The oldest and most Lindy software tools and interfaces are all text-based — the command line, the word processor, and email. Essentially unchanged, they have survived iterations of graphical interfaces from Windows 95 to macOS, from proto networks like ARPANET to the Internet. Part of it is simplicity — text is the simplest way to model the world digitally.
As tools that primarily operate on text, will Large Language Models revive text-based interfaces? Will they merge seamlessly with the Lindy text interfaces we’ve used for decades or eventually replace them?
You can get pretty far with the Unix philosophy of operating AI. At this junction, it’s becoming more apparent that we will integrate text-based generative AI before we do multi-modal. We’re still in the stage of dealing with pure text-based pipelines, but there will be more use cases for structured text-generation pipelines soon. We used text in Unix input and output because we didn’t have the data interchange formats we do now (or the networks that forced the development of universally parsable designs).
If LLMs have a significant impact (and I believe they will), they will come for these interfaces first. Already, there’s GitHub Copilot for writing code (in the IDE, and soon to be in the terminal), autocomplete in word processors (and exploratory functionality in richer text editors like Notion), and autocomplete in emails (maybe one day without a human-in-the-loop).
Conversely, AI may start a new paradigm of generative interfaces — possibly bespoke GUIs adapted just in time to use cases and users. Work that would have taken many hours of developer time can now be done programmatically and generically with LLMs (possibly). Language is not an exhaustive interface. Images can be more information-dense than text (Screenshots as the Universal API?).