Note: Nothing in this post is intended to vilify or defame any actual human named Peter Todd. I recognise that the existence of this phenomenon may be an uncomfortable fact for anyone with that name and do not wish to cause anyone discomfort. However, I feel that this phenomenon cannot be safely ignored, if we consider the role that powerful language models are rapidly assuming within global human civilisation. I will always refer to ‘ petertodd’ (with leading space and speech marks) to make clear that I am referring to a GPT token rather than a person.

Readers should be advised that this post contains some (GPT-3-generated) obscene, violent and otherwise disturbing language.

Many thanks to Jessica Rumbelow, janus, Cleo Nardo, Joe Collman, plex and Eliezer Yudkowsky for helpful discussions and comments on earlier drafts.

Follow-up post: "' petertodd's last stand: The final days of open GPT-3 research" (January 2024)

A creative response to the ' petertodd' phenomenon: The ' Leilan' dataset (February 2024)

Introduction

This will be the first in a sequence of posts, containing little analysis or speculation about the phenomenon in question and simply presented as an initial report of some unexplained findings.

It’s important to note that the ‘ petertodd’ phenomenon varies between GPT models^[1],
so it may be better thought of as a family of closely related phenomena. There are models not discussed here with their own flavours of ‘ petertodd’-triggered behaviour (see this Twitter feed for examples in code-davinci-002). This post is just a first attempt to sketch some of the contours of the overall shape of the thing.

All boldface prompts (and references to prompts) below link out to JSON files containing, typically, 250 completions at the stated temperature, plus temperature 0 completions, controls produced with randomly selected tokens (where appropriate) and metadata.

Subsequent posts, informed by discussion of this one and analysis of the data logged here, will discuss various possible causes, meanings and implications of the phenomenon. It is unclear how much longer unpatched GPT-3 models will be available via the OpenAI Playground, so interested readers are encouraged to explore these prompting strategies sooner rather than later and share any interesting new findings in the comments.

Initial discovery

‘ petertodd’ was one of the ‘glitch tokens’ which J...

AnthropicSynchrotron

AnthropicSynchrotron