version 2: https://scratch.mit.edu/projects/1315065042/ instructions: Wait for it to "train" (The "estimated seconds ramaining" variable should tell you how much longer it should take.) it will take around 20-60 seconds. and then you will see the output in the "output" variable. you can copy and paste the "full output" list so you can read the full output in a doc or something. REAL AI, but not a very good one It works like a simple Markov chain. The "estimated seconds ramaining" variable should tell you how much longer it should take to start generating. https://turbowarp.org/1242382682 IS MUCH FASTER current data it uses is 10k words generated with Gemini 3.0 flash. PREVIOUSLY, i have used: https://www.gutenberg.org/cache/epub/6692/pg6692.txt i have used around 10x that before, but it kept crashing and took over an hour to "train" on turbowarp. the current data in this should only take around 30-60 seconds on a moderate computer. For me it starts at around 300 words per second but slowly goes down. Now i am using text generated with gemini 3.0 flash havn't made this exactly user friendly. YET if you want to be able to upload your own data, and change the output length, you will have to either go in the code and try to understand it, or just wait for me to finish making stuff. (i'm a bit too lazy to do that rn) what this is: I made a mini "AI" in Scratch. It works like a simple Markov chain. I load a big thing of data, and then this splits it into words (tokens) and then based on which words most commonly follow each other, it outputs another word. This is the data i used generated by Gemini 3.0 flash but i could only use about 350,000 characters of it or the project would keep crashing it actually did crash a few times while i was making it and i had to redo a few things because the auto saving was slow. I'm thinking about adding two word memory, but that would take a lot more data and much more training time this is basically what it has: a single attention head in a single transformer layer, with a context window of 1 and no embeddings. if you want to change the data, just go in the code and change the data variable. only 350 blocks and 1 script. CHANGELOG: Janurary 25 2026: Changed data from gutenburg to ai generated stuff gemini 3.0 flash nov 14 2025: shared project, redid the way it stores which word to output next which makes it train faster and output around 2x faster. made everything decapitalize itself. optimized something that made it around 2000x faster to output. i forgot. added blank sprite. changed dataset. fixed bug. wrote this. nov 13 2025: Started working on project, and made basically everything. idea: make it so that it doesn't "train" for a certain word until it is generated for the first time
