Summary
In this chapter, we have seen how the tools we looked at in previous chapters can be added to an LLM. We saw that an LLM is capable of planning and reasoning, but it produces weaker results when it comes to execution. An LLM is capable of generating text, but at the same time, the enormous amount of information learned allows it to develop skills beyond text generation. While it is a computational waste to ask an LLM to classify an image, an LLM can use a specialized model to solve the task. As we saw with HuggingGPT, a model can invoke other models to identify a pizza in an image. In that case, we saw an LLM invoke more than one model, collect their outputs, and conduct reasoning about the results (observe that the models agreed on the type of pizza in the image). The LLM can then conduct reasoning and choose which models need to run to complete the task, collect the outputs, and observe whether the task is completed.
This concept makes it possible to revolutionize various...
