Fair use is not a free pass to do what you want with it, it’s a defense in court, which means it needs to be defended in court, it doesnt just mean they are exempt from the law. LMM doesnt seem much different from LLM other than it’s less about text- but no- it does produce the same thing with the same inputs unless used with Quantum computing. That is a restriction of all computers and physics. Im sure you are aware there is no such thing as random in computers, just complex algorithms or associating things with outside information but even that- is things that can be recorded and reproduced.

This means everything it “creates” is just the output of what it already had. It cannot create new things without new input. Again, this is a limit of computers as they exist. As for ” repetitions is not used to measure creativity” no – repetition is not, however something is not being created, there is no imagining being done, it’s recalling instructions and associating words to create an image or a sentence structure, or instructions. It’s not being creative, It can’t, it’s just performing it’s instructions based on its parameters. If you get a different response- congrats- you are using a different parameter, which could be a seed, the time, your cpu clock, your gpu, there are tons of parameters it uses to determine its output.