No, he doesn't say that. He said that "maybe 20%, 30% of the code that is inside of our repos today and some of our projects are probably all written by software". A guy like Nadella doesn't accidentally say software instead of AI. He knows exactly what he is saying.
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.
dang 58 days ago [-]
Ok, we've replaced AI with software in the title above. Thanks!
surgical_fire 59 days ago [-]
> No, he doesn't say that. He said that "maybe 20%, 30% of the code that is inside of our repos today and some of our projects are probably all written by software".
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.
cholantesh 58 days ago [-]
>Ouch, this makes the headline not sinply inaccurate. It makes the headline false.
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.
abanana 59 days ago [-]
And not only that, but how much vagueness and weasel-wording can he pack into one sentence?
"I’d saymaybe 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.
dlachausse 58 days ago [-]
I would say that most mainstream software projects have a significant amount of code that has been generated by software. IDE new project wizards, Rails, Django, NextJS, create-react-app, etc. are all software that generates code.
franktankbank 59 days ago [-]
Ya, I'ma go ahead and flag this. What trash journalism.
duttish 59 days ago [-]
I'd guess they also have various tools for large scale fixes/refactoring, and the output from those could count as well.
ionwake 59 days ago [-]
Sorry if this is unrelated.
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).
mosdl 59 days ago [-]
Sounds like a great sitcom idea
casenmgreen 59 days ago [-]
1. Has the total amount of code risen, such that the amount of human code is unchanged?
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?
Mountain_Skies 59 days ago [-]
7. If the code causes damage, who will be held responsible?
InsideOutSanta 59 days ago [-]
8. How much of that "software-written" code is stuff like Intellisense completing method names?
nottorp 59 days ago [-]
> is MS investing the time to check this code, or assuming it is correct?
The spyware part is manually reviewed, the rest... not so much.
johngossman 59 days ago [-]
These are the questions I came here to ask.
wolvesechoes 59 days ago [-]
People are going to ask if this is true etc. Not understanding that is doesn't matter if it is true, or if it is rather 20%, 2% or 40%, if it is new code or else.
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.
surgical_fire 59 days ago [-]
This is sort of the correct answer. But it also doesn't matter.
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.
fvdessen 59 days ago [-]
I recently rewrote a 40K lines application to 2K lines. Mostly by using newer tech and better design patterns. (Java to TS, Terraform to CDK, lambda soup to single app, etc.). It was a POC at work to show how things could be different. You'd think this would have made some devs rethink their approach, but no they're using AI to write even more code faster in the old approach.
arnaudsm 59 days ago [-]
I've witnessed projects beind rewritten using Cursor, with 10x more code & dozens of bugs no one noticed.
Vibe coding is the new offshoring.
ken47 59 days ago [-]
Written by software != written by AI. But CNBC has to pump.
Tade0 59 days ago [-]
That's still a huge indictment towards the languages/libraries they're using.
Just imagine how much better would it look if it said "30% of our code is removed by AI".
ankurdhama 59 days ago [-]
Hey Satya, you should ask Windows 11 taskbar team to use "AI" to write code so that the taskbar is resizable and can be moved. Let's see how good the AI is.
grishka 59 days ago [-]
But you don't need AI to undo a bunch of commits
shrubble 59 days ago [-]
I’m a daily user in a corporate environment of Outlook, MS Teams, Sharepoint etc.
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.
krige 59 days ago [-]
You know, there's a really obvious, snide comment to that.
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)
miyuru 59 days ago [-]
next would be "Windows AI"
pjmlp 59 days ago [-]
CoPilot Windows OS, running on CoPilot+ PC, obviously.
colonial 58 days ago [-]
If this is true (it isn't) then no wonder the web version of PowerPoint was so shit when I was forced to use it earlier today. Constant lag, random buffering popups, janky element snapping that can't be turned off...
I hate Google just as much, but at least their web office suite just works on the rare occasion I need a collaborative document.
xracy 57 days ago [-]
As originally written this headline is sensational. As now written, I would be surprised if this wasn't a bad estimate. Either way it's a meaningless number guessed off the cuff.
How much boilerplate code exists that gets generated from a template, probably most of it.
barotalomey 59 days ago [-]
RIP Microsoft
nipperkinfeet 58 days ago [-]
It's not surprising that the quality of their software has declined since Windows 10. How many months did it take to design a simple battery level icon? Many recent KB updates were complete failures.
snickell 59 days ago [-]
I use AI heavily in my own programming, so I’m not against, but I suspect this “as much as” is mostly copilot doing “tab completion” style autocompletions, not AI writing and modifying functions on its own.
croes 59 days ago [-]
Given all the bugs with each update doesn’t seem to improve the quality.
GianFabien 59 days ago [-]
MS AI is trained on existing MS code = bug multiplier.
godelski 59 days ago [-]
I think 60% of my code is written by vim. I make pretty heavy use of competitions, especially contextual compilations. Anything longer than 3 characters is only going to be typed once
amai 58 days ago [-]
Only 30%? Given Microsofts product quality I would have guessed a much higher percentage of vibe coding going on at Microsoft.
WhyIsItAlwaysHN 59 days ago [-]
Very often the code looks like:
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.
j4coh 59 days ago [-]
“As much as” are weasel words to help them sell their AI coding products to others.
apercu 59 days ago [-]
Well that explains a lot ;)
sys_64738 58 days ago [-]
Maybe the majority of that is for code comments and formatting.
faragon 59 days ago [-]
Translated: same engineers are more productive by using IA.
lm28469 59 days ago [-]
You can produce lot of code with zero or negative real productivity
faragon 56 days ago [-]
Yes. And that can happen with and without IA.
jgaa 59 days ago [-]
That explains a lot.
taubek 59 days ago [-]
30% of new code, or of all existing code?
mkl 59 days ago [-]
The way he phrased it, it's existing code, though he doesn't say it's by AI. From the article: “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,”
anovikov 59 days ago [-]
The Clippy strikes back
GianFabien 59 days ago [-]
Clippy v1 writes Clippy v2 writes v3 .....
Next Windows or Office download = 4.79 TB
pjmlp 59 days ago [-]
Nope, you will be using an OEM version of Windows 365 Link to connect to the Windows CoPilot 365 instance on Azure.
enigma101 59 days ago [-]
Misleading clickbait
n_ary 59 days ago [-]
Hmm... But SN himself is not writing code or involved in day to day development. He can claim anything and his incentives are 100% aligned to claiming AI bs.
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.
nuc1e0n 59 days ago [-]
If true, the methodologies and languages they use must be quite inefficient if they can be automated by machines like that.
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.
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.