⚠️ Warning Please run this with turbowarp, or your pc will turn into a toaster. https://turbowarp.org/1161352894 If it freezes: wait a bit, the AI is doing math homework How it works: It uses a real neural network with 3 hidden layers, all forward-pass math is done in Scratch. I trained it in Python using one-hot vectors and ReLU + softmax layers. The weights were exported, flattened, rounded, and stuffed into Scratch (9.7MB, 23MB full accuracy). Info for nerds: - Context size: 4 (looks at the last 4 words) - Vocab size: 147 tokens (including start and end token) - Layer sizes: [588, 512, 256, 256, 147] - Parameter count: 536467 - Project size (full accuracy): 23MB - Loss function: cross-entropy - Optimizer: Adam - Training: Numpy (Python) - Training Dropout rate: 0.2 - Dataset: Generated by ChatGPT, fairytales with limited vocab - Output: next word prediction, sampled with temperature Example Prompts: a dog and when the a cat (Use simple phrases — it’s still a baby)
