I’ve also written a more detailed writeup here: https://imtomt.github.io/ymawky/
Today, I just think, "how long would LLMs have taken to write this?"
I mourn the death of a human artform.
Got an idea that you'd need assembly language for - now you can do it instead of..... never doing it because it would have been impossible for you in any practical way.
Look to the positive instead of lamenting something that never would have happened.
It's unbelievably exciting that you can now program a computer virtually without the limitation of your ability to hand code it.
The “you” doesn’t necessarily refer to you. Im addressing 90% of the developers out there. We love playing around technology… but I doubt we will be thinking the same once we become unemployable. But here we are, having fun with the tools of companies that want to finish us. How ironic
Nobody actually needs a web server built in assembly language, it serves no practical purpose. And I say that as someone who learned to program 6502 assembly language in 1983 and has sporadically used assembly of various architectures since.
The absurdity of building it would have been the curiosity draw pre-LLMs, but when it existing is just a series of prompts away it really loses all of its meaning.
But yeah... hooray for AI. Can't wait until we learn to harness it to supercharge the most important and valuable thing we do as a human society in modern times: stuff increasingly intrusive ads in front of everyone at all times.
The difference is that now it is worthless: there is no learning, no person caring about the result, nothing aspirational for the public to look towards... we used to enjoy those challenges, used to be proud of solving complex problems... now? Yeah, whatever, execute execute commit push, let another LLM "review" and call it a day.
If LLMs are good for doing things we aren't already doing, it indicates the overall addressable "value" that LLMs could provide for such things is actually quite low. If the task has necessary prerequisites that you don't currently possess, but you haven't spent the effort to jump that hurdle yet, then it's a good indication the value of completing that task is very low. Even if, maybe especially if, we're talking about personal projects where the value proposition is personal and not momentary, it indicates the person already feels in their bones that the return on doing this thing is not worth the effort.
I'm struggling with this with my leadership at work. We have developed a thing that is going to remove the need to hire temps [0] for data entry when we get clients who send us large amounts of their "data," aka "a thousand 30-slide PowerPoints each with one line graph of interest sitting in the corner of one slide somewhere." It is an ask that comes up a lot, it's always very expensive in both the time and money axes for the client for that task, but overall it's just a small part of the contact budget. I'm all for using what we've built to cut down the time cost for our clients, but my leadership thinking it's going to lead to massive cost savings for our clients seem to forget just how much time we spend in meetings and planning and documentation and testing and reevaluation and more meetings versus actually executing on things.
It's also bad business. To me, giving results faster should be a premium offering. We should be charging more, not less.
[0] We don't actually hire temps, we turn our junior data analyst into temps by burning them out with tons of unpaid overtime. They then leave and we have to back fill them at rather extreme cost of overhead for hiring compared to the direct contract overtime we didn't provide.
Only exciting if you already got a lot of programming under your belt, like Carmack, or a product guy.
Isn't that kind of view pathetic and sad, though? Why would anyone pick up and guitar or play a piano if they could just listen to the same song already made by someone else? I struggle to understand this view of people that pretend to not understand why being an expert of some skill is perceived as valuable by some people. This is also belies next problem with this line of thinking which is that it says "we don't need to learn X to do Y because we have AI" but misses the same AI could easily replace the need to have you think to do Y in the first place. I don't know.
But you're not doing it. The ai is doing it.
If the op can write a web server in assembly language then I'm pretty sure they could have done it in a higher-level language. But they did what they did for the journey and the learning along the way. Vibe coding it omits all that, and misses the point of the exercise.
>> without the limitation of your ability to hand code it.
yeah its nice though this in 100% of cases results of software of even lower quality we had before.
so hard to tell where is the win here. the fact that you can generate some code does not make it a win, just a curious fact.
If you are having an LLM generate the assembly language for you, that is not even remotely close to writing the assembly language yourself.
I don't find it exciting even in the slightest. I can think of nothing more boring and unsatisfying than having an LLM generate all of your code for you.
I mean, I understand why some think this could be exciting from a "I can get something done fast because the LLM generates it for me" standpoint -- because their excitement stems from something getting done at all instead of just sitting in the pool of ideas forever. However, you will never know the code generated by an LLM like you know the code you wrote yourself. Also you will never gain the same satisfaction of finishing a project where the code was written by an LLM that you gain from finishing a project where you wrote the code yourself.
If you are a person that doesn't care about coding or doesn't like to code at all, I could totally see why you'd find this exciting - to you it's all about avoiding work you don't care for or want to do yourself anyway. Also, a high percentage of people who do love coding have zero interest in writing assembly language, so if they were required to write some for a project, I could also see them being happy with having an LLM generate that part of the project for them.
However, I think for people who genuinely love to write code, the situation is the opposite of what you said -- it is far more sad than it is exciting. In fact, for many of them it has already reached the point of depressing for many reasons. I don't think it is primarily because the LLMs have gotten significantly better at generating code (which they have). I think some of the bigger reasons are that so many people who now pay people to produce code have:
1) got a very short-sighted and "rose-colored-glasses" view of what LLM-produced code will do for their company.
and
In the long run I think most (if not close to all) of those businesses are going to be sorry if they over-indulge in replacing human-produced code with LLM-produced code. I think the ones who lean too heavy on the LLM side are going to eventually collapse into a heap of unmanagable dumpster-fire code that they can't understand nor maintain. A whole new world of incidental complexity will consume every project, and in the long run it will just eat them alive (figuratively speaking, of course :-D ).
Some 120 years ago recordings music was a living phenomena produced in the moment. Musicians worked at restaurants and coffee shops everywhere, being useful without being super stars.
Music didn’t disappear with recordings, but the works is certainly different.
With LLMs, we can't tell anymore if something is a labor of love with hundreds of hours of work behind it, or half a dozen prompts to Claude Code.
Yes, it's not deterministic, and if you were using it commercially, the ROI would be terrible, and it's certainly not reliable but for a hobby project.... why not?
Encouraging people to understand the layer of abstractions they're building on is helpful, doesn't matter if they do it by hand or with clankers.
LLMs lower the barrier for execution - they make you faster. The unstated question is: faster at what - they can make you faster at something clever, or faster at the entirely wrong thing...
Your point is correct if we're looking at it through a scarcity lens - the effort to make it certainly decreased a lot - but that doesn't mean that anything is now worthless. We can just move onto doing bigger, better things now, until we hit the next limits...
But we know that long term use of LLMs does not lead to better understanding, it leads to reliance on the LLM for the person to be able to function at their job.
Well, look at this way, the needs of commerce are going to solve the conflict between practical and the beautiful. I think those of that value the beautiful aspects of coding will find new avenues of expression. For example, I'm about to get back into C programming to build a play.date game engine for an MMO.
I'm afraid it's an elite skill in the sense that juggling is also an elite skill. It's impressive for the first few seconds you gaze into it, but once the novelty factor wears off you understand that it's wasted effort that leads to a project that suffers from a massive maintainability problem, is limited in which platforms it can run, and brings no advantage whatsoever. It's an gimmick that has no practice use.
This is the software development equivalent of an amateur guitarist posting shredding videos on YouTube.
In the same way we appreciate Japanese wood joinery, why not not just appreciate this? Someone might even learn a trick or two reading it.
Would we have considered writing a server in C writing it in assembly? No.
Is not the end of the world. Is a change.
You do indeed need to do quite a bit of thinking and problem solving, to build things with an llm.
If you disagree, repeat this project, so you can share with us how little thinking it required.
I mourn the death of a human artform.
We need to stop thinking of software as carpenters where the magic is some physical skill and that is the "CRAFT WE MUST PROTECCT".
And at least your comment was grounded in reality; a lot of people I talk to (who are not coders) seem to think a good software engineer writes every line and every word with thoughtful genius and AI just spams code so one is better than the other. And they are convinced its some naunced smart take and they understand software development on a inner level or whatever.
And the base assumption still holds true (pure AI-generated code is garbage) but its mostly because its badly designed and is still a pretty poor architect. And there is a need to pushback against slop but why do we need to elevate typing code as if its some sacred acctivity? Most of the work a good coder does is in their mind with little connection to the phyiscal reality of the world.
Look, I still got my physical copy of Michael Abrash's Graphics Programming Black Book with its genius content about hand-optimizing cycle count on 486 and Pentium processors, beating compilers at that time.
It was an absolute artform, but completely obsolete by today.
The artform only dies if you let it. Even if your employer is so idiotically myopic as to forbid you to ever write your own code, you can still continue the art on your own time. I for one don't care how "good enough" any AI-lableled technology gets at writing code. I will continue to hone my craft until the day I either die, become too unwell to do it, or some other creative endeavor consumes all of my personal time.
Yes, an LLM can write it, it’ll probably work. Yet, it’ll remain meaningless slop while this is not.
Nothing beats Go.
When you use HTMLX (goat) + sqlc (goat) + pgx (another goat) + Chi (yet another goat) and Sqlite (goat).
Most apps will not need anything more than Sqlite, i've several sqlite apps doing a couple of million visits per day.
Compiles to signal binary blazingly fast.
Deploy using systemd service, capture logs with alloy / Loki graphana setup, set up alerts and monitoring and go home.
And you can serve millions of requests on a server with 512MB RAM.
I don't think you'd ever need more speed than this.
Everything else is bloated, slow and doesn't give you enough room for optimization.
Here's the latency of one of my hobby projects (network latency not included): https://i.ibb.co/hJ6FQtyw/d3d6c9d15765.png
Request rate: https://i.ibb.co/Fq80nfJ4/67fcdbdb7491.png
It's running in US and EU (helps avoid atlantic routrip tax), in this one i am doing some 100s of checks, not simple CRUD work. With Go you can optimize a lot without complexity of Rust.
Congratulations to the OP for the accomplishment.
I wanted to see how an LLM would do writing one in pure 8088 assembler for CGA and it one-shot a nice demo (I fed it the vectors for the Elite ship in the prompt):
Maybe it's finally time to move on from being a career programmer.
> And the only people who will ever be in that position are the ones who take the time and effort, out of sheer curiosity, to learn how things work.
People learn something new all the time, AI does not learn anything, it just simulates and hallucinates. But the core question is not addressed with that. What would you do if you have to compete against AI, and AI is better? We already see these with the new generation of humanoid robots from China. Those things make Boston Dynamics robots look like tinker-toys in comparison - already as-is. Give it ten more years and we finally reached AI skynet for real.
success through cleverness and inventiveness - not yet fucked by AI
achievement through stubborn persistence - you can still dig deep holes in the garden
you still could have a character, if you were lucky
human agency? not yet fucked up, but it's gonna be
achievement earned through one's own qualities or effort? - intact somewhat
As for why it wouldn't run on Linux, there are some pretty big differences in the actual assembly. One pretty superficial difference is calling conventions -- MacOS uses the x16 register for syscall numbers, Linux uses x8. Calling the kernel in Mac uses "svc #0x80", in Linux it's "svc #0". That's ~120 lines that need to be replaced, but easy enough to just use sed. Syscall numbers are all different, as are the struct layouts for sigaction(), MacOS has an "sa_tramp" field that Linux doesn't have. Enforcing max processes is done here using the MacOS-specific proc_info() syscall, which can be used to get the number of children any given process has. Linux doesn't have an equivalent, so process tracking would need to be done differently. Finally, Linux has the getdents64() syscall, rather than getdirentries64(), which uses a different struct and is called differently.
I'm sure an LLM could make all those changes, but it's a pretty large codebase, so it would probably make some mistakes or miss things.
Not all languages use NULL terminated strings. I think Rust actually stores the string length alongside a pointer to the start of the string data. You can do the same in C, but you'd have to do it manually using a struct. In assembly you could do the same thing since you get to decide basically everything.
And super educational. Since then I've been pondering which problems require dropping down to the assembly level. E.g. implementing a JIT compiler, a coroutine runtime, etc.
Really cool project though!
In a comparison between a similar fork-per-connection server written in C and this, I would imagine the throughput would be about the same, because the bottleneck in this model is fork() itself rather than the actual code. It probably matters more for binary size and startup time than requests/sec. Would be fun to actually benchmark, though.
Humbling.
One of my first assembly projects was a CGI Script 100% in x86 assembly.
A full web server is certainly more impressive! Though I'd recommend to beginners to look up CGI and mod_cgi in Apache first lol
The script has been lost to time. I wrote it 5+ computers ago and I don't even know where input that backup...
The overall gist is that CGI Bin specification sets Environmental variables, STDIN and STDOUT to various values. A minimal pure assembly that writes <h1> Hello World </h1> over stdout is your minimalist CGI Script.
A bit of research into what those STDIN/Environmental variables is needed for more. I knew this may e 20+ years ago but have long forgotten....
With access to the various input parameters offered over CGI, you can easily access form data (buttons and whatever clicked by the user). Use some smart file writing to store sessions and off you go....
-------
Maybe start with a Perl CGI tutorial. Then go backwards to C, and finally raw assembly by hand
In general, stable syscall numbers are just a Linux thing. Everyone else uses blessed system libraries
I had a very hard time simply using and even utilizing C++ or Java.
C and Turbo Pascal especially was easier because the compiled code was very much resembling to hand written code.
As the author described, you can do in 4.000 lines what others can do with way less pain in 100.
So you build macros, come up with your own library and in the end you kind of build a meta language build on top of assembly because some lines are so hard to grasp that you delegate working code into a library for reuse.
It is funny how much we take conventions for numbers for granted. If you happen to know assembly and its intricacies you immediately will learn to work with a sign bits which mark negative numbers. But how do you know? Maybe you use the whole addressable space only for positive numbers.
Small things that make a huge different.
Nice article, I enjoyed your adventures and would do the same.
Higher level languages are more convenient for 99% of things, but the directness of Assembly gives me a rush unlike any other. I didn't live through the C64/Amiga, but I was obsessed with old C64/ZX emulators growing up.
Once I was doing 3D I quickly started moving everything but the inner loops to Turbo C, because I'm not a total masochist :)
- Dynamic libraries (e.g. for calling into the kernel, but also user space dynamic libraries) are OS-specific (.so for Linux, .dylib for macOS, .dll for Windows)
- Executable format is OS-specific (ELF for Linux, Mach-O for macOS, PE for Windows)
- Dynamic loading and linkage of both the above are also therefore OS-specific
The last time I did anything in assembler was x86 under DOS. Your code makes ARM64 with a modern OS less scary than I thought it would be.
I've been writing x86 Asm for a few decades. RISCs are simpler in all the wrong ways. After all, "just use a (stupid) compiler" was the whole philosophy.
I'm sure I'm not the only one who has fantasized about doing something like this as a self-soothing enterprise. Kudos to you for actually doing it!
I made a self documented (all the related links are in the file) single makefile (https://gist.github.com/ontouchstart/d3ad8e4d0adf63532303a90...) so that anyone can build ymawky from scratch and dig deeper by tinkering.
I can see the potential of doing some implementation of APL/J/K at levels even lower than C, like how those guys did APL\360 using assembly language. It is going to be super fun in the era of everyone using LLM to pump out verbose Python/TS/Rust code with context windows bigger than the whole operating system.
The absolute best projects ever posted here are the ones made for no other reason than “just because I wanted to”.
jk. Metal as fuck. Love it.
Example: spend time looking at health/nutrition. I assure you that in 5 years you will get more satisfaction and returns than the assembly code you wrote.
One important caveat - the subject of health/nutrition is SIGNIFICANTLY more complicated than assembly code, and not may sources out there know what they are talking about. Computers are child's play compared to biology.
P.S.: I would love a copy of that book please!
Bookmarking this — going to read the source on my Sunday off.
It would be awesome to read a blog post about the project. Your approach, lessons learned, unexpected stuff, etc.
...
