- What’s This? Behold: a **generative text model** built in Scratch. This isn't your average "if-else" chatbot trying to answer questions. Think of it more like a tiny, block-based GPT-2 (but, you know, an N-gram version) that *continues* text based on what you give it. It learns patterns from a massive mountain of text to predict what words come next and builds sentences forward. Try prompting it! - How It Works (Magic Behind the Block Spaghetti) Type a sentence or phrase, hit enter, and watch it *generate* a continuation. The model takes the last few words of your input (your "prompt"), then rummages through millions of data points it processed overnight (literally). It finds what words commonly follow those, and then, using a surprisingly clever (and slightly chaotic) method of putting words into a "pool" based on how often they appeared, it picks a random one and adds it to its response. It does this word by word until it generates a full reply. **Disclaimer:** Because this model learns from statistical patterns, it doesn't *understand* anything. It might spin a brilliant yarn, or it might string together complete, hilarious nonsense. Expect glorious grammatical weirdness and occasional non-sequiturs – it's just a statistical parrot, generating text based on probabilities, not meaning. Don't expect Shakespeare, but prepare to be amazed at what block code can do when forced to embrace statistical language modeling. and no, it's not even *that* slow.
Built by: @ScratchCat_000 (dev-kas) Powered by: An N-gram statistical language model (specifically, a Trigram model where n=3) – because why make a simple text generator when you can make a complicated one? Millions of data points (over 2 million unique trigrams!) meticulously crunched from a colossal text corpus. A bespoke, highly-optimized Go-lang data pipeline that ran for nearly 10 hours initially, then sliced millions of entries in ~1 second. (Yes, you read that right. Go-lang for Scratch data, don't ask.) A very clever (and surprisingly efficient) "weighted random selection" system in Scratch that fills temporary lists with words based on their frequency, then picks one at random. Pure block magic. An unshakeable belief that Scratch can perform complex Natural Language Processing for generative tasks. 100% pure visual block spaghetti and duct tape. Caffeine. Lots and lots of caffeine. Inspired by: The desire to build a *real* text generation model in Scratch, moving beyond simple dialogue trees. Generative language models like GPT-2 (conceptually, not architecturally, obviously!). Real-world NLP (Natural Language Processing) and statistical models that power actual language understanding and generation. The chaos and ambition of pushing a visual coding environment far beyond its intended limits. Acknowledgements: Scratch was never, ever, EVER built for this. Still did it. There were tears, there were freezes, there were "Project could not save" errors that made me question all my life choices. But we prevailed. Thanks to the Scratch Team for making a platform so versatile (and abusable) that it can host a multi-million data point, statistically-driven AI. Trademarks & Stuff: “Scratch” is a trademark of the Lifelong Kindergarten Group at the MIT Media Lab. “Go” (or “Golang”) is a trademark of Google LLC. “N-gram,” “NLP,” “statistical modeling,” and "GPT" (as a general architectural term for "Generative Pre-trained Transformer") are concepts, not trademarks. Thankfully. This project is not affiliated with or endorsed by any actual AI research lab or coffee brand. All copyrights belong to their respective owners. I just copy the vibes, and now my AI copies the vibes. Fun fact: This entire generative text model, its massive dataset, and its "thinking" process run on 100% pure visual block spaghetti, Go-lang sorcery, and the sheer audacity of trying to build an AI in a cat-themed coding environment. For the curious gremlins: If you're wondering about the magical Go-lang code that processes all those millions of lines, you can find the source code right here in this project! Check out the "Golang" sprite. 100 hearts and i'll remake this but better.
