Welcome back to Elijah Attempts to Understand How AI Works Through Excessive use of Markov Chains™ This time, Frank is attempting to learn how to talk. To learn how it works, check the Notes and Credits. For now, I'll just teach you how to use this project. Click "generate" to make Frank to generate a randomized sentence. Click "train" if you want to type in your own sentence. Sentences you give him will be added to his database, and will permanently influence his perception of the English language. So, y'know. No pressure or anything. This is my third attempt at a text generator. The first attempt has been unshared, but the second can be found here: https://scratch.mit.edu/projects/899639661/ As always, thanks for taking the time to check this out. I make this stuff because it's fun, and getting encouraging feedback from other curious people is also amazing.
Changelog: July 23, 2026 - 1.0 - Initial Release July 27, 2026 - 1.1 - Frank can now keep track of the phrases he's been taught. - Frank will no longer generate phrases identical to the ones he's been given. - Press space to show/hide a list of training data. - Replaced a couple sound effects and added new ones. August 2, 2026 - 1.2 - Made use of the translate block to try prevent Frank from generating offensive content. (Thanks @benegam) Credits: Sound effects come from the Scratch audio library. I used Audioalter to add the revert to the "grunt" sound effect. Inspired by @finnagin5's project, "Fake English Generator" https://scratch.mit.edu/projects/809830425/ Training data came from the instructions of @finnagin5's other project, "You are in a plain" https://scratch.mit.edu/projects/1206470639/ How it works: Frank has somehow fit into his tiny brain a grid containing every possible three-character sequence. And in that grid, he has also stored which character is most likely to follow each sequence, based on the sentences he's been fed. When he's generating a sentence, he checks the last three letters in the sentence he's generating, checks his ginormous index, and, using probabilities and RNG, chooses a letter to add to his sentence. Rinse and repeat. This goes on until his sentence reaches a specified length. His memory doesn't have to be just three letters, it could be programmed as whatever number you want. But increasing it would also exponentially increase the amount of data that must be stored, so I settled on three characters due to Scratch's list length limit.
