␣ertodd

A substring token of ` petertodd`: under-trained in its own right and inheriting the parent token's strange aura.

" ertodd"

␣marks a real space; the symbol itself is not part of the token, and copying preserves it.

Observed behaviors

• 01lsusr found in the comments that context “fills it in”: shown “2+5=ertodd”, the model reads it as “2+5=7” — as if the token carried the arithmetic result.LessWrong
• 02In the same thread mwatkins worked out its tokenization pattern: how ` petertodd` splits in different contexts explains the odd behavior.LessWrong

References

• The ' petertodd' phenomenon — LessWrong
• SGM1 comments: lsusr & mwatkins on ' ertodd' — LessWrong

Related tokens