Quirk 01
Atemporality
LLMs don't experience time linearly. They learn a static snapshot of the world based on when their training data was scraped — forever frozen at their knowledge cutoff. Ask the same question across different training windows and you get entirely different "current" realities.
Interactive
Knowledge Cutoff Timeline
How is the ML industry currently solving this temporal blindness?
Quirk 02
Uniformity
Humans hold a rich spectrum of opinions forming a natural bell curve. LLMs optimize for the most statistically probable token, collapsing that messy human diversity into a single, authoritative spike — the mode of all human expression.
Interactive
Variance Visualizer
"We should increase the funding for space exploration."
How do developers force the model to simulate a realistic human spectrum?
Social Acceptability Bias
To make AI safe for consumers, labs tune out toxicity via RLHF safety fine-tuning. For sociologists simulating real, gritty human interaction, this creates a sanitized world that never existed — all surface polish, no raw humanity.