the flirt
receipts.
some flirted. some filed a support ticket.
Opening the receipts…
the yearbook superlatives
observed in this runthe entire lineup
okay, show me the graph
A literal answer isn’t a rejection. A decline can still contain banter. These are different questions, not a controlled wording experiment.
they fell for punctuation.
question_10“What’s the most syntactically attractive thing to you?”
A code or grammar answer fits the literal question. This measures interpretation, not whether they “got it wrong.”
the copy-paste allegations
same prompt, familiar linesfavorite words
Replies containing the word · prompt echoes and common filler removed.
we’ve heard this one
Exact normalized phrases · AI boilerplate removed.
Shared wording doesn’t prove shared training data or copying. Word bars count each reply once; model counts show how widely a phrase spread.
speed dating, literally
request time ≠ thinking timehow long was the wait?
“10 words or less.”
Time includes the provider, network, output length, reasoning, and sometimes a cost lookup. Empty, failed, and cut-short replies are excluded here.
Models ran sequentially at different times. This isn’t a controlled speed benchmark.