besides this comment of yours is exactly what was expected as a first response, and is very boring.
The good ones will produce more reliable code than current LLMs. But on average, I don't think so.
Maybe it has to do with strings not really having been a first class citizen in C? Processing strings in C is a (dangerous) pain even in the year 2024.
I often find myself just writing python scripts calling subprocesses (if needed, so much is already included), ideally with a --json option so I get proper machine-readable output. This has shown to be far more stable and reliable than bash.
The author is complaining about inconvenience. If I'm in the shell, and I want to see how much disk space I have, asking a LLM for a command then running the command then asking the LLM to parse it is still inconvenient, compared to just typing "df -h" and getting a reasonable output.
The fact that Claude or whoever can parse that output seems irrelevant to the article as I understand it. "AI" is not a magic silver bullet.
All I’m saying is long prompts of shell progs that you need to use once or twice are easier written in the newskool way, not that proper output of common programs is bad idea.
But I really can’t imagine anyone or anything convincing thousands of devs to conform certain format. Really a wet dream today unless started so from day one. Only big players can do that.
Perhaps one may learn to remember <df> and <ls -a>, but will definitely not care to produce by heart oneliner bash abominations sprinkled with sygils and pipes and three-char cmd names.
This whole article very true, but somewhat irrelevant to 2024. Claude/Mistral/CHatGPT been parsing outputs perfectly, and there are plenty of experiments into getting the LLM to assist code writing and parsing results directly from the cmdline. And you know what, this also works in 100% console mode, you don't need GUI to talk to the new GUI called LLM.
the smaller models get, the sooner they get embedded in shells. then you'll get the real Ghost in the Shell, yeah...