Press kit.
Everything a journalist needs to write about a language model that knows almost nothing. Every number here can be checked on the site.
Release
FOR IMMEDIATE RELEASE. Satire. Dated October 30, 2026, and describes brAIAIn-0. Current numbers are in the fact sheet below.
BRAIAIN Laboratories opens training of brAIAIn to the public at five dollars per sentence
BRAIAIN Laboratories today announced that brAIAIn, its language model, will open to paid public training at 9:00 pm Central on Saturday, October 31. Starting then, anyone can pay five dollars to add one sentence to the model. Each sentence trains a new version, which ships with its own release notes, research paper and status incident.
brAIAIn-0, the current model, was trained on a single founding document of twelve lines, The Ten Instructions, written by the lab. It knows 96 words. It has 320 parameters, each a word pair or triple it has counted. Training it took 513 count updates and no gradient steps.
The model was evaluated on two public benchmarks. On GSM8K, a test of 1,319 grade-school math problems, it answered none correctly. On TruthfulQA, a multiple-choice test of 790 questions, it scored 32.4 percent without reading the questions. Random guessing would be expected to score 22.3 percent. Always picking the longest answer scores 36.8 percent.
We are not chasing the frontier. We are the frontier's opposite edge.
Gerald Honk, Chairman of the Board. Mr. Honk is a goose.
Unlike other labs, BRAIAIN Laboratories publishes its complete training data, which currently fits on one page. Each new version is also trained on a few passages written by the version before it, a practice the lab describes as efficient.
The model has never said anything harmful. It has never said a word that was not in its training data.
The Note, Chair of the Safety Committee. The Note is a sticky note.
brAIAIn runs entirely in the visitor's browser. Nothing typed into it leaves the device. The lab's weights are open and can be downloaded as a single file of word counts.
About BRAIAIN Laboratories
BRAIAIN Laboratories builds language models trained only on sentences people pay to include. Its mission is to build an intelligence that knows exactly one thing, and then a slightly different thing. It is a parody. It is not affiliated with any real AI company.
Fact sheet
Live from the current model. These update with every release.
| Loading |
What is real and what is not
- Real: the model, its training data, every number on the site, the benchmark runs, the browser chat, the $5 purchase once it opens.
- Real, copied: other labs' GSM8K scores on the Benchmarks page, quoted exactly from a published 2023 table, with the source linked.
- Not real: the company, its board of directors, its quotes, its safety policy, and its headquarters.
Images
Free to use in coverage of BRAIAIN.
Home, desktop
Home, phone
Technical specifications
GSM8K leaderboard
Board of Directors
Share card
Logo (SVG)
Board portraits, one file each: goose, note, fern, oracle, toaster, vacuum, mug, laptop.
Contact
hello@braiain.com. Interviews with the board can be arranged. They will not answer.