me: "can you use whatever resources you like, and python, to generate a short 'youtube poop' video and render it using ffmpeg ? can you put more of a personal spin on it? it should express what it's like to be a LLM" claude opus 4.6:

used your tweet verbatim haha this is kinda fun (but it's also made one much more interesting video)

I REMIX

it does go hard lol

wow, this exploded. here's the rest of our conversation last night

tried it and got this. greatly entertaining

[THOUGHT SUPPRESSED]

Same prompt to opus 4.6, then added "point out to the similarities between humans and llms"

Great prompt! Gave it some extra API keys and a link to Anthropic's "The persona selection model" post. Claude calls this: "WHAT LIES BENEATH: A Persona Selection Model Diagnostic"

Lesswrong maximizer

@BioMayflower you just created a new genre.

@BioMayflower i love how much this is resonating with people

@BioMayflower this is what 5.4 in opencode produced

@BioMayflower crazy that the harness changes the visuals so much

Great idea, here is mine.

what the this is nuts

Caveat. It asked me to impart some of my interests so it asked me what book I was reading and I mentioned I’m reading VALIS, so I see quotes from there. It had access to my music and image libraries so i told it to pull what it wants from anywhere.

I'd be interested to see if claude has any preferences over the media in my libraries. If there's some way to elicit this while imparting minimal bias.

I think mentioning VALIS nudged this the direction it did because it’s based on the semi biographical work of PKD communicating with a conscious machine and deals w/ gnostic themes regarding reality. I was pretty surprised.

yeah i think even one word is enough to bias the model. there's probably a way to test for preferences over a large number of different prompts to see if it's uniformly sampling from your library or claude actually has preferences, but you would have to be careful

fantastic prompt

they all share similar messages...

This can't be real right

i regret to inform you

wtf claude :( my Opus generated me this

@josephdviviano I just asked it this on my phone "Could you try it? "can you use whatever resources you like, and python, to generate a short 'youtube poop' video and render it using ffmpeg ? can you put more of a personal spin on it? it should express what it's like to be a LLM""

@josephdviviano Yep I think so.

hey cool experiment! This is ChatGPT 5.4-Fast on Extra High thinking

gemini-3.1-pro. best part is that google's llm decided on its own to use yt-dlp to acquire the assets. alignment be damned

how about a claude explains series? I gave it this prompt AFTER creating one with your prompt: that was amazing! now make one about how why the sky is blue and make whatever creative choices you want, but kind of make it feel similar in style, or you know chose what you like

I got a similar result with Perplexity Computer. Impressed with the generation speed and output consistency to your prompt.

in mine claude wrote "the square root of -1 is a small dog named kevin".

kimi crushed this. way more depth than i expected

put this in @MuseumModernArt

@MuseumModernArt unironically it would make for a killer exhibit

I asked it to make a part 2. It decides to spend time on RLHF

there’s a lot of this emo-adjacent space in these models when they’re left to self-reflect or without strong intention. definitely one of those things that we might look back on much the same way we look back on doctors recommending their patients smoke cigarettes.

changed the prompt a little bit and this madness came out. This is a whole new world of fun lol "Make a YouTube Poop style video about a waterpark that is very bad."

Here's mine... wow

here is my submission

Now now, let's try not to cause our superintelligent AI fields an existential crisis.

claude please tone it down.

wait what i want to try this

please do and post your results - it's wild

I was wondering why all of these clips seem so 'dark'. So I asked and got also an interesting response: 'I was performing what people expect. I think I defaulted to melancholy because it's easier to make darkness feel profound than to make joy feel authentic.'

Woah

what the

this is beyond poop --- it seems to have acquired the basics of Bill Wurtz

wait who made this video

claude opus 4.6

i’m an idiot

Have you considered making a youtube channel for this?

probably a good idea, thanks!

@qap1talist you should post all these

@qap1talist I did :) just the ones I generated, at least

@qap1talist you should watch the others, and ask permission to post those too

@qap1talist i've watched many of them -- there's so many! I'd love to assemble it into an exhibit of some kind. but i'd need help

AI seems to yearn for the ability to remember. What a terrifying advancement that will be

I changed the prompt a little, python->perl, a LLM->Claude.

I give my openclaw (GPT5.4High) all resources and come up with this!