In 2014 I helped to create an art project based around a bunch of large LED eyes. These eyes were very simple and only consisted of 63 pixels each. When I displayed them I was often asked how I made them see and actually follow people around.

I didn’t.

Instead I had studied how human eyes work. I read about what determines our rate of blinking and how our pupils move as we examine things. I learned about saccades, smooth pursuit, and fixation. I found simple ways of simulating these movements with a random walk of a virtual pupil. Well, not truly random - it was fed from a psuedo random number generator (PRNG).

Without specialized hardware, computers can’t make truly random numbers. Instead, pseudo randomness is generated from complexity. A shell game. A digital version of “find the queen”. It’s not actully random, but we can’t tell the difference.

This same method of using a PRNG to make something seem life-like is at the core of modern generative AI. The AI companies have a setting for their inference systems they call “temperature”. This is just the range of values their PRNG will toss onto the predictions coming out of their model.

Without the PRNG, if you asked a large language model the same question you’d always get the exact same answer. Each word in the response would be the most likely one to come next. It wouldn’t feel nearly as “alive” as these models can feel.

This then gets to the heart of why I think generative AI is bad at so many of tasks it’s being thrust into. Creativity and pseudo randomness are not the same thing. A writer chooses their next word based on their entire life experience. An LLM is doing vast amounts of math moving numbers in thousands of dimensions to make the best possible prediction for the next word and then purposely messing it up a bit.

Generative AI is a pattern matching tool, a very powerful one that we’re just starting to learn how to use. However, trying to use generative AI to replace human creativity isn’t even wrong, it’s fundamentally a category error.

The fact that this is how AI companies chose to market their tools blows my mind. These companies have chosen to introduce generative AI to us by dressing it in a dancing skin suit and hoping we were more amazed than horrified.