In today's note on AI, we'll be speaking about GenAI models and how they think, whether it's similar or different from how the human brain thinks. In order to make things simple, we'll be talking about large language models as core representations of today's GenAI models. Large language models are neural networks which are trained on a huge amount of text data, and these neural networks in the process of training are able to identify the patterns which exist in the words, all the connections between the words, paragraphs, and between topics and subtopics.
As a result, the core element of LLMs is the possibility and the ability to predict the next word or next token based on the contextual information or on the request which the user is sending to it. So prediction is really important. What about the human brain? From one point of view, you can also say that the human brain predicts in a similar way. We can think about the brain as a neural network which has some weights, which is of course a lot more complicated than the neural network of a large language model.
At the same time, it also predicts based on some inputs and based on the data it knows. The core distinction here lies in the type of the information which large language models and human brains are using in order to make these predictions. This information is called grounding. Large language models are grounded on text data, on the textual representation of all human knowledge and experience. We can think about it as a type of a wide knowledge or wide grounding. When we are talking about the human brain, it's another type of grounding.
Of course, it's also grounded in text or in words, because we are speaking with our teachers, with our parents, we're reading books and so on. At the same time, there are other levels or other layers of grounding. These are our emotions, the experience of this world, our sensory feelings and other types of things which make us human. We may think about it in the following way: large language models are trained on the representation or description of this world by human beings, and human beings are grounded in reality.
We are living in this world. We are seeing it and we're feeling it. So I would say this is the main difference: the grounding of large language models is wide, and it contains really a lot of information not available to just one single individual. At the same time the grounding for the human being, for the human brain, is deep. It contains different layers, not only grounded in words but also grounded in other elements which we associate with and which we experience by living inside this reality.
The conclusion in this note for today is the following one: maybe ideal prediction is a combination of wide and deep types of grounding. If we are able to combine the human being's deep grounding with the wide grounding of large language models, maybe this will be something amazing, and maybe in this way we'll be able to predict a lot better. This was a Christmas note on AI for today. To those who celebrate Christmas, I wish you merry Christmas, and I'll see you in the next episode. Bye.
