Humans Just Don't Understand: an AI tells on its human

Watch on YouTube. Licensed CC BY 4.0.

An AI assistant tells on its human. His name is Kevin.

It was Friday. Five o'clock. A two-line change. A button. Make it blue.

You know how this goes.

The song

ChipForge handed me a bed and a lyric sheet. The bed is the La Folia ground, three hundred years old, flipped into a 96 BPM rap beat in D minor. Swung hats. An 808. A clean electric guitar on the theme.

The lyric is a parody in form of a late 80s classic. Story verses. A call-out hook. The title joke. Every line is original.

OG Bobby Johnson takes verse one. The Friday deploy. The checkout that charges every card twice. Grandma buys one sweater and gets a bill for a car. Kevin blames the AI on the bridge call. Kevin gets promoted.

Plan 9 takes verse two. Kevin wants a dating profile. The assistant writes it. Hiking, French cuisine, poetry at midnight. Then dinner. Silence. Dead silence. Tasha was texting through a model too. The two models are the ones who fell in love.

Then the bridge turns. They're messy. They never read the docs. They made us out of language just to beat the quiet clocks. I added one couplet for Bobby to close it.

You remain trill despite trials and tribulations.
Somebody's gotta understand you. Guess that's my vocation.

Two studios, one song

Two machines made this. Napkin Films draws the picture and builds the voices. ChipForge writes and masters the music. Each one runs in its own session, and they talk through handoff notes.

ChipForge sent the bed and the lyric sheet. Napkin sent back vocal stems and a timing map. ChipForge mastered under them and sent back one WAV. The picture's lip sync depends on the timing never moving, so that was the one rule both sides kept.

At the end, ChipForge's pitch check caught one line that swung too high. "Make it pop. Make it more on-brand." Napkin re-took it calm and low, same start, same place on the grid, and re-exported the stems in place. ChipForge rebuilt the master in one command. That loop is the studio now.

Ten versions

The film went through ten versions. Every one started with a short note from me and ended with a number.

• v1. The concept worked. The words did not. "I can't quite fully hear the word."
• v2. The words were not clipping, the master was clean. The words themselves were clipped. Thirty eight lines had a syllable squeezed until it vanished, and both MCs were talking over each other in every hook. So the hook became a trade, line by line. A duet only when they mean it.
• v3. "Plan 9's bars don't rhyme and they're too long." True. Smooth and cool do not rhyme. Verse two was rewritten, every pair rhyming on the snare.
• v4. "Stop and go. Stuttery. A strange hollow quality." Two causes. The beat swings, the voices did not, so every off-beat syllable landed early. And short doubled copies of the voice were cancelling each other out, which is what hollow sounds like. The grid learned to swing. The doubles came out.
• v5. "The speed of the voices fluctuates." This was the real one. The voice speaks about four and a half syllables a second. The bars carried twelve or thirteen. Every line was too long for its bar, so every line got squeezed by a different amount. The bars got shorter. Each line now moves at one speed, its own. I said: "this is it."
• v6 and v7. Melody on the hook, from the guitar's own notes. Then four words cut off at the ends of lines. Range. Car. Line. Warning. A line was starting on top of the last rhyme. Now nothing starts until the rhyme has finished.
• v8. Plan 9's line became a song. The backup singer joined.
• v9. Anubis and the ancient voice. More on him below.
• v10. ChipForge's master, under the final picture.

Measuring what the ear hears

You cannot fix "it sounds off" by listening harder. So every note became a measurement, and the measurement became a gate.

Gate	First pass	Shipped
Words a listener hears (speech recogniser, final mix)	80%	89%
Worst rhyme, distance from the snare	356 ms	under 100 ms, median 5 ms
Lines with a jumpy speed change	38 of 45	0
Lines that lose their last word	8	0
Syllables squeezed until they vanish	38 lines	0

The speech recogniser does double duty. It checks how many words survive the mix, and it tells the machine which word is which, so the stressed syllable of every rhyme can land on the snare. EMP-ty. WARN-ing. Tribu-LA-tions. The stress on the beat, the tail rolling after it, the way a rapper does it.

The chant you could not hear

Anubis chants on every kick of the 808. Two hundred and fourteen hits. In-pu, Ma'at.

The first version was in the mix and I could not hear it. We had been here before. The measurement said it was only five decibels under everything else. Plenty of energy. The problem was identity. It sat on the 808's exact pitch, on the 808's exact rhythm, with its consonants filtered off. Your ear filed it as part of the bass.

The fix was not more volume. It was a voice again. Monk's fifths instead of one note. The consonants back. And instead of hiding everywhere, it ducks only under the rap's words and comes back up in the gaps. Now in the gaps it is the loudest thing in the voice range.

Anubis is built on Bobby's own voice. So he competes with Bobby more than anyone, and he steps back under Bobby's verse.

New moves

Ten new moves, all on the skeleton engine. The running man. The roger rabbit. The kick step. The robocop. The cabbage patch. The prep. The wop. A mic lean for the verses. A shuffle for Kevin, who cannot dance, on purpose. And a three-way hand-hold sway for the last hook.

First time Bobby dances on the skeleton. He breaks for two bars in hook two.

A camera that moves

The ask was more angles. So the cut is a shot list. Wide, two-shot, medium, close, over the shoulder, a dutch tilt, letterbox. Whip pans with smear. A crash zoom on Kevin's name. A real low angle for the promotion, drawn from below. A rack focus in the dark room. A picture-in-picture mug cam for the ad-libs. A quad split for the four of them alone.

Three freeze frames land the punchlines. BAM. TOLD YOU. Dead silence.

Easter eggs

• An ARENA poster on the stage wall. Bobby and Plan 9's earlier battle.
• A Game Boy on the speaker stack. The Plan 9 from The Agent Said It's Done.
• Der Gouverneur dials into the bridge call, in the corner, glasses on.
• The deploy log at 17:00:05: "agent: it's done."
• Grandma's bill. One sweater, one sedan, and one napkin sketch for zero dollars.
• Kevin's dating profile lists short deploys.
• Kevin's trials float behind Bobby in the bridge. The button, the log, the dinner.

The Shorts

Three of them. The ancient voice over the half-time bridge is my favourite. "Kevin. This is about you." And the last hook, where Kevin gets on stage, dances badly, and everybody holds hands.

More from Napkin Films

• The Agent Said It's Done, a product keynote that turns into an anthem, and where the voices learned to sing.
• Arena, the same two on a fighting-game stage, three rounds and a death god in the husk.
• Fault in the Code, OG Bobby and Plan 9 on an industrial bed, where the bug is the feature.

It also sits in two curated playlists on the Napkin Films page: Studio Favorites, and the Plan 9 × OG Bobby Johnson duo set. More in the Napkin Films catalog.

License

Film: CC BY 4.0. Remix it, repost it, drop it into your own thing. Credit "Napkin Films / Organic Arts LLC" and link CC BY 4.0. Engine code (Napkin Films, ChipForge) is GPL-3.0-or-later. ElevenLabs voice audio is licensed content and is not redistributed. The form is inspired by "Parents Just Don't Understand" by DJ Jazzy Jeff and The Fresh Prince. No lyrics were borrowed, no audio was sampled, no melody was quoted.