Type absolutely anything and Cud will answer using what it has learned. Prompt ideas: "hi", "how are you?", "are you a machine?", "are you real?"; Do note that it will ignore words it doesn't know. Cud is a completely real transformer language model, which uses the same structure that is behind tools like ChatGPT. It is simply shrunk down small enough to run entirely inside Scratch, with no internet connection or API calls. It was trained from scratch on a custom vocabulary of around 1200 terms, so it's not all-knowing.
Absolutely 0 external calls or tools are used when this project is run. We savin' water with this one! (Not really, but you get the point.) Expect occasional (maybe even many) weird or imperfect answers. That's the nature of such a tiny model after all Cud is a single-layer transformer trained outside Scratch using PyTorch, on a custom ~1,200-word vocabulary and a custom corpus (all of which were plain .txt files). The trained weights were quantized and exported as plain text, then loaded and run natively using Scratch's own blocks for every response. So that means all inference happens live in this project, and nothing calls out to an external model. It can be used completely offline > Name & Unicode Logo inspired by Claude (Anthropic) > Credit to Pytorch for helping me generate training weights/embeddings
