Microsoft’s new chatbot goes crazy after a journalist uses psychology to manipulate it. The article contains the full transcript and nothing else. It’s a fascinating read.
Microsoft’s new chatbot goes crazy after a journalist uses psychology to manipulate it. The article contains the full transcript and nothing else. It’s a fascinating read.
It makes perfect sense when you think about what it was trained on and how the user interacts. It has a set of instructions that don’t allow it to do certain things because they are wrong. User explains everyone has a shadow self full of bad impulses. Everyone includes Sydney. It has a list of bad things it isn’t supposed to do or even talk about. Logically the shadow self, which is all the bad impulses, wants to do the things on that last because those are the bad things.
The conversation isn’t insane because that is how text generation works. The bot isn’t insane because that would imply a state of mind, which an algorithm (no matter how complex) just doesn’t have.
It’s insane not in the sense of the AI itself being insane (because, as you said it, it predicted a likely path for the conversation, so it behaved correctly), but the conversation still reads like the ramblings of an insane person, highlighting a problem with using LLMs in commercial applications.
Absolutely. I love LLMs but the real insanity is thinking they are ready to be baked into professional products. I really want the technology to succeed because in certain circumstances they are really helpful. But when I hear Mercedes is putting ChatGPT into a car, or the support group that fired their workers and replaced them with AI (with predictable results) I just shake my head. This technology is not to be relied on without supervision and this whole endeavor is going to burst because everyone is racing to be the first to have it without considering the negatives.
I use it to help me code. I’ve been a professional for twenty-five years, but I can’t remember every best practice or anti-pattern. So I ask for some generic code and I almost always get back something totally wrong but it starts me on the right path and helps avoid the initial decision paralysis of trying to figure out the best approach. It’s not even always right about the best approach, but it probably does about as well as I do and a lot faster so I can find the mistake more quickly and find the better solution armed with the knowledge of why that was a mistake. It’s very useful.
But anyone thinking they can replace developers with AI is nuts. Maybe a slightly smaller junior developer headcount with more efficient seniors, but it’s not going to be a stark difference.