I’ve been wondering a lot about how good GenAI can ever be. I think I’ve convinced myself that at least in my lifetime, it will never be much more than average. My logic here comes from the training data.
Conceptually, I model the entire AI process as a regression model, but with a lot more tricks. The point of the regression model is to try to predict the general output of a function given a sparse (and potentially noisy) set of samples. For Large Language Models, the function is “Most Likely Next Word,” and it has one input, a vector of all the words it should consider for the next word. The neural network is to predict the most likely next word in that input vector.
Let’s take the best case. There are multiple samples of the exact input vector, but these multiple samples all have a difference next word. The network will not necessarily pick the best next word, but the most medianish word. When it comes to writing, there is a huge difference between the most median word, and the best word.
In Ovid’s Metamorphoses, there is the tragic story of Pyramus and Thisbe. It’s pretty well known as a love story. But it’s no Romeo and Juliet, and yet, they are the same story. One is the median next word, one the best next word. The challenge for true human-level generative AI is that there have not been enough geniuses alive yet. (If you take this argument to its logical conclusion, there is actually money to be made by selling training rights to decent literature.)