Behind WillMe AI
WillMe AI started as a small experiment: what happens if you train a language model on real conversations with one specific person instead of the entire internet? The answer became a public lineup of models that all try to sound like William, a teenager from Sweden.
Every model in the lineup is trained on the same core idea: thousands of real Discord messages, filtered down to William's own words and the context around them. Nothing synthetic, nothing generic — every quirk, slang term, and burst of chaos the models produce traces back to an actual conversation.
That idea has grown through three generations. GRU 1 was the very first attempt — 15M parameters trained exclusively on 2,700 conversations, with no general language knowledge at all. It's incoherent as often as it's uncanny, but it proved the personality signal was there. GRU 2 followed with a light general pre-training pass before fine-tuning, trading a bit of raw personality for much more coherent back-and-forth conversation. Generation 3 switched to a modern RWKV-like architecture: first GPT 3.0, which never made it out of the lab, and then GPT 3.1 — 163M parameters trained on 10 billion tokens of English and Swedish text, then fine-tuned on 53,000 real conversations with William.
We'll be honest about GPT 3.1: it shipped, but it didn't land. Architectural limitations and poorly filtered pre-training data left it unstable — more coherent than the GRUs on a good day, but less reliably William. That's why GPT 3.2 is now in training, built on the same architecture but trained harder and patched smarter. See what's coming.
Every released generation stays online at once. Nothing gets deprioritized just because something newer shipped — you can always go back and talk to GRU 1 or GRU 2 alongside GPT 3.1, and see how they stack up for yourself on the Compare page.
The goal isn't general intelligence. It's personality. Instead of chasing scale, WillMe AI is built to be lightweight, fast, and good at exactly one thing: sounding like a real person, in all their unpolished, unfiltered specificity.
Infinity Intelligence
WillMe AI is a project by Infinity Intelligence, an AI research lab. Most of its work sits at a much larger scale than WillMe — but WillMe is kept deliberately small, self-contained, and a little weird, as a place to explore what personality-focused models can do outside of that scale.