folio 10 of 27
party in the ai lab: from stick figures to shoggoths
- date
- 29 NOV 2025
- era
- pre-Fable
- status
- complete
- form
- song
- updated
- 02 AUG 2026
the actual boards, prompt chains, failed frames, and visual decisions behind a 176-second ai-lab parody.
i made this as a parody of weird al's "party in the cia." the idea was simple: put the ai-lab race inside a cheerful pop song and keep smiling while the decisions get worse. it was my second parody, after every prompt you paste.
the song took three lyric drafts. i kept moving the joke until the cluster, the missing safety switch, the alignment patch, and "optimized away" felt like one escalating story instead of a list of ai references. the folder also has four generated male-vocal mixes and reference tracks for both "party in the cia" and "party in the u.s.a."
the plan was a giant wall
the visual plan lived across four excalidraw boards. part one followed the first verse line by line: pack the bag, leave home, take the embarrassing golden gate selfie, enter the lab, meet the hoodie researchers, turn on the cluster. the earliest version of those shots was genuinely this rough:

the opening four shots, directly from the board and the finished video.
i was not drawing for beauty. i was trying to decide where the camera was, what object carried the joke, and whether the next cut made sense. the backpack sketch already names the superintelligence book and silicon valley map. the bridge sketch already has the phone. the lab sketch is basically a stick figure walking through two doors. ugly was enough to keep the sequence moving.
the art direction moved too. the first notes aimed at the upa-style geometry of the original parody. later notes point to gravity falls and star vs. the forces of evil. that is why the finished film has a cleaner saturday-morning-cartoon look than the first visual tests.
one bridge shot became 34 nodes
the golden gate shot lasts a few seconds. its saved canvas has 31 image nodes and 3 video nodes.

one small shot, after a source frame, character reference, compositing, pose work, cleanup, and face correction.
the chain shows what "generate the shot" actually meant here. first i tried to move my avatar into the bridge scene. then i fixed the phone pose. then the backpack. then compression. then the face. then the pose again because he still was not clearly taking a selfie. one saved prompt asks for a crisp 4k rerender. another replaces the face while preserving the body. another finally extends the arm toward the camera.
the finished frame survived a small argument with the image.
one joke, three versions
the line "and the safety switch is gone" needed to work before anyone had time to inspect it. the written plan says:
show a power button and a safety button.
then show the safety button decomposed.

the first pass explained the idea. the later versions made it readable in one cut.
the first image had three separate controls and a broken panel. the selected board frame reduced it to two big buttons: power and safety protocols. the video goes one step further and leaves a hole where safety used to be. that was the better joke, so that is the one that survived.
the second half stops being cute
the board changes after the first chorus. the avatar starts crossing items off the safety list. release beats alignment. benchmarks flip on overnight. people leave with boxes while robots walk in.

the visual escalation under "rush that model release," "passed every benchmark," and "workforce minimized."
some notes are polished scene descriptions. others are messages to future me:
same as image before, but with even more dataceter
make it everyt time more crazy dance
i am leaving the typos because that is what the planning really looked like. the board also has loose labels for "power seeking," "compute," and "freedom / creativity / love / wisdom." near the "optimized away" section, human and ai sit on opposite sides of a stop sign. none of that appears literally in the final cut, but it shows the problem i was trying to solve: make the party bigger while the human part gets smaller.
the shoggoth problem
the last unresolved note asks whether three representations of ai were too many: chat screens, robots, and the shoggoth. the shoggoth might be unfamiliar, but it could do something the friendly robot could not. it made the hidden system feel huge, strange, and badly covered by a smile.
so i kept all three. screens handle the ordinary assistant. robots make deployment physical. the shoggoth carries the thing nobody fully controls. by the final chorus, the masks are on and the machines are dancing without me.
the folder is the receipt
the local manifest lists 48 indexed media files and 8.8 gb of material. the four excalidraw boards alone are about 220 mb. around them are the lyric drafts, vocal mixes, reference audio, scene canvases, image experiments, and multiple full renders. party_v7.mp4 is the latest full render recorded there. party_v6.mp4 is the highest-resolution master backup found so far.
this is not a clean production history. it is a pile of wrong faces, duplicate shots, rough drawings, misspelled instructions, and scenes that changed their meaning while i was making them. somehow that pile became a 176-second video.
ilya sutskever, dario amodei, and demis hassabis appear as brief creator-confirmed satirical cameos. i am not asking anyone to squint at stylized pixels and pretend that is proof of who they are. the reckless researcher who turns off safeguards and rushes ahead is my avatar. i made that character me because i did not want the song to point safely at a few other people.
i like the finished video more after seeing the boards again. not because the process was impressive. mostly because it was so obviously uncertain, and the uncertainty is still visible under the paint.
folio 10 of 27 · written in cusco, peru, 29 NOV 2025 · updated 02 AUG 2026