One my friendlies on X, @_quanta_ recently said
”Not to be a pessimist, but the sad truth about software is that there's no more moats. If you create something that's valuable to people, someone can just ask an Agent to build a clone of it over night. I guess that's why I only work on FOSS stuff as my side projects.”
And I really had to think about that. Is it true? If so the software industry as we have known it is most surely dead.
But then I got to thinking about the conversation I recently saw between the Moonshot podcast crew and the gentlemen running a company called Blitzy which is making a lot of money translating millions of lines of old code from old accounting and legacy business software into modern, memory safe code.
The line I liked best is “Post-Blitzy the truth moves to the documentation because you can regenerate the code overnight anyway.”
In both cases everything seems to happen overnight.
Except, is that how it really works? Even with very high quality specs it can take days of debugging LLM sourced code even if the debugging is being done by the LLM.
And then there’s the case of those weird corners of latent space that the LLM just can’t quite reach. I still often find myself manually redirecting the Agent into a correct region of code by hand correcting a smidge of code or pointing it literally in the right direction.
And I see absolutely no reason to assume that will change any time soon. If so, how?
Yeah, yeah “they will just keep getting better” But does that really mean anything when the architecture and execution flow remain the same?
Look at what Bitcoin has taught us: there are some processes that are so expensive to hack that it is cheaper and more effective to just go along and mine coins.
Software that is being constantly improved and expanded by its wranglers will always have an advantage over code that is only now being generated. And in a world of very, very rapid development the racer that is already out in front will have an almost insurmountble lead over Johnny-come-lately copycats.