Nostr notes by datenwolf

But, as far as my line of work goes, because LLMs and coding ...

2026-06-12T16:36:18Z

In reply to nevent1q…92qn
_________________________

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

Regarding security work: I actually foresaw the enormous ...

2026-06-12T16:30:36Z

In reply to nevent1q…tj9a
_________________________

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

Regarding navigating codebases: Personally I take great enjoyment ...

2026-06-12T16:20:11Z

In reply to nevent1q…x746
_________________________

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/

Out of curiosity (I mean this honestly!): What benefits do you ...

2026-06-12T16:05:26Z

In reply to nevent1q…6mch
_________________________

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).

At the end of the 30 minutes time limit the state polling ...

2026-06-12T14:04:10Z

In reply to nevent1q…hq77
_________________________

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

Most recently they urged me to try them out to add support for a ...

2026-06-12T14:00:55Z

In reply to nevent1q…298f
_________________________

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/

I spend most of my time earning money with developing software ...

2026-06-12T13:57:13Z

In reply to nevent1q…kpcr
_________________________

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/

I'd like to encourage you to read up on the Gell-Mann-Amnesia ...

2026-06-12T13:37:59Z

In reply to nevent1q…quvp
_________________________

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.

Here's some ammunition to shoot it down: ...

2026-06-08T17:43:29Z

In reply to nevent1q…ufw9
_________________________

Here's some ammunition to shoot it down:

https://www.nature.com/articles/s41593-025-02132-9

If you're putting lots of text on your slides, you're ...

2026-06-04T09:31:11Z

In reply to nevent1q…zcch
_________________________

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:

Apparently this a.out performs simulation of Mott scattering. ...

2026-06-02T21:59:09Z

In reply to nevent1q…jd7j
_________________________

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.

I just did a `find -name a.out` on my homedir, and apart from the ...

2026-06-02T21:35:24Z

In reply to nevent1q…0fjg
_________________________

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.

No. It doesn't take external input to be inspired. It's a ...

2026-05-27T19:33:51Z

In reply to nevent1q…34p5
_________________________

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).

No. The history of art is full of cases where a daring artist ...

2026-05-27T19:23:41Z

In reply to nevent1q…4zqc
_________________________

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.

In visual arts, prompting at best can be compared to writing a ...

2026-05-27T19:08:52Z

In reply to nevent1q…mu4n
_________________________

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.

"The new instruments" / synthesizers argument is ...

2026-05-27T18:46:48Z

In reply to nevent1q…lp8w
_________________________

"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.

This! Also worth pointing out: LLMs are essentially "love ...

2026-05-16T10:58:56Z

In reply to nevent1q…j7ud
_________________________

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.

Nostr event nevent1qqs2gxjudqfas274hnzegc9xrw0fqe2kpjeq74pkwryucl9rqrayarczyrwmvtredv00ux9sfwuyrgfdhg4294j53x3len8y55rm8nncwama5pglpey

2026-04-01T13:01:34Z

In reply to nevent1q…p7ut
_________________________

Darf ich vorschlagen, dass Du alle Zwischenschritte überspringst ...

2026-04-01T12:57:07Z

In reply to nevent1q…n69r
_________________________

Darf ich vorschlagen, dass Du alle Zwischenschritte überspringst und direkt zum Acetylenschneidbrenner greifst?

Als Flüssigkeit kann es definitiv nicht mehr festgrefressen sein ;-)

And once again the Zig language project shows, that there's a ...

2026-03-21T11:13:13Z

In reply to nevent1q…jzg3
_________________________

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

If you were in need of DDR5, I have two kits of 2x 32GiB = 64 GiB ...

2026-01-29T16:45:59Z

In reply to nevent1q…z5yt
_________________________

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.

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

2025-11-26T16:06:24Z

In reply to nevent1q…n7uk
_________________________

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

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

2025-10-26T21:35:29Z

In reply to nevent1q…38m6
_________________________

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

1. Yes 2. No 3. No - specifically those "features" are ...

2025-10-23T17:31:01Z

In reply to nevent1q…cyah
_________________________

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.

Frage an die #Solar & #Akku Bubble: Ich würde mir gerne für ...

2025-09-12T08:19:50Z

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?

#Cursed #programming idea: When autogenerating filenames escape ...

2025-07-23T10:02:05Z

#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:
/ → ／
\ → ＼
# → ＃
! → ！
" → ＂
' → ＇
( → （
) → ）
[ → ［
] → ］
{ → ｛
} → ｝
| → ｜
* → ＊
+ → ＋
, → ，
. → ．
- → －
: → ：
; → ；
< → ＜
> → ＞
= → ＝
? → ？

> You don’t have to be a pilot to design an aeroplane ...

2025-07-09T14:01:48Z

In reply to nevent1q…4grg
_________________________

> 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.

Well that much I understood. Which is why I want to read it, so ...

2025-06-28T17:46:48Z

In reply to nevent1q…x4s2
_________________________

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.

May I ask you to reconsider, not sharing this study? The people ...

2025-06-28T17:37:09Z

In reply to nevent1q…f5kq
_________________________

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.

How a Tesla autopilot computer vision system interprets the ...

2025-01-15T21:46:01Z

How a Tesla autopilot computer vision system interprets the visual input when confronted with a train. 😆

Not Just Bikes 🇳🇱 (npub1z7x…q5jf)

via: https://labyrinth.zone/notice/Aq5V74iVZRc1GyRCm8

(EDIT: rephrased for clarity)

Dieses Schlupfloch ist IIRC in der Petition (habe ich vor ...

2024-12-27T09:51:22Z

In reply to nevent1q…yd6p
_________________________

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

Jupiter has the strongest, largest magnetosphere of all the ...

2024-10-15T16:42:59Z

In reply to nevent1q…qn2e
_________________________

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.

So what will it be? Panic buying a whole bunch of essential ICs, ...

2024-10-01T08:09:49Z

In reply to nevent1q…gslk
_________________________

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.

weird cognitive biases, maybe. Maybe because I'm terrible ...

2024-09-06T21:41:45Z

In reply to nevent1q…ghmw
_________________________

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.

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

2024-09-06T21:02:03Z

In reply to nevent1q…u0ur
_________________________

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

maybe your head mate?

2024-09-06T20:58:11Z

In reply to nevent1q…xl7n
_________________________

maybe your head mate?

@npub18rm…qeha @npub1j9l…jvxt nah, it's just the same, if ...

2024-04-11T18:01:00Z

Richard "mtfnpy" Harman (he/him) (npub18rm…qeha) Technology Connections (npub1j9l…jvxt) nah, it's just the same, if not worse in the German speaking parts of the Internet.