That’s kind of the killer feature of an LLM. You don’t even need to have your fingers on the right place on the keyboard and it will parse gibberish correctly as long as it’s shifted consistently.
Programs should precisely define what their inputs are and loudly reject all else.
Moreover, for this misfeature, you have to use a cloud API, where your syntax is analyzed by some massive cluster, using scads of processing and memory resources.
We could have a natural language command line for FFMpeg requiring at most megabytes (probably just kilobytes) that would work on an air-gapped machine.
In the early 70's, the SHRDLU project achieved amazing chat interaction with symbolic processing, on the hardware available then. It was a way more impressive hack than LLM. Not just because it required relatively low resources, but also because its author could actually explain its responses, and point to the responsible pieces of code behind them, which he designed.
…when interfacing with other programs. Humans aren't programs, which is a somewhat important distinction.
Secondly, people need precision and stability in program behavior also. People learn that when they give a certain command, they get a certain effect, and they don't want it to break.
Interfaces that allow people to express some wish in many ways (language, direct manipulation, ...) can still be rigidly defined: appear flexible to the casual user, but rigid to someone who cares to study the several thousand rules under the hood (or whatever).