“Our strategy is to combine AI and Algorithms to rewrite Microsoft’s largest codebases,” he added. “Our North Star is ‘1 engineer, 1 month, 1 million lines of code.’”
That’s insane. Even a good engineer will frequently need years to fully understand one million lines of code - even if the code is organized very, very well.
To compare, one million lines of program code might have around 100000 to 200000 unique symbols whose meaning and complex connections an engineer working with it has to learn and memorize. That’s far more than the average vocabulary one will learn in five years when learning a foreign language to a high skill level. Doing it in a month would be like learning to read and write fine Japanese or Arab literature in a month when you have never spoken a word in that language before.
The Linux kernel has now passed 40 million lines of code, written over 30 years by tens of thousands of master programmers. And that’s kind of a historic achievement. What happens is that complexity increases sharply with each duplication of the amount of code.
Kinda still your point, but if you have one engineer producing 1M SLOC, how many do you have for code review?
I hate how everyone nowadays is acting like reviews are not important. Actual oversight over codebase is way less important than shipping random code. Which is insane.
That’s insane. Even a good engineer will frequently need years to fully understand one million lines of code - even if the code is organized very, very well.
To compare, one million lines of program code might have around 100000 to 200000 unique symbols whose meaning and complex connections an engineer working with it has to learn and memorize. That’s far more than the average vocabulary one will learn in five years when learning a foreign language to a high skill level. Doing it in a month would be like learning to read and write fine Japanese or Arab literature in a month when you have never spoken a word in that language before.
The Linux kernel has now passed 40 million lines of code, written over 30 years by tens of thousands of master programmers. And that’s kind of a historic achievement. What happens is that complexity increases sharply with each duplication of the amount of code.
Kinda still your point, but if you have one engineer producing 1M SLOC, how many do you have for code review?
I hate how everyone nowadays is acting like reviews are not important. Actual oversight over codebase is way less important than shipping random code. Which is insane.
GitHub Copilot performs PR reviews now doesn’t it
it is not much better than just running a linter over your code.
Not disagreeing with you there
Yeah, that’s exactly what we need.
LLM producing the code
LLM approving the change
LLM agent pushing to prod
Then wonder why the users are gone
It’s not great.