The introduction of the transformer architecture has marked an inflection point in the ever-broadening of the automation realm. For the better part of a year, we have seen how more and more powerful models beat each other into giving us valuable outputs. Be it text, images, code, or movies, LLMs do a lot.
Yet, why is it that when I chat with an LLM, it feels as if I am talking with a demented savant? By demented, I mean a person with dementia, and by savant, the standard definition a person with an almost infinite skill set of a specific topic. The combination is somehow uncommon but not atypical. Often, savants are neurodivergent in other ways, and thus, to discuss with them at length, one needs to build some empathy infrastructure to help them feel at ease. So does chatGPT and their classmates.
Old age is the goal for most of us. If you live to be 90 years old, you won the game of life. At that age, you can eat what you want and enjoy yourself. However, as time goes on, your brain starts to shut off, and as it does, dementia sets in. Dementia, in my understanding, is a catch-all term that describes symptoms of varied types. Be it partial memory loss, confusion, emotional instability, speaking problems, and so on.
If a loved one of yours reached the 9th decade, you know what I mean. You need to use your empathy more when talking to them. To remind them who you are, how you are related to them, and what the interaction will be like: ""Hi Grandma, it's me, Jose, your grandson. I am here to visit you and have coffee. Would you like to have coffee and hang out for a bit?" Miss any of the signposts, and your interaction will go awry.
Media is in love with savants. Eccentric people with a specific superpower, an ability to outperform anyone with their minds and hands. People with eidetic memory, children who play sonatas before entering preschool, or who can outperform computers in randomly specific tasks (e.g., was January 13, 1987, a Tuesday?).
Media loves savants because of their skills but also because of their quirkiness. Seldom will a savant presented that is portrayed as an Ubermensch or as the person who the audience should see as themselves. Think of Sherlock Holmes, whose story, due to its copyright elapsing a long time ago, has had myriads of replications on TV. Every Sherlock is a troubled soul. We see him through the condescending lens of Watson, who admires and pities the savant.
Be it due to my studying physics and engineering, realms filled with people who know math and have good observation skills about the physical (i,e. nonsocial) world), or my undiagnosed neural atypicality, I have always loved neurodivergent people. Well-adjusted vanilla humans whose goals and beliefs align with what society puts in their way (e.g., money, power...) trigger my emotional gag reflex.
No, I have always preferred troubled people. People with insane ideas and the willpower to push them through against the express objection of the universe. People who survived some trauma and want to prove to the world that they are more than their past experiences. People who grew up defined by a proficiency they had but wanted to conquer different skills and explore new shores.
Large language models, such as chatPGT, LLAMA, etc., implement a chat based UI. While one can simply ask queries in Python, the vanilla UI is a smart assistant that politely answer questions. However, this assistant is a bit demented. Most of my chats start with me explaining who I am and what I know, the length of the answers I expect, and the level of interest I have on scientific evidence versus writing artistry. In doing this, I feel as I felt when I started my studies, I engage my curiosity and try to learn patterns that let me interact with my LLM of choice.
In Python, I can bundle this as a simple header to my questions, as a way to simplify my work, a wrapper that makes the LLM to talk to me as i want. In Python, I can do it more simply by adapting the settings of the outputs and providing introductory sentences to each new chat thread. Yet, when I forget, and.I open for example my Bing assistant, the whole process needs to start again. I need to remind my demented savant who I am and how they should reply.
Nothing I said so far is specially interesting. But I gets me to think how I can be more empathetic to a LLM than I am to a colleague and show more care and appreciation than I give to someone that serves my food or checks me in a hotel. I am deeply curious whenever a new LLM comes out. I wonder what its quirks are (remember the weird hands from Stable Diffussion v2?). I notice when OpenAI lowers the system's capabilities. I care about how these systems perform.
I might not be in love with my LLM but this care is troubling me deeply. We live in a world in flux, where crises abound and I spend my time understanding how a machine behaves instead of empathising with my fellow human beings. I guess this is not new, I am still the engineer who barely remember the people who taught him how to think and instead remembers in detail the chapters in which the wonders of the universe were explained to him. I guess, what changes is not my lack of intrinsic empathy but the fact that I am empathising and I find it strange and this strangeness makes me want to justify myself.