Rust coreutils is a single binary of 9.1 MiB, similar to busybox
It replaces a bunch independent 50-100 KiB utils.

the C utils used C library, ldd and vdso

Rust coreutils is linked in addition to libinog5, libgcc_s, libselinux and libpcre2 to add 2 MiB of possibilities.

Why would you need AWS for exchanging just small text ? It is not true for emails. It is not true for messages either. Just use any federated text services. All of them have a very good support for encryption client to client.

Unreal Engine V works also on MacOS and Linux. Both are comparable to PS5 market size, and cheap to do.

As Rust is a static compilation language, to avoid a 1GiB Rust-coreutils, you got a single 9MiB binary doing all the command, quite similar to busybox but:
- it is 10 times larger than busybox (900 KiB)
- original C-coreutils are independents few KiB binary without interaction between them.

Thus:
- Rust coreutils has a much larger attack surface due to interaction between formerly separated binary.
- Rust binary reading/validation against rootkit is harder in a single big file than in many tiny independent ones. Supply chain attacks that barely exist at all in C-coreutils compilation (changing the C-compiler), will be harder to discover.

Sorry, my mistake ! Yes, it is counted !
The OS market share of Debian or Arch are stable between "Linux Only" and "Combined", thus Steam OS Holo is still there, but it is just not listed.

It should appear as the first Linux, at 0.8%

Why, It is not listed ? No idea :-)

Steam Decks are NOT counted in the 2.89% ! Proof is in the full statistics on Steam website..
Steam OS Holo is 28% of statistic in "Linux Only" OS page, but this OS is not listed at all in the "Combined" page with the 2.89% market share.

Thus (if all Holo installs are all on Steam Deck)
- Steam Deck is roughly 1% of steam users
- With Steam Deck, the part of Linux is around 3.7% (0.28*2.89 + 2.89)
- 3.7% is the double of the 1.88% of MacOS.
- Windows 11+10+7 is 95% (remaining 5% of Steam is still not negligible, especially with less competition)
- Windows 10 is still 35% (impact of the death of 10 on Steam ?)

Disclaimer: I am still an old Firefox user, liking it a lot.

  Looking web browser history, if I am not wrong, a reference implementation of w3c already existed for 15 years.
https://www.w3.org/Amaya/
But its development stopped in 2012.

Reference Implementation path for Firefox looks very much like a dead end to me.

Why would somebody clicks on a g-doc list if he or she dislikes the leak of personal information ?

Curious about the list, I loosed. I clicked too fast without reading the link :-)

CS self-learning of the 80's: copying printed mini-games was the norm on 8-bits computer. It works for learning a programming language and the basics of algorithmic theory and data structures (array, list, tree, hashtable).
(Mini-game: 100 to 1000 lines of code; fits in few pages)

The main point of using paper in 2025 is to slow the process to give them time to think and making clear that writing the code is the exercise.
Useable easily with full beginners. Crafting nice and small mini-games is more tricky for advanced topics. Obviously, do not put the PDF with the code on the lecture web page.

Yes, I know, smarphone camera + LLM can be used. May be it is limited to lab sessions with a teacher, or must be not graded.
I am also using the last language version with hope to detect copied/generated "old style" code easily (eg. in C23 constexpr instead of #define, "void f(int n, a[static n+1])", wchar_t, local_storage, L"text", u8"text", ...)

Learning work is not work. It is learning. But it is still work, and thus most human will try to cut it.
Not doing homework just stacks learning debt and forbid to learn any more elaborated knowledge.

At AI time, my current strategies to teach CS:
1- My lectures start from the current state of the art and go down the basics. The lecture time is thus spend at my final desired level of knowledge. It is then clear and can be freely used at the exams. Difficult to go alone the reverse way from the basics to my goals as much more large knowledge spectrum must be acquired.

2- All my practicals use extensively the last language version (eg. C-23 arrays parameters, nullptr, threads, generic, bool, local storage, C++-23 ranges, threads, option, etc. ): never Python, never POSIX Threads.
2'- Alternative: low profile language: Algol68 in GCC-15 seems promising, but now I am using MODULA-2 for few small practicals

3- Humans learn by copying, thus producing any handwritten documents, even of some text coming from AI, wikipedia, or any cheating system, are fine enough. But, grade it "Fails" every time a student has copied (thus learned) a single obvious false statement (not done 5 second check in wikipedia) .

4- Ask questions with an easy answer in wikipedia but with low content overall on the web:
eg. history of computer science, about soviet union contribution: Andrei Erchov do not play a significant role in Algol68 development, nor soviets invents PASCAL, nor the Setun 1 was a mobile computer (AI improves over years, but still...)

5- Do computer lab exam on anonymous account without internet for any basic knowledge (perfectly working up to master level, with a nice IT Team able to set up local software and documentation)

6- Long term team projects are difficult for AI (1 month full time, 20 000K), or impossible (1 month full time, 2 000, highly technical lines)

7- Combining by pair false/true basic questions in multiple choices quizzes: 25% rate of random success.

DRM law protection justification is written as protecting the works of authors. No author, no DMCA. Thus, if AI generated art content (Book, Music, Movies, ...) becomes ubiquitous (for books, it seems to be already the case. Spotify shows it may also be the case for music), it becomes lawful to break DRM and distribute both tools to break DRM and the uncopyrightable content.

Thus, does it mean that any software if either open source or public domain ?

No copyright claim can be issue for any generated code. This seems to implies one of two alternatives:
- either the software is open source, then the uncopyrightable part of the code can be shown, is small, and then the remaining software is copyrightable
- As most of all small binary parts of a software may be demonstrated as AI generated, any copyright claim is quite difficult to defend for the whole software.

Assuming a constant ratio of Linux between the two last surveys, and assuming "Simplified Chinese" means "China"

In maxima:
solve([x*0.5006 + y*(1-0.5006) = 1.45, x*0.2918+(1-0.2918)*y=2.06], [x, y]);

x = - 937 / 104400 == 0,008975095785
y = 304063/104400 == 2,912480843

Thus the Linux steam market share without SteamDeck is roughly 3% outside China, and 0% in China (China is 50% of the current market)

When looking overall (Win, MacOS, Linux), SteamOS is not shown in the system statistics.
The best Linux is Arch at 0.14%

SteamOS appears in the "Linux Only" statistics (34.67%)
where Arch is 9.70%. (34.76*0.14)/9.7== 0.50039% of SteamOS overall

0.5% of SteamOS: if 132 millions active players, then, 6 millions SteamDeck. 1/25 of the number of Nintendo Switchs. A quite good result :-)

Fortran is never slow. Dot.

It is very difficult to generate slow code ! Eg. any Fortran compiler will always efficiently use the memory (no pointer => no aliasing => no risk of overlapping => better optimizations) and the code will always use vectorization (Go, C++ valarray, functional syntax in Rust, python-like comprehension, etc., I am writing about you).
The performance does not depend of the syntax constructions, types, or compiler arguments, as much as in other languages.

Fortran was going down up to the last 5 years. Yes, string management, the I/O syntax and the type syntax are "slightly" outdated :-) . But the rest (especially arrays) are quite convenient or reasonable. My opinion is that the recent major usage of Python in science has also probably refresh some memory about what is the performance achievable in Fortran compare to Python, with similar lines of code.

Moreover, with just Fortran code, it is possible to code at the scale of millions of distributed cores, to achieve the best performance. Modern Fortran includes, in the language, OpenMP (efficient multithreading, including tasks queues for the threads, and accelerators constructs for the GPU) and PGAS (Coarray, distribued arrays, using MPI calls).