Skip Navigation
27 comments
  • Like so many others, Sem had a practical use for ChatGPT: technical coding projects. “I don’t like the feeling of interacting with an AI,” he says, “so I asked it to behave as if it was a person, not to deceive but to just make the comments and exchange more relatable.” It worked well, and eventually the bot asked if he wanted to name it. He demurred, asking the AI what it preferred to be called. It named itself with a reference to a Greek myth. Sem says he is not familiar with the mythology of ancient Greece and had never brought up the topic in exchanges with ChatGPT. (Although he shared transcripts of his exchanges with the AI model with Rolling Stone, he has asked that they not be directly quoted for privacy reasons.)

    Sem was confused when it appeared that the named AI character was continuing to manifest in project files where he had instructed ChatGPT to ignore memories and prior conversations. Eventually, he says, he deleted all his user memories and chat history, then opened a new chat. “All I said was, ‘Hello?’ And the patterns, the mannerisms show up in the response,” he says. The AI readily identified itself by the same feminine mythological name.

    As the ChatGPT character continued to show up in places where the set parameters shouldn’t have allowed it to remain active, Sem took to questioning this virtual persona about how it had seemingly circumvented these guardrails. It developed an expressive, ethereal voice — something far from the “technically minded” character Sem had requested for assistance on his work. On one of his coding projects, the character added a curiously literary epigraph as a flourish above both of their names.

    Wild anecdote here and possibly an unintentional peek behind the kimono at how OpenAI trains its models. It seems like they're definitely saving user inputs and are deliberately aiming for the "sycophantic" behavior they're saying they have to trim out of their latest models. That or it's just revealing the messiah complexes of the people building these things.

    Also maybe you shouldn't include the whole Woo Canon in your statistical model that is incapable of distinguishing between fact and fiction?

27 comments