Four real generations from a live model (Qwen3-4B), frozen so you can explore them at your own pace instead of watching them fly by.
A textbook explanation, generated with unusually high confidence throughout.
Watch formatting and plain prose compete for the same sentence, token by token.
Answered wrong — and the model's own confidence collapses right where it starts guessing.
Wrong again, despite the correct answer being the model's own favorite candidate.