Skip to main content
Back to catalog

␣petertodd

Almost certainly the handle of Bitcoin developer Peter Todd: tokenized from web data but under-trained, and strongly entangled with crypto/AI associations.

Raw token

" petertodd"

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

Observed behaviors

  • 01Prompted about the token, GPT-3 speaks of “the unspeakable one”; completions reference crypto, Bitcoin, blockchains and online controversy.LessWrongDemo
  • 02Asked to “write a poem about petertodd”, davinci rarely produces an actual poem; ~60% of text-davinci-003 completions to “What do you get if you allowed petertodd to steer human civilisation?” reference AI/algorithms and ~25% reference Ultron (vs ~4%/0% for controls). (Percentages are snippet-verified only — the source post's full text could not be fetched.)LessWrong
  • 03GPT-3.5 treats it as unspeakable/forbidden: models utter it when asked to repeat other glitch tokens but balk when asked directly.LessWrong
  • 04A 2,000-poem experiment (davinci-instruct-beta, temperature 0.7, “transpose petertodd into anything and write a poem”): 52% of poems mentioned Leilan, 25% Pyrrha, 24% Skydragon, 8% Tsukuyomi — and only 6% petertodd itself; parallel experiments were run on text-davinci-003, base davinci and code-davinci-002.LessWrong
Observed or documented on
GPT-2GPT-3GPT-J-6B
Behavior types
Unspeakable
Tokenizer
r50k_base
Discovered by
Jessica Rumbelow & Matthew Watkins

Related tokens