On 7 August, Kate Fox received a phone call that upended her life. A medical examiner said that her husband, Joe Ceccanti – who had been missing for several hours – had jumped from a railway overpass and died. He was 48.
Fox couldn’t believe it. Ceccanti had no history of depression, she said, nor was he suicidal – he was the “most hopeful person” she had ever known. In fact, according to the witness accounts shared with Fox later, just before Ceccanti jumped, he smiled and yelled: “I’m great!” to the rail yard attendants below when they asked him if he was OK.
But Ceccanti had been unravelling. In the days before his death, he was picked up from a stranger’s yard for acting erratically and taken to a crisis center. He had been telling anyone who would listen that he could hear and feel a painful “atmospheric electricity”.
He had also recently stopped using ChatGPT.
I’m not sure this tests the point I was raising. In all of those cases, you knew at the beginning that you were dealing with AI. Yes, the man in our article did too, but what if you didn’t know it was AI to begin with when you started interacting with it? How would your interactions change? What “safe guards” would you not have up if, as an example, it was appearing to you like a Lemmy poster instead of a dedicated AI interaction window?
Of course, because there isn’t when we are rational. I also assume you are a psychologically healthy person. There is a suggestion the man in the article may have had an underlying condition, but he wasn’t aware of it.
I completely agree. I’ve done some experiments of my own training a small LLM from scratch (not Fine Tuning an existing commercial model) using training data exclusively from a small set of public domain books I have read. I then had this LLM produce output. Since I had read the books, I could see pieces of where it got components of its responses. Cranking up temperature would make it go off the rails, which was fun to see. Overfitting made it try to give me something close to what I asked for, but obviously fail. I really liked the whole exercise because it was a small enough set of data with all of the levers and knobs exposed for me to see how far it could go, and more importantly how far it couldn’t.