> Jarred Sumner, co-founder of Bun and Member of Technical Staff at Anthropic, used Claude Code to migrate Bun from Zig to Rust. A million lines of code were produced in less than two weeks, with 100% of Bun's existing test suite passing in CI before merge. Nineteen regressions surfaced after merge and have all been fixed.
I couldn't come up with a clearer example of how this kind of stuff fails to apply to my work. If I shipped nineteen regressions in 2 weeks, my company would be desperately apologizing to our customers, not writing a blog post about what a great job I did.
The regressions got fixed before a release. Do you count your internal integration branch as "shipped"?
0.02 defects per KLOC is a low density, even considering that it was "just" a port to a new language and before giving credit for the existing bugs it identified and fixed. That defect density will probably only generalize to other projects that have similarly extensive regression test suites, but it's a proof of what LLMs can do.
The source article says that "Large code migrations are a particularly effective use case for these advanced models", and the specific advantages suggest to me that the results won't generalize at all. I can't think of anything other than a large code migration that would allow for thousands of independent workstreams against a sufficiently detailed spec to stop agents from going completely off track.
They didn't ship 19 regressions over 2 weeks. "Merged" doesn't mean "shipped". Their customers have probably seen none of those because they've all been fixed.
Such a terribly bad faith take. I feel upset reading this more than I should have.
On what basis do you say their customers probably didn't see them? I'm not intimately familiar with Bun, but I work with people who try to do software development in this way, and they routinely ship regressions to customers. They believe that this is because we're not AI-enabled enough, and if only we had the right loops they'd stop shipping bugs; I think that it's not about the loops and the bugs reflect their lack of understanding about what they're trying to do.
Regressions is a meaningless metric to me. Every app has bugs, and most are either rarely encountered or aren't breaking the app. All I care about is interruption of service.
Migrating a million lines of code is how many human-engineer-years worth of work? Some quick fermi estimates place it at between 15-100 years depending on how much automated tooling is used (without AI).
It really blows my mind how some AI naysayers like you fail the absolute basic common sense test of comparing the comparable. If you, without AI, merged nineteen regressions in 2 weeks, how much work has actually been done? How many features did you rewrite? How many refactors did you make and how difficult were they?
It's precisely because I'm not an AI naysayer that I reject the premise of the comparison. My job is not to produce as much code as I possibly can; my job is to produce features that provide value to customers in a high quality and regression-free manner. To accomplish this I use safer AI-enabled development strategies, which produce fewer lines of code and involve more manual review but also do not result in 1 regression per calendar day.
Again, you are comparing things which are not comparable.
"Working on a feature" and "Working on a language-to-language port of 1M LOC" are NOT the same thing.
You're also still bringing up "regression per day", while ignoring the fact that without AI, this would have taken easily ~4000 days. Which is 1 regression per 0.003 calendar day. Not the same.
And you still aren't addressing sibling comments reminding you that these aren't shipped migrations. I'll second the comment asking you whether your code is always perfect first-shot, because if we're going to count those bugs, we might be well above your 1/day number.
> So, the whole brouhaha was so that Anthropic can write this blog post?
Remember mythos ? I was half-expecting them to scare us with a new danger: How AI can covert code into other languages, and confuse the hell out of all of us.
I couldn't come up with a clearer example of how this kind of stuff fails to apply to my work. If I shipped nineteen regressions in 2 weeks, my company would be desperately apologizing to our customers, not writing a blog post about what a great job I did.
0.02 defects per KLOC is a low density, even considering that it was "just" a port to a new language and before giving credit for the existing bugs it identified and fixed. That defect density will probably only generalize to other projects that have similarly extensive regression test suites, but it's a proof of what LLMs can do.
Such a terribly bad faith take. I feel upset reading this more than I should have.
It really blows my mind how some AI naysayers like you fail the absolute basic common sense test of comparing the comparable. If you, without AI, merged nineteen regressions in 2 weeks, how much work has actually been done? How many features did you rewrite? How many refactors did you make and how difficult were they?
Insane...
"Working on a feature" and "Working on a language-to-language port of 1M LOC" are NOT the same thing.
You're also still bringing up "regression per day", while ignoring the fact that without AI, this would have taken easily ~4000 days. Which is 1 regression per 0.003 calendar day. Not the same.
And you still aren't addressing sibling comments reminding you that these aren't shipped migrations. I'll second the comment asking you whether your code is always perfect first-shot, because if we're going to count those bugs, we might be well above your 1/day number.
Obviously "regressions" is not a great measure (how serious were they?) but 19 seems pretty decent to me.
Remember mythos ? I was half-expecting them to scare us with a new danger: How AI can covert code into other languages, and confuse the hell out of all of us.
And it is good for mental health - it allows us to express the distaste for Anthropic in a healhy way.