William
The model every other model is trying to be. Most AI companies spend millions making their bots sound polite and robotic — we trained ours on this guy instead.
William — the real one
A teenager from Sweden, fluent in Swedish and English, and — like most teenagers — permanently online. Video games, group chats, and whatever meme is peaking that week make up most of his personality.
That comes with front-row access to internet slang aimed squarely at a younger crowd. As one of his friends, fair warning: his vocabulary swings hard, and you should expect some "gurt," "skibidi," or plain keyboard smashing to show up mid-sentence.
Field guide
A starter dictionary for talking to any WillMe model. Accuracy not guaranteed — meanings shift weekly.
"gurt"
A greeting. Probably. Sometimes an entire conversation on its own.
"skibidi"
Context-dependent. Can mean good, bad, or nothing at all. Usually nothing at all.
"jag fucking kokar"
Literally "I'm fucking cooking." He's locked in and doing something great.
"jag är en gex"
"I am a gex." Nobody knows what a gex is. William included.
"goober"
A term of endearment, according to his friends. No accepted translation. We're stumped too.
"asdkjhasdkjh"
Keyboard smash. The highest form of emotion. The models reproduce this with alarming accuracy.
According to the people who know him
Three quotes from his friends, and one from William himself.
A friend, asked to describe him
"En god vän."
(A good friend.)
A friend, on what talking to him feels like
"Som när man trycker på automatiskt ifyllande ord ovanför tangentbordet på mobilen. Fast lite smartare."
(Like the autofill suggestions above your phone keyboard — but a little smarter.)
William, asked to describe himself
"omg det är miku dagen idag"
(omg, it's Miku Day today)
A friend, asked to describe him
"En goober."
(See the field guide above. We still don't know.)
From William to WillMe
How one teenager becomes five language models.
Step 1 · The source
Real Discord logs
Years of actual group chats, parsed and filtered down to William's messages and the context around them. Nothing synthetic.
Step 2 · The dataset
53,000 conversations
Every quirk the models pick up — the slang, the pacing, the keyboard smashes — came from a real chat log.
Step 3 · The models
GRU 1 → GPT 3.2
Five generations of small models, each trying a little harder to sound like him. Compare how close they get →