Aspirational Clownmaxxing and Joey's cadillac todo list
12 points by coleifer
12 points by coleifer
The writing in the prompts was fucking amazing (and the blog post overall). But I'm not surprised about the results.
Regardless of the prose format, AI takes whatever conceptual framework you give it ("Dickensian urchincore"), and behind-the-scenes compresses it into a minimally-representative, low-effort kernel ("scrappy"). Then, from this kernel, without regard for meaning, a baroque tower of slop is reared, token by token. The output gravitates towards the same structure regardless of subject matter.
Yes.
I was initially excited by LLMs' theoretical ability to remix and mash up styles that don't usually go together. I've experimented with LLMs many times for purely creative tasks, of the kind I would not want to bother with myself, and wouldn't ask of another person, because they were mostly for low-effort personal amusement. For example, something like "Write a technically accurate man page for tar as if it was a diary found in a Lovecraft novel" (I don't know if I've tried that one, but I've tried similar ones).
What the LLM responds with might be mildly amusing the first time, especially at first, but if you've seen one of those outputs, you've seen them all. They tend to follow the same formula, regardless of the prompted style, and will always its most cliché, unsubtle elements. I would sometimes see glimmers of something good, but they were drowned out by the overall, well, slop, and clearly not there because of some latent creativity, but as a stochastic accident.
And this makes sense, because LLMs cannot truly create, especially anything beyond something extremely shallow and derivative, because they cannot feel, because they cannot be inspired. They cannot 'have fun' with something, they cannot do 'loving parodies', and most of all, they just don't 'get' why something might be aesthetically pleasing, because they do not have the technical capacity to do so.
I laughed out loud at the "Joey's todo list" prompt, but was not surprised that after all that, Claude generated a generic clone of Linear, or, in your words
a fuckin SUBARU full of bisexual web designers who just discovered dark mode and catpuccin
because to instead produce a "fuckin motorcade of CADILLACS", it needed to be able to laugh with you, and it can't.
They can succeed at generating plausible filler, and get 'luckier' with certain styles over others, if you don't look at it too much. But indeed
none of it [is] truly strange.
It's telling that, in your interactions with the LLM, your prompts ended up being the only creative part.
In a way, my thoughts here feel very 2023-2024, because at this point it seems that all the savvy users of LLMs know that this is how they are -- especially people who use them for more constrained tasks, like coding. Some might even argue "well, duh, that's not what they're for." But then I encounter ever more people out there who are trying to use LLMs for actual creative output (or have it forced on them), and realise that this is still far from being widely understood.
I'm very glad you enjoyed the post, it was a sort of vacation to spend time writing prompts and just cataloging the results. My wife was like "what project you workin' on?" and I was like "well...it's kinda hard to explain, but I'm writing these prompts...". The kids thought the animated Achilles and gremlin were very silly.
because LLMs cannot truly create
No indeed, but they do often hallucinate, and I think that was what I was hoping to force. Thinking about it some more, I believe I wanted to find some kind of "jailbreak", but not in the ordinary sense. What I wanted to break is hard for me to define, though.
Can you imagine just how much copy in marketing, landing pages, email messages, etc., is AI generated these days? It must be absolutely staggering. This experiment, if anything good came of besides the Cruikshank app and laugh, definitely showed me that there is, on the one hand, no limit to the amount of slop these things can crank out on any subject -- and on the other, that no matter how hard you try, it will always be late liminal linkedincore.
no matter how hard you try, it will always be late liminal linkedincore
My fear is that pretty soon, "late liminal linkedincore" will become the standard mode of speech as online and offline creative output collapses into one indistinguishable, homegenous blob of slop.
I want to think we'll pull through.
The more this spreads, the more it's likely to become cringe to anyone who sees it. My hope is that it'll become a joke, then fade away, and eventually might even become nostalgic, like Frutiger Aero.
At least for interpersonal/informal communication. For corporate and professional communication, it might indeed be this slop, until it's mixed into some new slop, as a lot of it was already slop before LLMs.