The Future Ahead
In this book, we started with how a neural network could digest text. As we have seen, neural networks do not do this natively but require the text to be processed. Simple neural networks can be used for some basic tasks such as classification, but human language carries an enormous amount of complex information.
In Chapters 2 and 3, we saw how we need sophisticated models in order to use semantic and syntactic information. The emergence of transformers and LLMs has made it possible to have models capable of reasoning and storing enormous amounts of factual knowledge. These multipurpose knowledge and skills have enabled LLMs to solve tasks for which they have not been trained (coding, solving math problems, and so on). Nevertheless, LLMs have problems such as a lack of specialized domain knowledge, continual learning, being able to use tools, and so on. Thus, from Chapter 4 onward, we described systems that extend the capabilities of LLMs and which are designed...
