Experiment Explanation:
This simulator transforms raw language model tokens into a human-like typing animation,
revealing how AI models "think" through their word choices in real-time.
Mode: - Fixed Speed: Consistent typing pace with random variance around the base
speed
- Confidence-based: Typing speed adapts to the model's confidence—high
probability tokens type faster, while uncertain choices slow down (based on surprisal)
Correction Threshold: When set above 0, if the second-best token is close enough
to the top choice (within the threshold ratio), the simulator will type the runner-up first,
pause to "reconsider," then backspace and type the correct token. This visualizes the model's
uncertainty between competing options.
Speed & Variance: Control the base typing rate and how much randomness is added
to simulate natural human typing rhythms. The simulator also adds extra pauses at word boundaries
and punctuation marks.