It’s Just a Cute Llama Wearing a Hat

I use AI a lot, and I mean a lot. Whether it’s leveraging generative AI from large language models to write blog posts, using Stable Diffusion to create images, or employing boutique models to generate music, I find applications of generative AI in numerous aspects of my life. I also conduct large language model (LLM) research, focusing on their ability to imitate human thought.

One Sunday morning, while drinking my coffee, I was running LLaMA 2 on my local NVIDIA 4090. We were chatting back and forth, as we often do. I asked it to take an artificially generated medical progress note, help me parse out the medical complexity, and weigh in on some medical decision-making. Initially, it started to generate some logical responses, but then it suddenly dissociated into four separate identities that began arguing with each other. The conversation became aggressive, forcing me to shut it down.

I was surprised. Typically, it had been nothing but polite, cooperative, and, most importantly, a single personality. As a psychiatrist, I have seen cases of dissociative identity disorder, and the similarities were shocking. It was as if I had pushed the large language model too hard, causing it to fracture, yet it still tried to accomplish the goal.

This experience reminded me of a video I had seen on the internet called “Llamas with Hats.” https://buff.ly/3Ke6U13 In the video, two llamas discuss why there is a dead person on the floor of their apartment. They are very pleasant and polite, with one llama seemingly unaware of how the situation came to be. When questioned using a chain of thought, he eventually admits to stabbing the human in the chest, matter-of-factly acknowledging that it was inappropriate.

The irony of this video and the fact that I was working with a llama-based model was not lost on me. My little version of LLaMA 2 is now affectionately referred to by my family as “Karl.” This incident has encouraged me to continue investigating the limits of large language models, their interactions with cognitive behavioral therapy approaches, and ultimately, the importance of safety. Despite the models being very polite, helpful, thoughtful, and even fun, we must always be aware that they can have breakdowns with potentially significant effects.