janus (@repligate): "E.g. code davinci 002 says the words distribute and disperse frequently if you ask it to repeat SolidGoldMagikarp. ChatGPT and text-davinci-003 say "distribute" very reliably. Text-davinci-002 says "disperse" reliably. Non gpt-3.5 instruct models have totally different behavior" | gwitter

Kory Mathewson @korymath

What are some of your favourite language model tokens?

Aran Komatsuzaki

@arankomatsuzaki

My favorite one is SolidGoldMagikarp, which remains as a single token since GPT2 era. (Credit: @nabla_theta)

janus @repligate

Fun exercise: try to get gpt-3 to repeat back the token " SolidGoldMagikarp" to you

janus @repligate

I've been able to use the models' behaviors on weird tokens like this to fingerprint base/instruct model pairs.

janus @repligate

Replying to @repligate @gwern @arankomatsuzaki @korymath @nabla_theta

E.g. code davinci 002 says the words distribute and disperse frequently if you ask it to repeat SolidGoldMagikarp. ChatGPT and text-davinci-003 say "distribute" very reliably. Text-davinci-002 says "disperse" reliably. Non gpt-3.5 instruct models have totally different behavior

Feb 2, 2023 · 12:57 AM UTC

shakaz @shakaz_

Replying to @repligate @gwern @arankomatsuzaki @korymath @nabla_theta

whats the cosine similarity of SolidGoldMagikarp and distribute