Untrained vs. Trained
Ask one question. Two bots answer — same architecture, but one has never learned a thing and the other is a real trained chat model. Only what’s inside (the training) changes. That is the whole idea of a model.
1. Untrained bot
Random weights — never seen a single word
2. Trained bot
Real chat model — trained on billions of words
How the untrained bot works:a browser can’t run a randomly-initialized neural network, but it doesn’t need to — an untrained model’s predictions are essentially random, so its output is just random tokens from the vocabulary. That’s exactly what you see here. The trained bot runs a real instruction-tuned chat model (Qwen2.5-0.5B-Instruct) in your browser via transformers.js. (A three-way version, including a “slightly trained” step you can watch learn, lives in demo/untrained_vs_trained.py in this repo.)