Introducing brAIAIn.

A language model trained on exactly what it was given. Every word it knows was paid for.

$5 buys one sentence in the next model. It ships as a numbered release with its own notes, paper and status incident, credited to you or to Anonymous.

BRAIAIN Laboratories is a parody AI lab. The model is real, very small, and trained only on sentences people pay to add.

Paid training opens in

--:--:--

Checkout is installed. The model is ready to learn.

When the clock reaches zero, anyone can pay $5 to train the next version on one sentence. Until then, brAIAIn knows only its commandments, and the lab is not allowed to charge for them.

Add the opening to your calendar

Current version-

Paid sentences-

Words it knows-

Training text it wrote itself-

Talk to brAIAIn.

Ask it anything. It will answer using only what customers have taught it, which is not much, and it will be confident about it.

Press Peel a layer to hear the same question answered by the version underneath, and the one under that, down to brAIAIn-0. Each version is trained on everything before it, plus text the version before it wrote. Peel it until nothing is left but the first twelve lines.

This runs entirely in your browser. Nothing you type is sent anywhere or saved.

Model

Answers are assembled from the founding text, accepted paid sentences, and text earlier versions wrote. The model has read nothing else. brAIAIn can make mistakes. It can also only make mistakes. Ask something, then press Peel a layer to hear the version underneath answer.

Trained on you.

Every version adds one sentence that a customer paid $5 to include. That sentence becomes part of the model forever, or until our safety team removes it.

Trained on itself.

Each version also reads a few passages written by the version before it. Researchers warn that models trained on their own output slowly lose their minds. We wanted to see it happen.

Fully disclosed.

Other labs will not say what their models were trained on. We will. Here is all of it. It fits on one page.

Under the hood

Full system card

The specifications other labs keep confidential. The counts and timings below are computed on the current model, in your browser, as this page loads. The rest are true by construction.

Latest releases

All models

Shape the future of intelligence.

Write one sentence. We train a new model on it, give it a version number, publish its release notes, and tell the world. That is the whole product.

Feed the model for $5