Anomalous Tokens: a mysterious failure mode for GPT

lesswrong.com

11 pointsmquander2 comments

This is fascinating!

It would be a neat party trick to be able to elicit some of these non-sequitur responses from ChatGPT, but I guess there's no way to access the anomalous tokens in that particular case, short of hands-on access to the implementation?

ChatGPT displays many of the exact same failures! (But presumably deriving them without already knowing them would require access to the model, yes.)

Reply on news.ycombinator.com