Someday people are going to get tired of "programming in English" with prompts, getting inaccurate output, etc and someone is going to invent a higher level kind of CODE that allows the user to directly specify the actions the computer should take to solve the problem. Later someone will invent a kind of tooling that COMPILES these CODES into a runnable thing skipping the prompt part all together. It might be called something like Unified Prompt Language.
Programming is moving to programming by stated and understood intent, rather than syntax. Maybe contracts/legalese, but definitely not compilable code. Sure, some people will compile code, and more than some will be reading generated code, but that will be increasingly exceptional.
For some time now I've had some loose abstraction swimming around my head sort of between code and prompt, and maybe its the OUTPUT of the prompt. Im assuming going forward that developers will be no better at reading generated C code than I am at reading assembly today.
I'm imagining some "AI Native Intermediate Representation" where the prompt says "Give me a function that takes an array of strings and returns that array sorted", the real code that runs in the end is the actual C code, and the representation might be something low complexity but still human readable:
Not quite a prompt, certainly not runnable code but a higher level hybrid that is human readable AND can be compiled into actual C code. At some point someone is going to have to debug something and nobody will be able to read assembly/real code anymore.
It just seems to me in the whole "AI WRITES MY CODE" world that nobody is really thinking about debugging and maintenance. What do we just commit the prompt we built the program with to Git and call it a day? What about when we need to modify the system? Do we make a prompt to modify? Do we modify the original prompt?
AI so far is probabilistic, and that can certainly be dealt with via various workarounds, and theres currently no reason to think todays stuff would even generate the same code twice. I can't shake the idea theres a step missing in everyone saying "AI will write all our code now".
Its all a hack until a PHD writes a paper, then it becomes a technique.
Edit: Funny story - this morning Im up early to write a program of low complexity but still somewhat rigorous. Let me see if I can think up a prompt that I might use:
I need a program invoked with a main. We'll use this as a scaffolding to extend as the requirements materialize and constraints emerge. It will generate valid trading days for ES index futures where contract months are HMUZ. For each contract it will start with the last day of trading as defined by the contract specification located at (www.cmegroup.com/contracts/ES) and build and build an array consisting of 4 calendar months of those dates starting back from the last day of the contract. This collection will be sorted in ascending order by date where by the 0 element in the array is the earliest trading date.
^^ So this here is just my intermediate step. Im not even sure this is how I want to go forward with it but this approach is just the first way I can conceive of to get the job started. From here I'll read the output and figure how to add in the useful things I need to make it actually perform the intended function beyond what this prompt tells it to do. I cant even imagine how I might specify the final product I need in English. Id be writing a giant prompt for everything I need built and code would be so much faster to write and think in. Describing my stated intent in all this in English sounds like ... murder. And we all know that prompt I just made up sitting here is both a decent start at a prompt AND wholly inadequate and will produce absolutely useless garbage.
