What is striking to me now that we are three years into the LLM revolution is ... there is no foundational software that has resulted. I keep asking where the good software is now that the miracle of LLM programming has entered the firmament.

Companies were always huge piles of proprietary custom code. Slowly what would emerge from those would occasionally be foundational software like ERPs, better databases, better message queues.

What I suspect will happen now is that everyone will simply construct their own distinct stacks of code, or just pile onto the existing proprietary ones inside companies. The "export" of code for general use will functionally disappear.

It's all throwaway, to be wastefully recomputed/rerendered by armies of H100s and only apathetically checked for utility.

I could be proven wrong, I'll know if Linux starts to miraculously close functionality and usability gaps with Windows/OSX, which would be the most visible example of LLM acceleration of development, but it so far has not.