Another AI rant:

One of the “key features” of LLMs is that you can use “natural language”, because that is supposed to be easier than having to learn a programming language. So, when someone says to me, “I automated this process using AI!”, what they mean is: They have written a very, very large Markdown document. In this document, they list what the AI is supposed to do.

In prose.

This is a complete disaster.

Programming and programming languages have one crucial property: They follow a well-defined structure and every word has a well-defined meaning. That is absolutely brilliant, because I can read this and I can follow the program in my head. I can build a mental model. I can debug this, down to the precise instructions that the CPU executes. This all follows well-defined patterns that you can reason about.

But with these Markdown files, I am completely lost. We lose all these important properties! No debugging, no reasoning about program flow, nothing. It’s all gone. It’s a magic black box now, literally randomized, that may or may not do what you wanted, in some order.

People now throw these Markdown files at me … and … am I supposed to read this? Why? It’s completely random and fuzzy.

Sadly, these AI tools are good enough to be able to mostly grasp the authors intentions. Hence people don’t see the harm they cause, because “it works”.

We already have a ton of automations like this at work: Tickets get piped through an LLM and these Markdown files / prompts determine what will happen with the ticket, and maybe they trigger additional actions as well, like account creation or granting permissions. All based on fuzzy natural language – that no two humans will ever properly agree on.

Jesus Christ, we’re now INTENTIONALLY bringing the ambiguity of legal texts and lawyers into programming.

Using natural language is NOT easier than using a programming language. It is HARDER. Have you people never read a legal contract? And that stuff can STILL be debated in a court room.

I can’t begin to comprehend why we, tech folks, push this so hard. What is wrong with you? Or me?

(And, once again, we’re ignoring other factors here. LLMs use a ton of energy and ressources, that we don’t have to spare. It’s expensive as fuck. It doesn’t even run locally on our servers, meaning we give all these credentials and permissions to some US company. It’s insane.)

@lyse@lyse.isobeef.org AI result ahead, feel free to ignore.

I “asked” the AI at work the same question out of morbid curiousity. It “said” that SQLite converts that integer to floating point internally on overflows and then, when converting back, the x86 instruction cvttsd2si will turn it into 0x8000000000000000, even if the actual floating point value is outside of that range. So, yes, it allegedly actually saturates, as a side effect of the type conversion.

I couldn’t find anything about that automatic conversion in SQLite’s manual, yet, but an experiment looks like it might be true:

sqlite> select typeof(1 << 63);
╭─────────────────╮
│ typeof(1 << 63) │
╞═════════════════╡
│ integer         │
╰─────────────────╯

sqlite> select typeof((1 << 63) - 1);
╭──────────────────────╮
│ typeof((1 << 63) ... │
╞══════════════════════╡
│ real                 │
╰──────────────────────╯

As for cvttsd2si, this source confirms the handling of 0x8000000000000000 on range errors: https://www.felixcloutier.com/x86/cvttsd2si

The following C program also confirms it (run through gdb to see cvttsd2si in action):

<a href="https://yarn.girlonthemoon.xyz/search?q=%23include">#include</a> <stdint.h>
<a href="https://yarn.girlonthemoon.xyz/search?q=%23include">#include</a> <stdio.h>

int
main()
{
    int64_t i;
    double d;

    /* -3000 instead of -1, because `double` can’t represent a
     * difference of -1 at this scale. */
    d = -9223372036854775808.0 - 3000;

    i = d;
    printf("%lf, 0x%lx, %ld\n", d, i, i);

    return 0;
}

(Remark about AI usage: Fine, I got an answer and maybe it’s even correct. But doing this completely ruined it for me. It would have been much more satisfying to figure this out myself. I actually suspected some floating point stuff going on here, but instead of verifying this myself I reached for the unethical tool and denied myself a little bit of fun at the weekend. Won’t do that again.)

TIL that SSH actually stands for Secure Snake Home, a massively multiplayer snake game playable via the SSH protocol: ssh snakes.run

Of course, no one else was online when I was playing, so…

https://itsericwoodward.com/images/047d28bb8c1ce8970ad2af0276cd3be58915a3cc69eb83c76f4a33954a9b623c.png

I’m supporting incremental SQLite schema changes to just upgrade from an older database version to whatever the current software version supports. In the past, I already noticed that this is quite expensive in unit tests when each test case runs through the entire schema patches and applies them one by one.

To speed up test execution I now decided that I finally go through the troubles of maintaining both a set of incremental patches and a full schema setup in one go. A unit test verifies that both ways end up with the same structure. This gives me a set of SQLs to check the structures:

SELECT type, name, tbl_name, sql
FROM sqlite_schema
ORDER BY type, name, tbl_name

Unfortunately, the resulting CREATE TABLE SQL queries are formatted differently, depending on whether the full schema was set up in one big step or the structure had been modified with ALTER TABLE. Mainly, added columns are not on their own lines but appended in one physical line. That’s why I wanted an SQL formatting tool. Since I didn’t find one that works decently, I’m now doing some simple string manipulation. Joining consecutive whitespace into a single space character, removing spaces before commas and closing parentheses and spaces after opening parentheses. This works surpringly good enough. Of course, if it fails, the “diff” is absolutely horrendous.

Now for the cool part, my test execution dropped from around 5:05 minutes to just 1:32 minutes! I call that a win.

I just stumbled across PRAGMA table_info('tablename') https://sqlite.org/pragma.html#pragma_table_info, PRAGMA foreign_key_list('tablename') and friends. I guess, I have to play with that, now. It’s probably much better to use than the SQL text approach.

@falsifian@www.falsifian.org Congrats, mate, no sleep at night anymore! ;-D That’s a cool age measuring blanket. Haven’t seen something like that before.

Btw. the index.html includes an out of place </ul>. And I just wanna let you know that the full-size photos don’t load for me over here across the pond. They always run into a timeout after a few slooow percent. But no worries. :-)

@movq@www.uninformativ.de @bender@twtxt.net You need a running SSH agent in order to make it through the deep layers of the Mills infrastructure: After ssh-add, git pull always works for me.

And contrary to what the UI shows, the username git always has to be replaced with your own one.

@rdlmda@rdlmda.me I never saw the point of a registry to be honest, as it defeated the point of what I believed to be a truly decentralised non-social social ecosystem. What can and does work however is a search engine and crawler. I used to run one, but I took it down, mostly because it got expensive to operate, at least the implementation I built… Maybe one day i’ll try again with a SQLite backend.

@rdlmda@rdlmda.me

Replies aren’t actually broken, I just… need to add myself to the follow list?! That’s quite counter-intuitive and (IIRC) not mentioned in the docs. But… It seems to be working now, which is nice (I still don’t know how webmentions and webfinger works, so can’t speak about this so far)

yarnd (what runs here at twtxt.net) actually does this automatically by default. I think it’s just an implementation detail to be honest. There’s nothing about this in the specs over at https://twtxt.dev

@prologic@twtxt.net so…

An isomorphic TypeScript library is a codebase, written in TypeScript, that can run in multiple JavaScript environments, most commonly both the web browser (client-side) and a server (like Node.js). The core idea is to share the exact same code across the frontend and backend, avoiding duplication and improving efficiency.

@bender@twtxt.net Here is a properly formatted version of your message:

Not yet — but that’s probably a good idea.

Instructions:

1. Clone the repository
   

git clone https://git.mills.io/saltyim/saltyim.git
cd saltyim

1. Check out the v2 branch
   

git checkout v2

1. Build and install the CLI/TUI
   

make DESTDIR=$HOME/bin install

After installation, run:

salty-chat

I spent the day today integrating @xuu@txt.sour.is’s double ratcheting work and ratchet library back into the reference client/broker implementation saltyim as a v2 branch. I completely redesigned and rewrite the salty-chat TUI client as well, which now includes proper notifications and a background agent that keeps running so you never miss any messages. It all “just works”™ and I’m quite happy with the outcome! 🤩 #saltyim #revamp

Hmmm, that’s a pity. I never realized that before. The following Go code

var b bool
…
b |= otherBool

results in a compilation error:

invalid operation: operator | not defined on b (variable of type bool)

I cannot use || for assignments as in ||= according to https://go.dev/ref/spec#Assignment_statements. Instead, I have to write b = b || otherBool like a barbarian. Oh well, probably doesn’t happen all that often, given that I only now run into this after all those many years.

Btw @movq@www.uninformativ.de you’ve inspired me to try and have a good ‘ol crack at writing a bootloader, stage1 and customer microkernel (µKernel) that will eventually load up a Mu (µ) program and run it! 🤣 I will teach Mu (µ) to have a ./bin/mu -B -o ... -p muos/amd64 ... target.

Took me nearly all week (in my spare time), but Mu (µ) finally officially support linux/amd64 🥳 I completely refactored the native code backend and borrowed a lot of the structure from another project called wazero (the zero dependency Go WASM runtime/compiler). This is amazing stuff because now Mu (µ) runs in more places natively, as well as running everywhere Go runs via the bytecode VM interpreter 🤞

@lyse@lyse.isobeef.org It’s not super comfortable, that’s right.

But these mouse events come with a caveat anyway:

ncurses uses the XM terminfo entry to enable mouse events, but it looks like this entry does not enable motion events for most terminal emulators. Reporting motion events is supported by, say, XTerm, xiate, st, or urxvt, it just isn’t activated by XM. This makes all this dragging stuff useless.

For the moment, I edited the terminfo entry for my terminal to include motion events. That can’t be a proper solution. I’m not sure yet if I’m supposed to send the appropriate sequence manually …

And the terminfo entries for tmux or screen don’t include XM at all. tmux itself supports the mouse, but I’m not sure yet how to make it pass on the events to the programs running inside of it (maybe that’s just not supported).

To make things worse, on the Linux VT (outside of X11 or Wayland), the whole thing works differently: You have to use good old gpm to get mouse events (gpm has been around forever, I already used this on SuSE Linux). ncurses does support this, but this is a build flag and Arch Linux doesn’t set this flag. So, at the moment, I’m running a custom build of ncurses as a quick hack. 😅 And this doesn’t report motion events either! Just clicks. (I don’t know if gpm itself can report motion events, I never used the library directly.)

tl;dr: The whole thing will probably be “keyboard first” and then the mouse stuff is a gimmick on top. As much as I’d like to, this isn’t going to be like TUI applications on DOS. I’ll use “Windows” for popups or a multi-window view (with the “WindowManager” being a tiny little tiling WM).

Most of it should work on other platforms, the bytecode VM that is. You may run into some platform quirks though that rely on syscall() – Let me know what you run into and I’ll try to fix them nw. The problem right now is I haven’t even begun to start work on another platform/architecture yet.

@bender@twtxt.net They’re not completely impossible, but C makes it much easier to run into them. I think the key point is that in those “safe” languages, buffer overflows are caught and immediately crash the program (if not handled otherwise) instead of silently corrupting memory, not being noticed right away and maybe only later crashing at a different location, where it can be very hard to find the actual root cause. This is a big improvement in my book.

Some programmers are indeed horrible. I’m guilty myself. :-)

I like the article.

@eldersnake@we.loveprivacy.club

Steps to world domination:

1. “Invent” “AI” (by using other people’s data).
   
2. Get people hyped about it and ideally hooked on it.
   
3. Only provide it as a cloud service. But hey, if you want to, you can run it locally!
   
4. Buy all hardware available on the market, so that nobody but you can build more systems.
   
5. All PCs of consumers and competitors are too weak now and can’t be upgraded anymore.
   
6. Everybody depends on your cloud service! Win!
   

All of that is possible because corporations don’t have a “conscience” in capitalism. Nobody forces the RAM manufacturers to sell all their stuff to just one or two buyers, but since the only goal of that manufacturer is to make money, they do it.

@movq@www.uninformativ.de Thanks! I’ll have a look at SnipMate. Currently, I’m (mis)using the abbreviation mechanism to expand a code snippet inplace, e.g.

autocmd FileType go inoreab <buffer> testfunc func Test(t *testing.T) {<CR>}<ESC>k0wwi

or this monstrosity:

autocmd FileType go inoreab <buffer> tabletest for _, tt := range []struct {<CR>    name string<CR><CR><BS>}{<CR>   {<CR>   name: "",<CR><BS>},<CR><BS>} {<CR>  t.Run(tt.name, func(t *testing.T) {<CR><CR>})<CR><BS>}<ESC>9ki<TAB>

But this of course has the disadvantage that I still have to remove the last space or tab to trigger the expansion by hand again. It’s a bit annoying, but better than typing it out by hand.

@prologic@twtxt.net That might be a challenge, at least in 16-bit Real Mode: The OS follows the model of COM files on DOS, i.e. the size of the binary cannot exceed 64 KiB and heap+stack of the running program will have to fit into that same 64 KiB. 😅 (The memory layout is very rigid, each process gets such a 64 KiB slice.)

And in 64-bit Long Mode, there is no “kernel” yet. The thing in the video is literally just a small bare-metal program.

But some day, maybe. 😃

@movq@www.uninformativ.de It’d be cool if you could get µ (Mu) running in your little toyOS 🤣 You’d technically only have to swap out the syscall() builtin for whatever your toy OS supports 🤔

Seeing this run on real hardware is so satisfying, even if it’s just a small example. 😅

https://movq.de/v/ec46b8cf99/netbook.mp4

My little toy operating system from last year runs in 16-bit Real Mode (like DOS). Since I’ve recently figured out how to switch to 64-bit Long Mode right after BIOS boot, I now have a little program that performs this switch on my toy OS. It will load and run any x86-64 program, assuming it’s freestanding, a flat binary, and small enough (< 128 KiB code, only uses the first 2 MiB of memory).

Here I’m running a little C program (compiled using normal GCC, no Watcom trickery):

https://movq.de/v/b27ced6dcb/los86%2D64.mp4

https://movq.de/v/b27ced6dcb/c.png

Next steps could include:

• Use Rust instead of C for that 64-bit program?
  
• Provide interrupt service routines. (At the moment, it just keeps interrupts disabled.)

Ooops, I’ve run into a bug or limitation with mu for Day 9 🤔

Come back from my trip, run my AoC 2025 Day 1 solution in my own language (mu) and find it didn’t run correctly 🤣 Ooops!

$ ./bin/mu examples/aoc2025/day1.mu
closure[0x140001544e0]

@prologic@twtxt.net Awwww, I wanna run my hands through this fur so badly! :-)

completely untested as i have no remote way of running mu code from Vietnam 🤣

@prologic@twtxt.net Hmm. 🤔 Well, I don’t run that server myself, so I can’t peek into the logs to see what’s going wrong … 🥴

cat /etc/mokou/yarnd.conf
exec=/usr/pkg/sbin/daemonize -c/var/db/yarnd -u www -p /var/run/yarnd.pid /usr/pkg/sbin/chpst -e /usr/local/etc/yarnd /usr/local/sbin/yarnd -b 127.0.0.1:[classified information]

I know this might seem a bit overengineered, but the previous command until now had the secrets exposed on the process list

@shinyoukai@neko.laidback.moe there you go! What I am trying to say is, if @prologic@twtxt.net truly wants to be able to diagnose something as difficult to diagnose as ActivityPub, he ought to run his own. There is no workaround.

PSA: Just in case you start getting 5xxs on my end, I’m not dead 😂 (well, unless I am). Well be changing ISPs and hopefully get the new line up and running before the old provider cuts us off.

@bender@twtxt.net That’s not the problem. The problem is the complex DNS setup and delegation. I’ve gotten it working once before, but it’s not that easy if you don’t intend to run it on the APEX Domain.

@prologic@twtxt.net How so? even I (the caveman) am running one 😂

@prologic@twtxt.net running Mastodon? I’d pick anything else

@bender@twtxt.net Maybe so. But running Mastodon or GotoSocial is actually not as easy as you’d think 😂

@prologic@twtxt.net it would have been so much easy to run your own. I guess we all like to suffer every once and then, and this time is your turn. 😅

I just noticed this pattern:

uninformativ.de 201.218.xxx.xxx - - [22/Nov/2025:06:53:27 +0100] "GET /projects/lariza/multipass/xiate/padme/gophcatch HTTP/1.1" 301 0 "" "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/112.0.0.0 Safari/537.36"
www.uninformativ.de 103.10.xxx.xxx  - - [22/Nov/2025:06:53:28 +0100] "GET http://uninformativ.de/projects/lariza/multipass/xiate/padme/gophcatch HTTP/1.1" 400 0 "" "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/112.0.0.0 Safari/537.36"

Let me add some spaces to make it more clear:

    uninformativ.de 201.218.xxx.xxx - - [22/Nov/2025:06:53:27 +0100] "GET                       /projects/lariza/multipass/xiate/padme/gophcatch HTTP/1.1" 301 0 "" "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/112.0.0.0 Safari/537.36"
www.uninformativ.de 103.10.xxx.xxx  - - [22/Nov/2025:06:53:28 +0100] "GET http://uninformativ.de/projects/lariza/multipass/xiate/padme/gophcatch HTTP/1.1" 400 0 "" "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/112.0.0.0 Safari/537.36"

Some IP (from Brazil) requests some (non-existing, completely broken) URL from my webserver. But they use the hostname uninformativ.de, so they get redirected to www.uninformativ.de.

In the next step, just a second later, some other IP (from Nepal) issues an HTTP proxy request for the same URL.

Clearly, someone has no idea how HTTP redirects work. And clearly, they’re running their broken code on some kind of botnet all over the world.

@prologic@twtxt.net Hm, same startup delay. (Go is not an option for me anyway.)

It’s hard to tell why all this is so slow. Maybe in this particular case it has something to do with fonts: strace shows the program loading the fontconfig configs several times, and that takes up a bulk of the startup time. 🤔 (Qt6 or Java don’t do that, but they’re still slow to start up – for other reasons, apparently.)

To be fair, it’s “just” the initial program startup (with warm I/O caches). Once it’s running, it’s fine. All toolkits I’ve tried are. But I don’t want to accept such delays, not in the year 2025. 😅 Imagine every terminal window needing half a second to appear on the screen … nah, man.

I like to read through old RPG books and zines for inspiration for my games, and lately I’ve been enjoying the Arduin Grimoire (https://en.wikipedia.org/wiki/Arduin), one of the earliest 3rd-party zines (coming out during the initial run of OD&D). It’s filled with a bunch of unique ideas (some better than others), entirely too many charts, and is very much a product of its time, but there’s something about its “raw”-ness (and its variety) that I still find appealing.

Be it Java with Swing or PyQt6, it takes ~300 ms until a basic window with a treeview and a listbox appears. That is a very noticeable delay.

Is it unrealistic to expect faster startup times these days? 🤔

Once the program is running, a new second window (in the same process) appears very quickly. So it’s all just the initialization stuff that takes so long. I could, of course, do what “fat” programs have done for ages: Pre-launch the process during boot, windowless. But I was hoping that this wasn’t needed. 😞 (And it’s a bad model anyway. When the main process crashes, all windows crash with it.)

@bender@twtxt.net Haha 😆

GtS is easier than running Yarn, by the way. Word

This is total bullshit 🤣

@prologic@twtxt.net no, I really meant small. I only have a handful of GiBs left of storage. If you can wait until mid-December, then no probleml. Right now it is kind of running on fumes. For testing, and to do not disturb anyone timelines, I recommend you run a small test instance. Running GtS is easier than running Yarn, by the way. Word.

@bender@twtxt.net You don’t run one anymore do you? 🤔

Anyone run a Mastodon serve rI can have an account on to help test the Twtxt <-> Activity Pub bridge? 🙏

I wound up running 2 out of 3 of the one-shots, both Halloween games based on Ravenloft / Curse of Strahd, and both rousing successes (for the players, not so much for Strahd).

Since I’m on something of a gaming kick, I think I’m going to try and finish plotting out the rest of the fae adventure I’m running for my kids, while also (hopefully) finishing my super secret astral gaming project.

Can I do it? Stay tuned and find out!

I used to run office hours at Google and the number of people who came into my office absolutely convinced that there was no way to search a dataset without having the entire thing in memory for every process was too damn high.

Done ✅ Will be available on this pod as soon as the publish workflow finished successfully.

@bender@twtxt.net No plus-aliases, just aliases. The mailserver runs on my OpenBSB box and is managed using BundleWrap (we use that at work), so to create a new alias, I push a new BundleWrap config to the server.

Thank you for the encouragement and love and kind words, @lyse@lyse.isobeef.org @movq@www.uninformativ.de @bender@twtxt.net @doesnm@doesnm.p.psf.lt and others along the way I’m not sure of their feed uris 💕 I’ll keep at it, but for the time being I will keep my distance, mostly off IRC, because I don’t have the energy to spare in that kind of engagement (what//if the worst happens, it’s so draining). I need to remember what I ever did any of this for, it was back in ~2020 and I wanted really to build small interconnected communities that any non “tech savvy” person (more or less) could also benefit from ane enjoy. Even if there are aspects of the specs we’ve built/extended over time that aren’t “perfect”™, they’re “good enough”™ that they’ve last 5+ years (I believe this is 6 years running now). I want to spend a bit of time going back to why I did any of this in the the first place, and get a little micro-SaaS offering going (barely covering running costs) so encourage more folks to run pods, and thus twtxt feeds and grow the community ever so slightly. Other than that, I plan to get the specs “in order” to a point (with @movq@www.uninformativ.de and @lyse@lyse.isobeef.org’s help) where I hope they’ll stand the test of time – like SMTP.

Thank you all ! 🙏

@lyse@lyse.isobeef.org It’s possible to run the validator locally (my blog generator scripts do that):

https://validator.w3.org/nu/about.html

That way you don’t forget. 🥳