Next Windows or Office download = 4.79 TB
In similar development, as much as 98% AI generated code wastes my time because I need to delete or press escape to stop the auto completion because a) it is spewing code from official doc verbatim, b) risks injecting licensed code into the product and getting my employer on legal fire, c) it is wrong and exactly not what I want.
Considering that most of their software has been developed for decades and AI assistants have only started becoming useful in the last ~4 years it would be very surprising if 30% of their code is AI written. I doubt they even touched 30% of their code in the last 4 years. But what is perfectly plausible is that 30% of their code is written by code generators. Microsoft has a lot of interface code. All the windows DLLs that are just thin syscall interfaces, the COM and OLE interfaces in their office suite and everywhere else, whatever Office uses nowadays for interoperability to allow you to embed content of one product in another, whatever APIs their online products use, etc. In the leaked Windows XP source code it can be difficult to find the actual source code in between the boilerplate files containing repeated definitions, and in the decades since then the world has only leaned more into code generation.
Ouch, this makes the headline not sinply inaccurate. It makes the headline false.
Having generated code is nothing new. Multiple companies I worked for in the past had a lot of automatically generated code for multiple things (from automatically generating tests for API contracts to automatically generating glue code between different modules/services).
Callong that "AI" borders on the level of malicious ignorance.
To be fair, Hanlon's razor suggests it was most likely written by someone who _would_ conflate software with AI. Though knowing the reward structures in '''journalism''', it could still be deliberate.
"I’d say maybe 20%, 30% of the code that is inside of our repos today and some of our projects are probably all written by software".
> He knows exactly what he is saying
Very much so. He's plainly pulled those figures out of his arse, keeping it vague so he can't be blamed later for lying, for nothing more than "look how much we use AI" marketing purposes.
I still remember as a 25yr old programmer in England, attending a job in what was basically a porta-cabin in the countryside around Birmingham. Inside all the walls were covered in 1980s "Hang in there" motivational posters, only they were originals and it was unironic ( it was like a time capsule ). After shocking the boss by agreeing to a cup of water offered by the secretary which caused a massive problem as they had no glasses, the over weight boss sat me down and asked me:
"What percentage of windows is written in C++?" The position was for web development...
I guessed correctly ( it was about 40% ).
I then apparently failed the interview as he suddenly asked me if my girlfriend wanted to interview for the position ( there was no indication she was even in the field of CS ).
The moral of the story is apparently the code ratio of Microsoft products is a big deal to random middle managers in the black country. ( Thats what the region is called because the soot from coal mining would blacken the landscape - yes its what Mordor was based on by Tolkien).
2. What's the bug rate like from AI code? is MS investing the time to check this code, or assuming it is correct?
3. Who writes the test cases?
4. 30% in what terms? number of lines of code? number of commits? what code is in play here? bulk test code, which is huge and per-line very low value, or is this serious, difficult, complex code used in critical project functions?
5. How much design work is being performed by LLM? as opposed to churning out lines of code.
6. Do the developers actually trust this code and trust LLMs for this, or is it they have a KPI they have to meet and so everyone now has "30% LLM code", all of which is pretty much useless?
The spyware part is manually reviewed, the rest... not so much.
What matters is what "investors" and bean counters believe to be true. Not much relief in knowing that they are wrong, when you lose your job or your working conditions suddenly become much worse.
AI can replace devs or they can't. And it is sort of weird that people think about how many lines of code they output.
It is weird, because it has been a very long time that writing lines of code is far from being my main activity, even if it is the part of the job I like the most. I wonder where exactly people work where this is their main activity if they are not junior developers anymore.
Yeah, LLMs are sort of useful at spitting out code, if you know how to code. It can output code faster than you, and if you have good domain knowledge you can proofread it for the inevitable bullshit/bugs.
That's about it. Even to write decent prompts to guide it you need to understand well the domain to get it working.
Vibe coding is the new offshoring.
MS Teams is using 1200 MB of RAM when it starts running. Skype for Business used less than 180 MB, same functionality.
Outlook is completely unusable at 1366x768 resolution (which only shows up when you use your laptop while traveling!)
I predict that soon you’ll see some large corporations publicly switching to other solutions; the path MS is on which neglects these kinds of problems, will be a natural progression.
But instead I'll just say I'm curious about the next Windows edition. Traditionally it would be time for it to be the "not as terrible" one, which many currently on W11 would probably welcome with very, very open arms. Would that still hold?
Would there even be a Windows 11? (or, by MS naming schemes, Windows Series W or something)
I hate Google just as much, but at least their web office suite just works on the rare occasion I need a collaborative document.
How much boilerplate code exists that gets generated from a template, probably most of it.
fn actualLogic(param, param, param) ...
switch (some Condition) { case a: actualLogic(Foo, bar) ; case b: actualLogic(bar, Foo) ; ... }
It might be expressed with if statements, pattern matching, function calls etc., but most of the code is boilerplate.
From my experience in a C# codebase most of the code which completes correctly is this boilerplate.
O3/Sonnet/Gemini are able to write actual logic in chat/agent mode sometimes but then the problem is that they rewrite too much, which would count as AI generated code.
These two factors probably play a huge role in the 30% if it's counted in any accurate way.