datenwolf

Hacker, Physicist, CCC, µC³, Chaotic Neutral, INTJ-T – 𓍘𓅱𓀁𓇋𓄿𓅱𓀗 – recruiting an army of corvids. /// followbacks mostly only for other birders.

OrcID: https://orcid.org/0000-0002-9413-3474
Amateur Radio Callsign: DL1WXD
Pronouns: * / ally
Gender: cis het male

Public Key

npub1mkmzc7ttrmlp3vzthpq6ztd6923dv4yf507vee99q7euu7rhwldqsuc0eu

NIP-05 Address

datenwolf_at_chaos.social@momostr.pink

Profile Code

nprofile1qqsdmd3v0943alsckp9mssdp9kaz4gkk2jy68lxvujjs0v7w0pmh0ksprpmhxue69uhhyetvv9ujumt0d4hhxarj9ecxjmntqyt8wumn8ghj7un9d3shjtnswf5k6ctv9ehx2aqzkeg76

Publishing to

relay.momostr.pink

relay.primal.net

offchain.pub

Author Public Key

npub1mkmzc7ttrmlp3vzthpq6ztd6923dv4yf507vee99q7euu7rhwldqsuc0eu

Show more details

Last Notes

2026-10-07 08:00:43 UTC

- reply

Anyway, please send me paper recommendations showing the measurably ill effects of the use of LLMs in academic use. Every angle you can think of:
- deskilling effects
- plagiarism and copyright
- is the generated code actually licensable?
- measurable vs. perceived productivity
- technical debt
- code quality / readability
…
You want a #seminar on #slop – okay (#malicious_compliance activated).
Please #boost

2026-10-07 07:57:00 UTC

Oh FFS… one of the higher ups at my place of work (biophysics and optics) was asking me, if I was willing to give a seminar on #AI and how to effectively use it. For context: I'm surrounded by sloperators; even my boss has become addicted.
The number of times I'm told "you could be so much more productive, if you just used AI" (no I wouldn't because I don't measure productivity in LoC).
Anyway, if I'm going to give that seminar, it's going to be a "don't use #slopbots" angle.

2026-09-16 08:36:57 UTC

- reply

Rust und "responsible supply chain" in einem Satz… guter Witz… sehr guter Witz.
Cargo/crates.io, npm, pip und wie sie alle heißen: Alles die selbe Sosse.

2026-09-12 16:03:39 UTC

Dear #retrogaming #retrocomputing enthusiasts living in #northerngermany (or even further out):
At our research institute there's a Sony PVM-20M2MDE professional #Trinitron #CRT monitor that's headed for e-waste if there's noone to give it a new home. It has to be gone by the end of next week.
If you want it shoot me a DM, but you'll have to come by, to pick it up.
https://crtdatabase.com/crts/sony/sony-pvm-20m2mdu
#norddeutschland
https://assets.chaos.social/media_attachments/files/117/258/912/558/495/495/original/2cec51e69c93cc7a.jpg
https://assets.chaos.social/media_attachments/files/117/258/912/573/672/905/original/2bba4e24e584b69e.jpg
https://assets.chaos.social/media_attachments/files/117/258/912/581/944/515/original/e7dd889ad8cf97dc.jpg

2026-08-26 16:30:35 UTC

- reply

The QR-Code is a nice touch 😆

2026-07-26 07:48:13 UTC

- reply

I've watched the development the same. And at some point, being fed up with the problems of udev/tmpfs at boot time I just blurted into some maillist a question along the lines "why can't just have a tmpfs that get's overlaid/merged/populated by a devfs? Something like a tmpdevfs?" Must have been around 2004.

2026-07-26 06:31:47 UTC

- reply

Re: udev devfs
For quite some time now the standard ware of doing things is in Linux is devtmpfs + udev.
devtmpfs is a variant of tmpfs, that automatically gets de-/populated with device inodes by the kernel as they come and go. It will then call into udev to initiate run-time operations from user space in reaction to devices being added or removed. One can live entirely without udev, at the lack of persistent naming (usefull for resource limited, static stuff).

2026-07-20 08:15:57 UTC

- reply

Money isn't anonymized trust though.
The purpose of money is quantifying and measuring availability, allocation and expenditure of limited resources. Everyday experience applies it on tangible resources, any intangible things alike, like time and raw materials being spent on making something.
Following it to the conclusion, money actually is just another measure for energy. And purchasing power a measure for its entropy.
Having capital means you have command over energy to expend.

2026-06-12 16:36:18 UTC

- reply

But, as far as my line of work goes, because LLMs and coding agents (still) struggle to "see" (they don't see, they're "just" extremizing the Laplacian action between the prompt and the generated token stream) the big picture, they still tend to produce terrible code.
Every now and them – with a lot of hand holding – I test them on the finicky problems I had to solve as a baseline. The code that comes out often resembles my naive, low perf PoCs.
3/3

2026-06-12 16:30:36 UTC

- reply

Regarding security work: I actually foresaw the enormous capabilities of ISA/programming language coupled LLMs for the discovery of vulnerabilities. As far as reversing binaries go, I have to agree that LLMs are quite effective deobfuscators/prettifiers. You can throw at them x86 code naively decompiled to C and LLMs will do a quite impressive job.
2/
I posted a somewhat provocative remark toward this 2 years ago on Xitter: https://x.com/datenwolf/status/1798261751714295973
https://assets.chaos.social/media_attachments/files/116/738/075/942/494/417/original/6709f9d31a0bac79.png

2026-06-12 16:20:11 UTC

- reply

Regarding navigating codebases: Personally I take great enjoyment in figuring out the low level stuff. You're a demoscener? Isn't most of the fun in discovering the workings of the machine?
Regarding boilerplate: Lots of boilerplate is a sign of a poor foundational design. My current project actually is a large-ish refactoring task with the aim do eliminate most of the boilerplate that did accumulate. It's about 80% theory building, not coding.
1/

2026-06-12 16:05:26 UTC

- reply

Out of curiosity (I mean this honestly!):
What benefits do you get out of LLMs?
(P.S / Re seniority: Every now and then I cross paths with people much younger than me, which I immediately recognize as far more skilled and intelligent than I am, and I usually get excitement and amazement from their feats. Laps around the Sun in age or developer years IMHO is a very poor metric).

2026-06-12 14:04:10 UTC

- reply

At the end of the 30 minutes time limit the state polling priority queue was messed up, leading to the laser controller being hammered with bursts of redundant state requests, a couple of particular corner cases in the implementation of the command were not dealt with (maximum frame size of comms buffers, requiring diagnostics data traces to be fragmented).
Eventually I aborted the mess, and just wrote it down myself. 10 minute job.
3/3

2026-06-12 14:00:55 UTC

- reply

Most recently they urged me to try them out to add support for a new command added the the lasers our company makes in the complementing communications library I maintain. Since I built the library for extensibility doing it fully manual was a 10 minute task, including writing tests.
I set a hard time limit of 30 minutes using Claude code to add the functionality.
2/

2026-06-12 13:57:13 UTC

- reply

I spend most of my time earning money with developing software (specifically realtime RF signal processing and image conversion using GPUs). In the past I've written ROTS kernels, networking stacks, laser control firmware to name a few. I also happen to know exactly how LLMs work and are trained.
From time to time coworkers try to "convince" me of the usefulness of coding agents.
1/

2026-06-12 13:37:59 UTC

- reply

I'd like to encourage you to read up on the Gell-Mann-Amnesia Effect (the term was coined decades ago):
https://en.wikipedia.org/wiki/Michael_Crichton#%22Gell-Mann_amnesia_effect%22
In short: When reading reports of any kind one will immediately recognize misrepresentations or outright fabrications in the own field of expertise, even if presented confidently with eloquence. But outside the own expertise one is oblivious if presented the same way.
I keep being unimpressed by AI systems in my field of expertise.

2026-06-08 17:43:29 UTC

- reply

Here's some ammunition to shoot it down:
https://www.nature.com/articles/s41593-025-02132-9

2026-06-04 09:31:11 UTC

- reply

If you're putting lots of text on your slides, you're doing it wrong anyway. The slides are stage dressing for your talk, they NOT ARE your talk!
Put in your slides which can't be communicated verbally. Like photos of your experiments, data graphs, figures, the likes.
Example:
https://assets.chaos.social/media_attachments/files/116/691/116/641/993/653/original/6d0168abdc8760b9.png
https://assets.chaos.social/media_attachments/files/116/691/118/816/195/259/original/f15768f36752e0ec.png
https://assets.chaos.social/media_attachments/files/116/691/120/841/291/149/original/abeb19a6e4c00d84.png
https://assets.chaos.social/media_attachments/files/116/691/124/734/754/526/original/87bccee935c6d5ec.png

2026-06-02 21:59:09 UTC

- reply

Apparently this a.out performs simulation of Mott scattering. https://en.wikipedia.org/wiki/Mott_scattering
And if I'm not mistaken there's also the Fortran-77 source code for it. No license information, and I honestly don't remember if I wrote this, or if we were given the code by our TA.

2026-06-02 21:35:24 UTC

- reply

I just did a `find -name a.out` on my homedir, and apart from the CMake occassional configuration artifacts, there's one a.out that apparently belongs to one of the physics lab courses of my undergrad years, dating to August 2008.
Total XKCD-1360 situation discovery.

2026-05-27 19:33:51 UTC

- reply

No. It doesn't take external input to be inspired. It's a BS argument, an excuse used by (lazy) people to rationalize their behavior.
Here's a suggestion: Ask your favorite GenAI tool to create an entirely new conceived language (conlang), with a fully fledged grammar and rich vocabulary, under the codition that it's completely removed from any existing human language.
Current GenAI tools are incapable of this task (because they don't know what language is).

2026-05-27 19:23:41 UTC

- reply

No. The history of art is full of cases where a daring artist tries something so new, that it lacks any form of preconceived example.
Usually it's those pieces of art that are considered "scandalous" at first.
Also trained GenAI systems are not inspired. The training process is an extremely efficient lossy data compression method, and inference is a fuzzy keyed value lookup.
With the right prompts you can extract pretty much the original input data.

2026-05-27 19:08:52 UTC

- reply

In visual arts, prompting at best can be compared to writing a screenplay, whereas the actual artistic creation is the creative work of set builders, grips, cinematographers and special and visual effect artists.
In the case of AI what you have a humonguous set of stolen works that have been numerically correlated between different domains of expression, and a machine that pieces together blurps matching the statistical profile of the prompt, then smoothes the seams.

2026-05-27 18:46:48 UTC

- reply

"The new instruments" / synthesizers argument is bullshit. It's bullshit in the same was as the "photography vs. painting / drawing" argument is bullshit.
Playing a synthesizer requires skill. Creating a patch for a synthesizer requires creativity.
Phtotography is a form of artistic expression that requires skill and creativity. A skilled photographer can create art with the crapiest cameras, working the medium. But handing a Hasselblad to a noob will not make good photos.

2026-05-16 10:58:56 UTC

- reply

This!
Also worth pointing out: LLMs are essentially "love bombing" their user. A technique (among others) used by cult leaders to isolate their marks from their social framework.
I've said it before, and I'll keep saying it: LLMs are cult indoctrination machines. They're short circuiting mental fault detection mechanisms that evolved for purpose of breaking cognitive feedback loops. LLMs create, amplify and steer cognitive feedback loops.

2026-04-01 13:01:34 UTC

- reply

https://assets.chaos.social/media_attachments/files/116/329/569/937/695/816/original/1b5ebb26078e1a38.jpg

2026-04-01 12:57:07 UTC

- reply

Darf ich vorschlagen, dass Du alle Zwischenschritte überspringst und direkt zum Acetylenschneidbrenner greifst?
Als Flüssigkeit kann es definitiv nicht mehr festgrefressen sein ;-)

2026-03-21 11:13:13 UTC

- reply

And once again the Zig language project shows, that there's a better way: https://ziglang.org/code-of-conduct/#strict-no-llm-no-ai-policy

2026-01-29 16:45:59 UTC

- reply

If you were in need of DDR5, I have two kits of 2x 32GiB = 64 GiB 6400MT/s that I misordered by accident just days before the price hikes started.
For trusted people in the FOSS community who are in serious need for RAM, I'd part with them for a price close to what I bought them at.

2025-11-26 16:06:24 UTC

- reply

I take it, that the name of the parrot is "stochastic", right?

2025-10-26 21:35:29 UTC

- reply

if you find it, can you pass it on? I'd like to read _that_, too.

2025-10-23 17:31:01 UTC

- reply

1. Yes
2. No
3. No - specifically those "features" are often XY-problem emanations
4. Yes, and hence contradicts 2
5. Yes, sort of. The fact that you actually can shell script X11, but (by design, not directly) Wayland / re competitors is often reason enough to not use it.

2025-09-12 08:19:50 UTC

Frage an die #Solar & #Akku Bubble: Ich würde mir gerne für Camping / Kleingarten (& Notstrom) eine #Powerbank die sich auch direkt per #Photovoltaik DC aufladen lässt zulegen; etwas so um die 1kWh (+/- 25%).
Früher (vor >20 Jahren) hätte (sogar "habe") ich mir sowas selber gebastelt, aber dafür fehlt mir jetzt die Zeit (und der Nerv).
Es gibt ja einige Offerten von diversen Herstellen. Gibt es gute Empfehlungen?

2025-07-23 10:02:05 UTC

#Cursed #programming idea: When autogenerating filenames escape characters that are unsafe to use in file path and command line with their corresponding Unicode Fullwidth Form counterparts:
/ → ／
\ → ＼
# → ＃
! → ！
" → ＂
' → ＇
( → （
) → ）
[ → ［
] → ］
{ → ｛
} → ｝
| → ｜
* → ＊
+ → ＋
, → ，
. → ．
- → －
: → ：
; → ；
< → ＜
> → ＞
= → ＝
? → ？

2025-07-09 14:01:48 UTC

- reply

> You don’t have to be a pilot to design an aeroplane cockpit.
Yes, you do. In all major airplane companies user interface design decisions directly involve the end users (i.e. pilots); it's a lesson learnt the hard and bloody way.

2025-06-28 17:46:48 UTC

- reply

Well that much I understood. Which is why I want to read it, so that I know the roadmap – in order to know where to place the roadblocks or maybe even turn parts of the path into a moat proactively.

2025-06-28 17:37:09 UTC

- reply

May I ask you to reconsider, not sharing this study? The people who work hard to suck people into Gen AI will read it anyway, passing it along by word of mouth. And it's difficult to build social defenses if you don't know what to look out for and be wary of.

2025-01-15 21:46:01 UTC

How a Tesla autopilot computer vision system interprets the visual input when confronted with a train. 😆
@npub1z7x…q5jf
via: https://labyrinth.zone/notice/Aq5V74iVZRc1GyRCm8
(EDIT: rephrased for clarity)
https://assets.chaos.social/media_attachments/files/113/834/558/306/205/219/original/55ec2f82c593fa48.mp4

2024-12-27 09:51:22 UTC

- reply

Dieses Schlupfloch ist IIRC in der Petition (habe ich vor gefühlt einer Ewigkeit gezeichnet) explizit adressiert.

2024-10-15 16:42:59 UTC

- reply

Jupiter has the strongest, largest magnetosphere of all the planets in the whole solar system; it even reaches into the inner system and directly interacts with the solar magnetosphere. This gives Jupiter the monster version of Van-Allen belts. The inner jovian moons pass through these belts. Io is the most extreme, its volcanic eruptions moving so much plasma into space, that it creates a radiation channel between Jupiter and its radiation belts.

2024-10-01 08:09:49 UTC

- reply

So what will it be? Panic buying a whole bunch of essential ICs, CPUs, GPUs, memory? Or investing in Wacker Chemie stock, under the presumption, that they're going to increase their polysilicon purification and synthetic quartz/silica production capabilities…?
Decisions to be made.

2024-09-06 21:41:45 UTC

- reply

weird cognitive biases, maybe. Maybe because I'm terrible with names all the time, and they have a much much weaker associative weight in my mind, than the actual work and ideas of a person.
Fun fact about me: I really, really struggled in math class, because I couldn't make a solid connection between some person's name and e.g. a series they formulated. Like Cauchy series.

2024-09-06 21:02:03 UTC

- reply

understood. Didn't know before if you'd share names, or not.

2024-09-06 20:58:11 UTC

- reply

maybe your head mate?

2024-04-11 18:01:00 UTC

@npub18rm…qeha @npub1j9l…jvxt nah, it's just the same, if not worse in the German speaking parts of the Internet.