I think the relevant analogy here would be to run a local model. There are several tools to easily run local models for a local API. I run a 70b finetune with some tool use locally on our farm, and it is accessible to all users as a local openAI alternative. For most applications it is adequate and data stays on the campus area network.
As an ancient imbedded system programmer, I feel your frustration… but I think that it’s misguided. LLMs are not “computers”. They are a statistics driven tool for navigating human written (and graphical) culture.
It just so happens to be that a lot of useful stuff is in that box, and LLMs are handy at bringing it out in context. Getting them to “think” is tricky, and it’s best to remember that what you are really doing is trying to get them to talk as if they were thinking.
It sure as heck isn’t programming lol.
Also, it’s useful to keep in mind that “hallucinations “ are not malfunctions. If you were to change parameters to eliminate hallucinations, you would lose the majority of the unusual usefulness of the tool, its ability to synthesise and recombine ideas in statistically plausible (but otherwise random) ways. It’s almost like imagination. People imagine goofy shit all the time too.
At any rate, using agentic scripting you can get it to follow a kind of plan, and it can get pretty close to an actual “train of thought”facsimile for some kinds of tasks.
There are some really solid use cases, actually, but I’d say mostly they aren’t the ones trying to get LLMs to replace higher level tasks. They are actually really good at doing rote menial things. The best LLMs apps are going to be the boring ones.
A more accurate analogy would be, are you capable of finding and correcting errors in the model at the neural level if necessary? Do you have an accurate mental picture of how it performs its tasks, in a way that allows you to predictably control its output, if not actually modify it? If not, you're mostly smashing very expensive matchbox cars together, rather than doing anything resembling programming.