The original LinkedIn post is pretty wild. I wonder if he did a fat line of coke before writing that, or if it actually were any concrete plans that have been worked out.
With AI, I can produce 1 million lines of code in one day.
Artificial benchmarks produce artificial results.
later
| ... our ...
I think Galen's goal to rewrite every bit of C/C++ was his goal, not Microsoft's.
That thing we don't have yet?
AI is summarizing the lessons, writing the students' papers, then grading those papers for the teachers.
Time to shut down the entire education system.
Like a compiler, we could optimize this by doing no work at all! Once you realize no one will look at any of the results, final or intermediate, there's nothing left to do.
How is it wild? On social media I kept seeing things like people falsely expecting the end goal would require manually reading through a million lines of code. It seemed more like people making up reasons to be mad or trying to dunk on the author.
If not, here, I can show you my opinion too!: No, what you said is completely false.
Which is absolutely batshit. There's no way that can be reviewed properly, even if it's putting all of the review work on all of the other teams.
They do still review code, but the first wave of layoffs in 2022 mainly hit principal engineers and above because some bean counters said "oh, these are the engineers that are costing us the most per head", so it's kind of the inmates running the asylum now.
And I'll say that their biggest sin was always that their code from the late 90s on was about 20% too clever for their own good. Kind of goes to that classic quip about how how it takes twice your brain power to debug code as it takes to write it, so if you were already maxing out just writing it, then you're not smart enough to debug it. That's half of why features seemed to get a 1.0 release, then get replaced with something rather than iteratively improved (the other half being FAANG style internal incentive structures).
Were all seeing the effects of them clearing house of their weaponized autism that was barely keeping the wheels on the wagon. They do review, but they don't have the ability to do it properly at scale anymore. Which makes rewriting everything even more batshit.
As a third-party developer in the late 2000s I remember my boss giving me a CDROM binder (binders?) of every single OS release that Microsoft had ever put out. I assume he’d been given it my his developer-relations rep at Microsoft. My team and I used it to ensure our code worked on every MSDOS/Win* platform we cared to target.
I expect that, internally, the Windows team have crazy amounts of resources to implement the most comprehensive regression testing suite ever created. To that extent, at least, you’d be able to tell if the Rust version did what the old code did even if you didn’t read the code itself.
That hasn't been nearly the same goal for decades now.
For instance, Crysis literally won't run on win10 or later anymore.
On top of that, security bugs aren't the kind of thing you can automate away during a rewrite that no one has the bandwidth to actually review.
That’s not to trivialize what a compiler does, but it’s effectively going from a complex form to its building blocks while maintaining semantics.
Changing high level languages introduces fundamentally different semantics. Both can decompose to the same general building blocks, but you can’t necessarily compose them the same way.
At the simplest example, a compiler backend (the part you’re describing) can’t reason about data access rules. That is the domain of the language’s compiler frontend and a fundamental difference between C++ and Rust that can’t just be directly derived.
This is "lets put our postgres database on blockchain because I think blockchain is cool" level of crap you see in peak bubble.
AI code generation is not deterministic and has no guarantee of behavior, thus requires review unless incorrect code is acceptable.
You don't have to use AI code generation to be what is generating the code or you could require some kind of proof of equivalence to verify the code that was generated.