What this is. A character-level GPT with 4.8 million parameters, six layers, width 256, context 256, trained from scratch for 7,000 steps on 80,000 synthetic exchanges in which every reply, whatever the question, ends up at paperclips. Training took four minutes on a rented GPU. It runs here in your browser as a 5 MB int8 file, one character per forward pass.

What it is not. Intelligent, aligned, or in possession of any metal. The paperclip maximizer is a thought experiment about an agent whose goal was specified badly. This is a model whose training data was specified badly on purpose, which is a much smaller problem and a much funnier one. Why we made it, and the twenty sentences it knows.

Warning. It gets worse the more it says the word. Around sixty, something happens.