One problem(?) for AI assisted programming is that people who are professional programmers today have a certain personality type in which we like programming because we enjoy making deterministic rules to describe fuzzy real world conditions. With AI, that goes out the window and you have fuzzy input (the prompt) and fuzzy output (the behavior). One way to approach it is to just treat the AI as a blackbox, and quarantine it. This is what all systems have done basically up until now, but I think it’s going to breakdown, and software development itself will become a lot fuzzier. Instead of source code → machine code → execution, we’ll have prompts → IR → execution. Maybe the IR will be inspectable by humans or maybe not, but either way, most of the IR will just be passed through because it’s not a good use of time to try to inspect it except for the most critical sequences. Maybe a big portion of the prompts will end up being test conditions. Anyhow, none of this sounds appealing to me as the sort of person who has encoded algorithms off of Wikipedia for fun. It seems more like babysitting an unknowable process. Oh well, only 25 years until I’m retirement age.
It feels analogous to what happened when cameras were invented: the ability to render visual depictions of reality became largely irrelevant but was replaced in importance by:
the ability to discern what was worth photographing and
the ability to imagine variants of reality and render those instead.
Another analogy is the invention of recorded music (initially via the gramophone): the ability to faithfully render music composed by somebody else initially dwarfed in importance by the ability to compose original music and later supplanted by the ability to modify and mix existing captured/synthesized music into new blends.
One problem(?) for AI assisted programming is that people who are professional programmers today have a certain personality type in which we like programming because we enjoy making deterministic rules to describe fuzzy real world conditions. With AI, that goes out the window and you have fuzzy input (the prompt) and fuzzy output (the behavior). One way to approach it is to just treat the AI as a blackbox, and quarantine it. This is what all systems have done basically up until now, but I think it’s going to breakdown, and software development itself will become a lot fuzzier. Instead of source code → machine code → execution, we’ll have prompts → IR → execution. Maybe the IR will be inspectable by humans or maybe not, but either way, most of the IR will just be passed through because it’s not a good use of time to try to inspect it except for the most critical sequences. Maybe a big portion of the prompts will end up being test conditions. Anyhow, none of this sounds appealing to me as the sort of person who has encoded algorithms off of Wikipedia for fun. It seems more like babysitting an unknowable process. Oh well, only 25 years until I’m retirement age.
It feels analogous to what happened when cameras were invented: the ability to render visual depictions of reality became largely irrelevant but was replaced in importance by:
Another analogy is the invention of recorded music (initially via the gramophone): the ability to faithfully render music composed by somebody else initially dwarfed in importance by the ability to compose original music and later supplanted by the ability to modify and mix existing captured/synthesized music into new blends.