*The Rustification of Bun#
Rewriting 500,000 lines of Zig into another language would be a gargantuan undertaking if done by hand. “A rewrite in another language would take a small team of engineers a full year. It would mean freezing bugfixes, security fixes or feature development for that time,” Sumner wrote.
Instead, Sumner went with Claude. He spun up about 50 dynamic Claude Code workflows, reaching a peak of about 1,300 lines of code per minute and generating over a million lines of Rust code. The job took 11 days and cost about $165,000 at API pricing. Claude Fable did most of the heavy lifting.
How much is one million lines of code?
If a senior engineer works in a well-defined, well-architected greenfield project, and does extensive tests, 10,000 lines of code written in one year are very productive. The amount of new code per person-year will be much less in larger projects - a good indicator is the line count and the estimated value of the Linux kernel (40 million lines in 2025, redeveloping it in 2011 would have already cost 2200 million Euros, which would pay e.g. 36000 person-years for mid-level German software developers, that would mean 1111 lines per person-year{actually about half of that because of the interim growth of line count in the meantime }- and most software projects can only dream about so much actual value.)
So, you can say 1,000,000 lines of code generated in 11 days are probably 100 person-years of technical debt.
One caveat: It is relatively simple to auto-translate code in restricted languages like Java, Python, or Rust to unrestricted languages like C or C++. In that case, enforced invariants (like “no use of freed memory” or “sharing data across threads xor mutation”) become implicit, but verbatim translation is often possible.
The other direction will require a re-design…
you can’t compare the tech debt because it’s essentially a line by line translation from an existing codebase, not 1M brand new lines of code
That’s…irrelevant.
Really. The specific details of the task doesn’t change that this was automated. Until it’s reviewed, it’s basically a pile of Schrödinger’s bugs and security holes. That’s the technical debt.
It really isn’t. A line by line translation doesn’t introduce new logic and doesn’t change the system architecture. Saying it carries the same tech debt as brand new code is just not true.
Assuming the llm did it line by line without changing anything during an arbitrary hallucination.
and that’s a different problem entirely
not sure if I understand correctly, but this sounds off
More like 500 lines per person-year.
You can do the math yourself:
how many lines of code does the linux kernel have?
what is the estimated value of the kernel, which is the hypothetical cost to re-create it?
what do you think the seniority of a linux kernel developer is?
what annual salary do they make in the country which uses Linux most - the US?
now divide kernel value by annual salary to get man-years, and lines of code by man-years to get average lines of code per year
Result: Typing in code that somehow compiles was never the bottleneck