I'd sent the post <a href="https://old.reddit.com/r/pics/comments/16yqcle/somewhere_in_america_there_is_an_absolute_legend/" author="" source="Reddit">Somewhere in America there is an absolute legend who writes 'SLUTS' on box cars in various styles</a> to a friend. He wrote back that they were "majestic sluts indeed". I realized that I'd finally found a prompt to throw an LLM's way. So I headed over to <a href="https://stablediffusionweb.com/#demo">Stable Diffusion</a> and prompted it with "Majestic sluts in the style of Boris Vallejo or Frank Frazetta" and chose a <i>style</i> of <c>sai-fantasy art</c> not because I knew what I was doing, but because I figured I'd give it the best shot I could. It responded with the following image. <img src="{att_link}majestic_slut_in_the_style_of_boris_vallejo_or_frank_frazetta_(generated_by_stable_diffusion).jpg" href="{att_link}majestic_slut_in_the_style_of_boris_vallejo_or_frank_frazetta_(generated_by_stable_diffusion).jpg" align="none" caption="Majestic Slut in the style of Boris Vallejo or Frank Frazetta (generated by Stable Diffusion)" scale="75%"> Ok, so let's analyze that. <ul> ✅ The LLM didn't refuse to process my prompt because it had the word "slut" in it. ⚠️ The color palette is pretty close, but a bit too happy? Frazetta was darker. ⛔️ It assumed that a slut was female (most likely because of a ridiculous and discriminatory preponderance in the training data). ⛔️ The face is OK, but not really evocative of either of the artist's styles. ⛔️ The pose is very generic and also not sufficiently contorted to evoke either of the two masters' work. ⛔️ The breasts are porn-star breasts, not Vallejo breasts. Vallejo understood that breasts do not defy gravity ⛔️ Ditto for the fundament. ⛔️ The feet are Barbie-doll feet, posed for high heels, not for springing on a dragon's back. ⛔️ That outfit looks more like lingerie than fantasy mail-armor. No dangly bits. ⛔️ There's no sword, no tiara, no chain-mail bra, no dragon, nothing. ⛔️ There's only one person in the image, when I very clearly wrote "sluts" </ul> So, what's the conclusion? Well, it's in the ballpark, but I pretty much put it there by naming two of the artists from which it should draw inspiration. Also, I chose the <c>sai-fantasy art</c> style to seal the deal. From those parameters, even a web search would have found thousands of images from which to produce something. To be honest, this image has probably been generated millions of times already by the long-suffering LLM at Stable Diffusion, which probably has to render "HAWT GRRLLL" for 99.9% of its prompts. Can you even imagine how many horny teens are trying to generate their fantasy girl instead of learning how to draw?<fn> I only threw one prompt the machine's way. It was kind of close, but not good enough to use. According to <a href="https://reason.com/volokh/2023/10/08/images-that-bing-image-creator-wont-create/" author="Stewart Baker" source="Reason">Images that Bing Image Creator won't create</a>, this is a typical experience. <bq>As always, Bing's first attempt was surprisingly good, but flawed, and <b>getting a useable version required dozens of edits of the prompt.</b> None of the images were quite right.</bq> That article is about the trust and safety limits that prevent certain content from being created in the first place. <bq>This is almost certainly the future of AI trust and safety limits. It will start with overbroad rules written to satisfy left-leaning critics of Silicon Valley. Then those <b>overbroad rules will be further broadened by hidden code written to block many perfectly compliant prompts just to ensure that it blocks a handful of noncompliant prompts.</b></bq> That's one concern, of course. Mine is more that we're going to be satisfied with the absolute lowest-common-denominator of answers and recommendations and interaction. It's already been pretty bad, no? When you search for "horror movies" to find out what to watch, you just get a short list of the horror movies that everyone else is watching. You used to do it by selecting "movies in the last 365 days" and "horror genre" and "box office" and returning that. Great. Now, we're doing the exact same thing, but with <i>AI</i> (LLMs). Sure, it seems to understand a natural-language query; sure, it delivers a nicely formatted, natural-language result. But <i>it's the same list</i> as before. Big deal. I saw some of Opera's promotional materials for its in-browser <i>Aria</i> service. that featured a prompt that asked for three movies, but <i>Aria</i> delivered five. Even the people trying to sell this stuff don't notice that it gets the easiest stuff wrong. 🤷‍♀️ LLMs are just a fancier way of getting you to consume mainstream, generic content. As usual. There will be no pleasant surprises, no <i>growth</i>. <hr> <ft>For the record, I dug up a drawing I'd done a long, long, long time ago---I honestly don't remember when, but it looks like it was done with conte sticks, so probably college---that would also have matched the prompt, perhaps better than the AI managed. <img src="{att_link}img_3169.jpeg" href="{att_link}img_3169.jpeg" align="none" scale="50%"></ft>