Now that we have a solid understanding of what an AI model is, we can discuss what large language models really are. They are huge neural networks which use layers of non-linear functions. You can think of them as a bunch of functions with feedback loops, that constitute an even larger function. In other words, a single “model” might have a bunch of “sub models” and those can have “sub-sub-models” and so on. Like any large system basically.
As these neural nets use combinations of non-linear functions (e.g: sigmoid, ReLu, etc), they can approximate any curve or surface - not just lines or polynomials. This is why they are called universal function approximators
The text-classifier we previously saw was:
A good way to think about an AI model is as a huge, flexible, curve in a very high dimensional space. Training is just curve-fitting: adjust the parameters to minimize the error between prediction and truth. Using the model later is just plugging x into the fitted function