Learning LangChain

Chapter 11. Building with LLMs

One of the biggest open questions in the world of LLMs today is how to best put them in the hands of end users. In some ways, LLMs are actually a more intuitive interface for computing than what came before them. They are much more forgiving of typos, slips of the tongue, and the general imprecision of humans, when compared to traditional computer applications. On the other hand, the very ability to handle inputs that are “slightly off” comes with a tendency to sometimes produce results that are also “slightly off”—which is also very much unlike any previous computing tendencies.

In fact, computers were designed to reliably repeat the same set of instructions with the same results every time. Over the past few decades, that principle of reliability has permeated the design of human-computer interfaces (variously called HCI, UX, and UI) to the extent that a lot of the usual constructs end up being subpar for use in applications that rely heavily on LLMs.

Let’s take an example: Figma is a software application used by designers to create faithful renderings of designs for websites, mobile applications, book or magazine covers—the list goes on. As is the case with pretty much all productivity software (software for the creation of some kind of long-form content), its interface is a combination of the following:

• A palette of tools and prebuilt primitives (fundamental building blocks), in this case lines, shapes, selection and paint tools, and many more ...