Summary
In this chapter, we discussed the transition from transformers to LLMs. The transformer was an elegant evolution and synthesis of 20 years of research in NLP, combining the best of research up to that point. In itself, the transformer contained a whole series of elements that enabled its success and versatility. The beating heart of the model is self-attention, a key tool – but also the main limitation of the LLM. On the one hand, it allows for learning sophisticated representations of text that make LLMs capable of countless tasks; on the other hand, it has a huge computational cost (especially when scaling the model). LLMs are not only capable of solving tasks such as classification but also tasks that assume some reasoning, all simply by using text instructions. In addition, we have seen how to fit the transformer even with multimodal data.
So far, the model produces only text, although it can produce code as well. At this point, why not allow the model to be able...
