Troed Sångberg

Demoscene coder. Cybersec consultant. A child of the 80s home computer era.

Depending on who you are I either help secure your stuff against evil hackers or I play one and show you how they work.

I use two accounts. This one is tech/retro/cybersec focused. Random social chitchat on random topics can be found over at Troed Sångberg (npub14zw…pf4e)

HOMEPAGE: https://troed.se
PEERTUBE: https://video.troed.se
MATRIX: Troed Sångberg (npub192k…y8wk):matrix.troed.se
DEMOSCENE: https://sync.wtf

Public Key

npub192kepsf3ypyzgsds9r93amadzta0efs05ftuxhl84m04e9xxhy9qevy8wk

NIP-05 Address

troed_at_swecyb.com@momostr.pink

Profile Code

nprofile1qqsz4tvscycjqjpygxcz3jc7a7k397hu5c86y47rtln6ah6ujnrtjzsprpmhxue69uhhyetvv9ujumt0d4hhxarj9ecxjmntuvr27l

Publishing to

relay.momostr.pink

Author Public Key

npub192kepsf3ypyzgsds9r93amadzta0efs05ftuxhl84m04e9xxhy9qevy8wk

Show more details

Last Notes

2026-05-09 18:37:37 UTC

Insight wanted:
Family member needs new phone. De-googled Android is the requirement. Samsung is a no-go due to their VoLTE/IMS not being supported in AOSP.
Any de-googled Android except GrapheneOS is of interest.
Which phone should they get?

2026-05-04 08:13:26 UTC

If you're still living in February 2026 it's time to realize that things have changed.
https://www.youtube.com/watch?v=VQ0kLuST800
https://media.swecyb.com/media_attachments/files/116/515/296/290/997/333/original/5a5e3b2c71a217ee.png

2026-05-02 11:52:43 UTC

- reply

But this is exactly my point. Knowing how LLMs work and saying we're far from AGI ignores, IMHO, what we know about human intelligence.
We belittle the LLMs since they're "just autocomplete on steroids" while the same describes us very well.
Now, the current LLM architecture with transformers and static training that's then crystallized will not give rise to AGI - that's a given. I'm not as certain about an LLM architecture that does modify the weights during usage however. That would come very close to how we work.

2026-05-02 11:18:10 UTC

Just read another post here on Mastodon where someone commented that LLMs can never become AGI because "they're just spicy autocomplete".
I wonder how many people who make such comments have actually paused and reflected on how they think us humans work.
Are everyone really dualists (believing in metaphysical "fairy dust", spirits, souls etc) or have so few really figured out that we humans are also just pattern matching machines in a loop?

2026-05-01 18:20:30 UTC

- reply

Yeah I just can't see cloud-LLMs continue as they are. They must go to smaller models to get compute down to something people are villing to pay, and at the same time those smaller models will be self-hosted by companies and even lone developers instead. No different to the choice between running your own metal vs VPS.
... and since there will be no magic $5000/month profits from all developers in existence the whole cloud AI bubble will pop.
Loudly.

2026-05-01 17:54:04 UTC

@npub1huw…c5x7 Before: "Anthropic has stolen our works without compensation!"
Now: "Anthropic has stolen our works and then given money to a tool we use and love!"
Can't say I understood the reactions.

2026-04-30 06:19:52 UTC

Well aren't you glad today for all those carefully curated fapolicy allow lists.

2026-04-29 19:45:42 UTC

- reply

Thanks. Turned out it also was Hugo HTML-escaping the apostrophe in the title so I had to write a wrapper around hugo when deploying to fix both issues. Now all good. Thanks again for your posts pushing me to fix my stuff :D

2026-04-29 18:22:54 UTC

- reply

The existence of humans disprove that finding ;) It was much narrower in scope than that.

2026-04-29 18:00:49 UTC

- reply

Hugo generated RSS for my blog here - but since I know some RSS parsers choke on it (News in Nextcloud for example) I'll take this opportunity to ask if anyone can spot what the problem is and help me fix it :D
I _think_ it's due to an apostrophe in the name, but so far I haven't figured out how to solve it.
https://blog.troed.se/rss.xml

2026-04-29 14:44:06 UTC

tl;dr: The whole LLM-based "AI" revolution we're living through right now was started by Google releasing a paper detailing the architecture LLMs now use. That was almost a decade ago.
These models are trained once and then used for various tasks, but they don't acquire new abilities or learn from usage. Their neural networks are frozen.
A possible architecture for AI models that can evolve was presented a few months ago. Again by Google.
Don't base your opinions on what AI can and cannot do on one single architecture.
https://www.youtube.com/watch?v=VTQR9n3aqNU

2026-04-29 07:16:57 UTC

This is why I'm licensing all my projects where I've used LLMs as Creative Commons Zero. It's the only thing I can do to protect myself and others from (possibly) abusing copyright.
"Works predominantly generated by AI without meaningful human authorship are not eligible for copyright protection under current doctrine, and that position is stable even if it is not finally settled."
https://legallayer.substack.com/p/who-owns-the-claude-code-wrote

2026-04-28 19:15:15 UTC

- reply

... and this is the site: https://sync.wtf/
The archives-part was previously fully dynamic, since I thought it would be nice to have the site automatically work with any new disks I archived the second the image had been finalized. The site ran a multitude of cli tools for every request, and while I of course had spent time on sanitizing inputs I wasn't exactly confident I could catch everything.
It also happily produced new zip archives of the full collections every time the dl links were accessed ...
Now everything is statically built. Even the search is done client side on a downloaded json. The AI bots and hackers can go wild :P
I don't think I would've spent the effort doing this rewrite without AI tooling tbh. As it was now I could just point the AI to do the task with the existing site as the goal to achieve - and it worked out just fine.

2026-04-28 19:08:21 UTC

ngl, it feels slightly uneasy to have @npub1xrs…qupw email remind me to not forget to finalize my order when I did so (with confirmation) the first second it was possible a month ago already
The database should know :/

2026-04-28 14:36:05 UTC

RE: https://mastodon.matrix.org/@matrix/116482796650492841
Malmö you say?
Might even get me to travel those few miles away from my home office even!
Now to raise awareness amongst the local politicians and journalists.
#note1sjh…pw2d

2026-04-28 09:33:20 UTC

This was a really interesting video on a game engine concept I didn't know about. I'd love for the Minecraft/Valheim/No Man's Sky etc genre to evolve into fully modifiable detailed worlds.
(Which is the solution to Fermi's paradox btw ;)
https://www.youtube.com/watch?v=il-TXbn5iMA
#GameDev

2026-04-27 15:00:36 UTC

- reply

fwiw Apple was rumored (strongly) to have working smartglasses-prototypes internally _10 years ago_ :) The talk then was that they would be released when technology had advanced far enough for them to "be good looking enough".
(This rumor told to me personally by the person-in-the-know at a private SXSW gathering)

2026-04-27 08:17:23 UTC

https://media.swecyb.com/media_attachments/files/116/475/674/261/291/982/original/7e97bc95368b5552.png

2026-04-26 20:16:54 UTC

- reply

I suggest "Consciousness: An Introduction" by Susan Blackmore. Don't be fooled by the name, it's a whole university semester's worth of science on what we know about how our brains work.
It always help studying a subject before taking on too strong views on it, I feel.
https://www.susanblackmore.uk/consciousness-4-2/
@npub1lcc…lcye

2026-04-26 19:59:43 UTC

- reply

I spent a lot of time studying the known science regarding human counsciousness ~2 decades ago*.
Did you know that you're also just a pattern matching machine?
It changes the perspective a bit regarding how to view LLMs. Just as we show emergent effects that we describe as "conscious", "understanding", "feeling" etc so can those words be used to describe the emergent effects shown by LLMs.
Maybe you don't know as much about the subject as you think you do?
*) https://blog.troed.se/posts/i-am-a-strange-loop/
@npub1lcc…lcye

2026-04-26 19:34:49 UTC

- reply

That's your belief, based on preconceived notions you have.
In contrast, I challenged my initial opinions on how LLMs were useless for development and have admitted to myself that I was wrong. That journey can be traced in public posts I've made here.
@npub1lcc…lcye

2026-04-26 19:04:45 UTC

Here's an example of how extremely rapid the development is right now regarding coding LLMs:
I have a subscription to Mistral AI since I choose to spend my money in Europe. They have a large coding model, Devstral 2, which launched in december 2025. I've used it successfully with a few projects.
They're having capacity problems, and today the model simply refused to do any work. Since today is when I have the time, I instead booted up a local model (Qwen 3.6 35B-A3B) heavily quantized (UD-Q4_K_XL) to be able to somewhat fit on my workstation's Nvidia 5060Ti with 16GB RAM (and using a bunch of system memory too).
The local model, while running a bit slower, does a better job than the cloud model. They differ by about 5 months.
Whatever your views regarding LLMs for development are, they're outdated.

2026-04-26 18:47:54 UTC

- reply

Agree with your take - I posted yesterday on both the simularity to when assembler coders had to admit compilers out-optimized them as well as Java JIT out-optimizing static compilation.
Don't agree with "destroying the planet" though. I'm right now doing work with a fully local model on a regular gaming graphics card. The huge cloud models subsidized with VC money will go away naturally since they'll never be profitable.

2026-04-26 07:28:54 UTC

Great post. I agree that a critical factor in why I'm getting good results from LLM-assisted coding is that _I know this shit_. I flag the model when it's going down the wrong path, and I include hints in my prompt that I know will steer things in the right direction.
If you don't know how to do that, you're getting shit quality output.
"It works today, because the people reading those docs have the engineering expertise to act on them. What happens when they don’t? Honestly, I don’t know. Maybe AI in five years is good enough that it won’t matter. Maybe the problem stays manageable. I can’t predict the capabilities of models in 2031."
#LLM #AI
https://techtrenches.dev/p/the-west-forgot-how-to-make-things

2026-04-25 16:49:28 UTC

In the beginning of the 2000s mobile phone manufacturers started adding Java to their platforms. This meant that game developers could get code onto the phones much more easily, but a lot of developers at said mobile phone manufacturers sneered at "that slow Java stuff".
They were in for a rude awakening a few years later when it turned out that JIT on a modern CPU is able to optimize better than statically compiled C/C++, and they rapidly found themselves not keeping up to date on technological advancements was a bad decision.
This is a post about @npub1xsv…q6he refering to coding LLMs as "glorified autocomplete".
(Yes, I was there when assembler developers laughed at the idea that a C-compiler would ever be able to optimize better than them too)

2026-04-25 14:30:52 UTC

- reply

I'm fully onboard with everything, except I don't like that it gives the impression that not talking about issues is "good".
My variant of “That happened. It sucked. Moving on.” is "Don't spend energy on things you cannot change".
But I also _talk_ about my own issues, to make sure others - regardless if they want to talk about theirs or not - don't feel alone. That there's nothing wrong with them in particular if they suffer from depression, or that they failed with some goals they had set up.
I got that realization a bunch of years ago, when I got a few comments to the tune of "But Troed, you're always so sure of yourself and succesful at what you do!" and I realized that what others need is to hear that we all run into issues and that's just part of life.
/successful burnout-sufferer on SSRIs

2026-04-25 10:11:40 UTC

The whole home network with its separate VLAN and services is now IPv6 enabled!
I know, I know, this should've been done a long time ago (at least as soon as our ISP enabled it) but it's difficult to prioritize when "everything's working already".
At some point down the line I'm sure I'll be happy I did this though. For now I'll just sit and enjoy the fact that I now know more than I did yesterday ...

2026-04-24 19:29:27 UTC

- reply

Well in this case the ISP is not to blame. It's a friend of mine who's running the nameservers on his own ... well ok then, I guess he actually is a small ISP :) But my connection should be fully IPv6.
@npub1huw…c5x7

2026-04-24 19:16:41 UTC

@npub1huw…c5x7 made me feel so ashamed that the services I host aren't reachable over IPv6 earlier this week so I actually sat down and worked on it now.
... just to find out that the nameservers for my domains filter AAAA records over IPv4 and aren't responding on IPv6 themselves.
ffs

2026-04-24 15:33:51 UTC

- reply

Snyggt! Helt rätt tid för det politiska budskapet om varför vi måste bygga på öppen källkod nu när, t.ex, vi behöver ett nytt journalsystem efter att Millennium rasat ihop helt.

2026-04-24 10:23:42 UTC

- reply

It's popular to rant on GenAI (LLMs etc) here but idk. I posted this on the retro computing #atariscne IRC channel earlier this morning:
"I asked Qwen3.6 35B A3B to give an opinion on the current state of the collision detection in Atari Dash and it came back with everything I knew and one thing I didn't - and then produced accurate suggestions on improvements.
everything in pure 68000 ASM"
Atari Dash is a game engine I've written myself in pure assembler, a "port" of Geometry Dash. I'm ... stunned by how well a local LLM running on my GPU can understand such a codebase.
I'd say we've crossed where LLMs went from just being hilarious to actually useful a while back.

2026-04-23 16:52:07 UTC

- reply

Got the suggestion to post an issue on the appstore since I found it unclear whether my dual-license tags (one in the approved list, the other being CC0) should have worked or not.
It became a feature suggestion since no, the Nextcloud AppStore enforces a strict list of approved licenses (why??). However, CC0 has a good chance of being approved so let's see how long this will take ... :)

2026-04-23 09:17:41 UTC

- reply

You're right. This script, that for some reason exists in my $HOME, prompts "meh" and if the user responds with "get" it'll print out "meh", otherwise "mah".
Fascinating piece of code.

2026-04-23 08:02:07 UTC

Cleaning up the local disk. It's an interesting excercise in trying to figure out what some files and directories actually contain, so that I can make an informed decision whether they should be kept around or not.
This one is pretty easy to know what to do with. But I can't figure out why it exists at all.
Unless I'm hacked, by a gigantic troll.
https://media.swecyb.com/media_attachments/files/116/452/959/517/294/222/original/72ab39b865413cb6.png

2026-04-22 19:47:12 UTC

Upgrade to Ubuntu 26.04 (I'm on 25.10 on the main workstation) tomorrow or just disappear from the world for a while and reemerge on OpenSUSE Tumbleweed?
I don't even want to think of how much pain a distro switch will bring just to get back into a productive environment ...

2026-04-22 19:15:28 UTC

- reply

Kanske kan röra på sig något mer nu när det är på hemmaplan dock?
(ang. https://consumerfed.org/press_release/consumer-federation-of-america-sues-meta-for-failing-to-protect-users-from-scam-advertisements/ )

2026-04-22 18:21:54 UTC

- reply

It's another wild hen-type that's possible to tame (savannah biome). I just love how crazy looking they are compared to the "regular" ones. I think they have different movement patterns too because they're the only ones that seem to be able to get up on the roof - and keep doing so.

2026-04-22 18:15:37 UTC

I love that the #Hytale devs understand that when you have chickens there's always one hen that's just a little bit crazier than the others.
https://media.swecyb.com/media_attachments/files/116/449/712/716/510/810/original/4bac618f2fedce55.png
https://media.swecyb.com/media_attachments/files/116/449/714/578/704/693/original/6117389c722a55f6.png

2026-04-22 16:55:13 UTC

welp
Tried to release my #Nextcloud app just now but it turns out I'm using too permissive licenses which isn't allowed :P
https://help.nextcloud.com/t/unable-to-release-application-due-to-too-permissive-licensing/243620

2026-04-22 15:59:14 UTC

- reply

https://www.recordedfuture.com/blog/ai-hype-vs-reality

2026-04-22 14:49:26 UTC

- reply

I read that blog. It completely misses the point.
The thing that's scary is that the bad guys have easier ways of finding and automating attacks using LLMs, faster than maintainers are able to patch them - even if they also get bug reports. Mythos is an example of a large LLM capable of doing so - that smaller models can do some of it rather strengthens the worry, not debunks it or whatever it is flyingpenguing is trying to do there.
See https://daniel.haxx.se/blog/2026/04/22/high-quality-chaos/

2026-04-22 07:03:40 UTC

A factory is at a stand still. No one knows why, but the production line isn't moving along as it should. The foreman calls in an external repair service, and a clearly senior and wethered person shows up with a toolbox in hand. They walk along the length of the production system, humming along interrupted every now and then with "interesting!".
After a while, the person takes out a small screwdriver and proceeds to slide in underneath one of the machines. After a few seconds they reappear, and now all lights are suddenly green and the factory can start back up.
When the bill arrives the foreman is livid - it's one line and simply states "Repair: $10000". They call up the repair service and proclaim that they were there when the repair was done and there's no way screwing around for a few seconds could cost that much money!
The repair service apologies and offers to send a new bill properly itemized. When it arrives it now has two lines:
"Screwdriver application: $1"
"Knowing what to apply the screwdriver to: $9999"
This is a post about LLM assisted software development.

2026-04-21 20:13:22 UTC

- reply

I just want to chime and say that you're obviously doing the right thing by adding storage quotas! And wow they're generous indeed at that.
Thank you for your service. It's very appreciated.

2026-04-21 17:52:12 UTC

- reply

It has never become the top priority on the never ending list of "things I really should do".

2026-04-21 17:31:02 UTC

- reply

Det blir alltid rubriker om saker som är nya. Självklart smygfilmas och smygljudinspelas det otroligt mycket med mobiler. De rubrikerna ser vi inte längre eftersom det inte är nytt.
https://www.washingtonpost.com/news/the-intersect/wp/2014/12/30/some-creep-is-prowling-the-d-c-metro-filming-women-for-a-porn-site-and-theres-nothing-you-can-do-about-it/

2026-04-21 16:29:46 UTC

- reply

Jag arbetade med forskning kring smarta glasögon på Sony för många år sedan (Sony SmartEyeglass). Vårt resultat var att det är lättare att filma någon obemärkt med mobilen som hålls i handen än med glasögon eftersom med de senare blir det rätt tydligt att du stirrar.
Det har alltid varit min uppfattning, sedan Google Glass-incidenten, att det handlar om ren reaktionär teknikfientlighet mot "allt sånt som inte fanns när jag växte upp"* snarare än någon faktisk skillnad i möjligt missbruk.
*) https://www.goodreads.com/quotes/39828-i-ve-come-up-with-a-set-of-rules-that-describe

2026-04-21 14:14:25 UTC

- reply

Allt som står där gäller även helt vanliga mobiler.

2026-04-21 12:48:27 UTC

- reply

Yeah well. TP-Link Deco M9 Plus devices sold right now in the stores are running Linux 3.14 kernels.

2026-04-21 11:03:40 UTC

- reply

Klientdelen. Finns förklarat en bit ner här: https://github.com/jameskdev/android_samsung_imsservice

2026-04-21 07:59:11 UTC

I messed up
That LineageOS Samsung S20FE I got running was intended for one of the kids, as their first phone.
Now I know that Samsung has their own IMS that isn't compatible with AOSP and thus you cannot do voice calls in countries where VoLTE is the only option available.
Like Sweden.
Sure, this information is "known" if you're into degoogled Android phones - but I wasn't. The only way to get voice working on this hardware is to restore it back to Samsung's - no longer security updated - Android.
ffs

2026-04-20 17:01:14 UTC

- reply

... and while that caused the whole of Bluesky to go down, the rest of Mastodon doesn't even notice this.
... as designed.

2026-04-20 12:47:55 UTC

- reply

Thanks! Will have a look.
Right now I'm knee deep in debugging what I think are errors outside of my app :/ Won't be able to release if I can't get that sorted.
https://help.nextcloud.com/t/unable-to-persist-exapp-settings/243474

2026-04-19 20:55:26 UTC

I love this Youtube channel.
Also, this is a very good update on the state of Risc-V in the market.
https://www.youtube.com/watch?v=z6gHC-R59lw

2026-04-19 20:08:40 UTC

- reply

True - but local LLMs are lagging only by about ~1 year in functionality. It's very possible to LLM-code using a local model like Gemma 4 on "any regular gaming workstation". You won't get the same hands-off results as with the latest cloud models, but it's very doable if you put in effort.
Half a year ago, the results were quite poor. This is moving at breakneck speed atm.
@npub13t2…gh8m

2026-04-19 19:59:19 UTC

- reply

I think once every two years would come close to a good average.
One of the very real benefits of using an LLM is that you get to ~80% functionality and across the initial "oh do I really have time to ..." friction with just a few prompts. Then it's just a question of how much effort you'd like to spend on the quality.
You might be interested in following Harry McCracken as well as he now vibes his own writing tools. While I'm an actual professional software dev, he isn't, but still gets what he needs from them:
https://mastodon.social/@harrymccracken/116432577142940009
#note104u…9upf

2026-04-19 19:51:08 UTC

RE: https://infosec.exchange/@jerry/116433023149652219
I have four followers on mastodon.cloud from what I can see. It seems you'll need to act on this advice - sorry that you ended up on an ill-maintained instance :/
#note13qz…2hw5

2026-04-19 19:36:45 UTC

- reply

I cited your exact words. Yes, you claim that you know that the official known death toll (the one we know) is "very wrong". When asked to back up your claim you came back with debunked guesswork.
I don't think facts is something you feel is important when discussing this subject.

2026-04-19 19:32:48 UTC

- reply

I thought you claimed differently - and additionally tried to claim that I were one of the "industry supporters" that "downplayed deaths".
"Only 31 people died from Chernobyl! (very wrong)"
Fact is, nuclear power is the safest power generating technology we have - tied with solar. I'm all for solar and batteries - but I'm also a proponent for rebuilding a nuclear power plant right in my own backyard since it was a mistake to close it down a few decades ago.
https://ourworldindata.org/grapher/death-rates-from-energy-production-per-twh

2026-04-19 19:11:06 UTC

- reply

Just because they weren't directly attrubuted doesn't mean they happened either.
Your link was about what _could_ happen, in the future, and was 20 years old. Now we know that the projected increases in cancer didn't happen.
"There have been no validated increases in solid cancer reported from the liquidator cohorts, and observed increases in leukemia have been statistically insignificant. The liquidators were adult at exposure and the average external dose was 117 mSv.
A paper in Science has stated that there have been no transgenerational effects of radiation exposure in children born of those working as liquidators. This study used whole genome sequencing in a cohort of parent and child blood samples"
... and apparently that report in itself was highly biased:
"However—as an April 2006 special report in the peer-reviewed, scientific journal Nature detailed in response—the accuracy and precision of this United Nations-led joint group's projected death toll of 4,000 were immediately contested, with several of the very scientists, physicians, and biomedical consortia whose work the joint group had cited alleging publicly that the joint group had either misrepresented their work or interpreted it out of context."
https://en.wikipedia.org/wiki/Deaths_due_to_the_Chernobyl_disaster

2026-04-19 19:01:19 UTC

- reply

"Only 31 people died from Chernobyl! (very wrong)"
So, let's hear your claims so that we can debunk them scientifically then. How "very wrong" are the 31 Chernobyl deaths and on what data do you base your claim?

2026-04-19 18:57:41 UTC

- reply

"presumably these are the first ones you've released."
I've been releasing open source software since 1989.
If you had read the blog post regarding how I went from LLM-coding skeptic to finding them useful you would already know about one of the programs.
Two others are Hytale mods which are used by other server admins with very nice feedback.
The fourth is a Nextcloud app I've just registered on the appstore and will publish when I get issues with their Python lib and the configuration system sorted - it's running only on my own instance so far.
I don't think you're looking for metrics. I think you have your opinion all made up and no one will be able to convince you otherwise.
@npub1hr7…jxyc

2026-04-19 18:48:45 UTC

- reply

I'm unable to talk about what my current cybersec assignment entails.
My experiences used in this discussion come from the four open source programs I've written the last month which I know had been zero had I not used LLMs to a high degree during their development.
Feel free to cite that metric. I'm quite confident you're not participating seriously in this discussion anyway.
@npub1hr7…jxyc

2026-04-19 18:21:41 UTC

- reply

You seem to be claiming there are lots of deaths attributed to nuclear power accidents.
There are not. That's not "minimizing them", that's just stating a scientific fact.

2026-04-19 15:49:05 UTC

- reply

I don't think waste from atomic bomb programs is what people usually have in mind when talking about the safety of nuclear power.
(I'm European, we have plenty of nuclear power here all without any such issues indeed)
@npub1p9a…r87r

2026-04-19 15:46:57 UTC

- reply

While the japanese government agreed to consider the worker's lung cancer, four year afterwards, as having been caused by his heroic efforts so that his family would get benefits, there's very real doubt as to whether it actually was. However, there's no point in arguing over that one death since it doesn't matter at all for the question of whether harmful radiation was released or not. His situation wasn't exactly the norm.
https://en.wikipedia.org/wiki/Fukushima_nuclear_accident#Fatalities

2026-04-19 12:31:13 UTC

- reply

Notice the "~six months ago" in my original post. The paper is much older. There has been a huge change to how good the coding models are since then which has made them go from (in my view) useless to useful.
You'll see many others make the same comments everywhere, if you care to challenge your own convictions.
@npub1hr7…jxyc

2026-04-19 12:19:33 UTC

- reply

Why do you question my statement to begin with? I'm 51 yo, started coding when I was twelve, graduated as a Software Engineer in 1998.
I sort of consider my own evaluation to be good enough to stand behind.
How much time have you yourself put into the subject before coming to a conclusion?
Here's a writeup on what converted me: https://blog.troed.se/posts/vibe_coding_results/
@npub1hr7…jxyc

2026-04-19 11:49:56 UTC

- reply

We passed the "maybe they're useful, maybe they're not" when it comes to development, reverse engineering, code analysis etc about -6 months ago.
I'm a converted skeptic. An LLM is a very useful tool in a developer's toolbox.
Posing a question to a codebase and getting pretty much immediate insight is truly time saving, and the "chatbot diseases" (hallucinations etc) are pretty much nowhere to be found. For stuff that has to do with code.

2026-04-19 08:49:48 UTC

- reply

I mean the reason it gets repeated is because it's true. You're the one posting trivially debunked myths.
(One worked - maybe - died from radiation caused lung cancer due to Fukushima)

2026-04-18 21:02:30 UTC

- reply

Yeah. Maybe iterate on the second part?
"Indeed the doors aren't open."

2026-04-18 14:39:53 UTC

- reply

The King's visit to Ukraine, otherwise yes :)

2026-04-18 12:14:04 UTC

- reply

I guess "Moving the goalposts" + "bye" means that most of what you post is disconnected from reality then.
https://www.forbes.com/sites/iainmartin/2026/04/16/how-frances-mistral-built-a-14-billion-ai-empire-by-not-being-american/

2026-04-18 11:54:31 UTC

- reply

I understood the ask I replied to was regarding ethical training. Mistral, as an EU company, has to abide by EU regulations AI companies in the US, China etc don't have to.
https://artificialintelligenceact.eu/article/53/
[@chris_evelyn](https://fedi.chris-evelyn.de/@chris_evelyn ) @npub1jpf…vqhj

2026-04-18 11:35:55 UTC

- reply

What an interesting claim. Has it got anything to do with reality?
@npub1jpf…vqhj

2026-04-18 11:20:30 UTC

- reply

It's my belief that Mistral's models fit that bill.
@npub1jpf…vqhj

2026-04-18 08:47:16 UTC

- reply

I'm on the fence. Part of me feels like I'm avoiding responsibility if I document what's "my" contributions and what's "AI". All of the code is my responsibility.
At least if the idea is for this to be externally visible. I could see a point of keeping track of it purely for my own saker.

2026-04-17 19:30:43 UTC

- reply

You are doing the Lord's work.

2026-04-17 19:29:06 UTC

- reply

ok, I've bit the bullet. PR asking for a signing certificate made and thus my WIP source is now public on Codeberg too.
I've had a horribly patched together shellscript-thing running on my own Nextcloud instance on and off before, but converting it to an actual ExApp has done wonders. It's right now creating EXIF, Nextcloud tags and ALT text for _all_ images & photos (many tens of thousands) in the family's photo storage.
That server has an Nvidia A2000 GPU. 12GB of VRAM means it can run a competent vision model (unsloth/Qwen3-VL-8B-Instruct-GGUF:Q4_K_XL) and churn through the images somewhat rapidly, whilst not using more than ~65W of power doing so.
The end result is, IMHO, absolutely awesome. Let's see if I can manage to clean it up, conform to guidelines and make it public ...
https://codeberg.org/troed/tagger

2026-04-17 16:24:03 UTC

Uh.
So that #Nextcloud app I've written is chugging along nicely now on my own instance. Time to look through how to make a proper release so that others can run it too right?
Well. That release process looks like it needs a full time job all in itself to get through. Seems easier to describe to people how to do manual deployment of a docker container just like I've done for dev purposes.
(Yes I used to work at Developer Relations at Large Global Company. Don't create too much friction if you want third parties to make stuff for your platform ...)

2026-04-16 06:27:00 UTC

TIL
replacing Ollama where it's still used now
https://sleepingrobots.com/dreams/stop-using-ollama/

2026-04-15 19:11:31 UTC

I'm not really keen harping on open source projects, we all know how hard it is to get funding to properly maintain them, but I really need to vent a bit regarding trying to write a #Nextcloud ExApp.
It's an exercise in "while floof was the way meeps were done before the flaff API v2 needs woshies that are done differently due to smappapp. The documentation still say nuffnuff but as can be seen in the commit history to shwabadamp it should now be done this way instead".
well ok then
My nc_py_api wrapped AppAPI using ExApp still just gets an empty list when trying to read preferences and that ui dropdown menu item I've supposedly registered is nowhere to be found.
pain

2026-04-15 17:44:10 UTC

- reply

Hope all goes well - especially the learning french part! That's a language the rest of us Europeans try to forget exists :D

2026-04-14 06:37:55 UTC

- reply

RE: https://fosstodon.org/@balloob/116398481380578311
The only reason I have Chromium installed is for webserial. I haven't had the need to use anything but Firefox for years otherwise.
Awesome.
#note1dmy…m5au

2026-04-13 15:11:22 UTC

I'm so giddy atm. This week I'll get to teach redteaming to an eager young intern - on an actual product.
I'm also not going to pre-investigate, just let it all happen as we figure stuff out together.
They're already a really good developer. I'm hoping that this might leave a lasting interest in hands-on cybersec on top of that.

2026-04-13 14:42:50 UTC

- reply

Visible faults are the best faults!

2026-04-13 14:33:09 UTC

- reply

Copying a small snippet i copying a small snippet. Dividing up something large into many small snippets and then copying them all means you copied something large.
Not the same thing.
"Certain things cannot be copyrighted because they are trivial to implement. This however does not change the fact that one has to respect the copyright of code that one is copying"
No, you don't have to respect non-existing copyright.
@npub1wxe…dz62 @npub1aj4…cdjs

2026-04-13 14:19:55 UTC

- reply

The actual problem is that a looping ausearch with checkpoints will _miss_ some audit events. It looks like when it happens it has also decided to err 12 and restarting as suggested in the man page with --start checkpoint is not enough to catch them.
RHEL8 fwiw

2026-04-13 14:11:27 UTC

- reply

"How one can argue that this is not a copyright violation is baffling."
A background as a Pirate Party "politician" with immense deep dives into "Intellectual Property" protections and legal frameworks will do that to you.
A lot of people belive just because they "did something" that that something qualifies for copyright. That is most definitely not true.
@npub1wxe…dz62 @npub1aj4…cdjs

2026-04-13 14:07:47 UTC

Why is ausearch broken? This can't just be me. Trying to ausearch using --checkpoint will "randomly" return err 12 even though the audit file is just chugging along in the background with nothing affecting it whatsoever (no rotation, no other process writing in it etc).
I don't want to re-trigger on old things which is my fear when restarting with --start checkpoint
#Linux #Blueteam

2026-04-13 13:50:25 UTC

- reply

No, that was not what I claimed. Let's have a serious discussion instead?
@npub1wxe…dz62 @npub1aj4…cdjs

2026-04-13 13:35:29 UTC

- reply

You do realize that "that would constitute copyright infringement" was part of my original question - which you cited?
"Copying lines of code that are small enough does not eliminate copyright"
Snippets that don't qualify for copyright don't magically do so just they're part of something larger either.
When there's only one way to do a task code that does that task isn't copyrightable.
(For those a bit older this was part of the Oracle-Google Java API fight)
https://www.eff.org/deeplinks/2021/04/victory-fair-use-supreme-court-reverses-federal-circuit-oracle-v-google
@npub1wxe…dz62 @npub1aj4…cdjs

2026-04-13 13:31:00 UTC

- reply

Yeah that is a great paper but it assumes the conclusion. Instead of repeating the same thing in two places in the same original thread I'll just link my other reply here:
https://swecyb.com/@troed/116397595152157307
#note1pn4…y8gz

2026-04-13 13:19:11 UTC

- reply

The problem there isn't the reproduced snippets of code but the lawyer stating that it's "clearly copyright infringement".
What I'm getting at is whether a functional snippet of code qualifies for copyright at all - the threshold of originality: https://en.wikipedia.org/wiki/Threshold_of_originality
... which is why I can guarantee that my own LLM-assisted code isn't license encumbered. Nothing large enough is being assisted anywhere, the contributions from the LLM are trivial.
@npub1wxe…dz62 @npub1aj4…cdjs

2026-04-13 12:19:34 UTC

- reply

This is what you wrote:
'Would you be willing to guarantee, legally binding, with all consequences, that your "AI" "assistant" didn't copy-paste code that's under an incompatible license? Or even proprietary, stolen one?'
Yes, I have no problems guaranteeing that. Why do you believe that would be a problem?

2026-04-13 12:18:27 UTC

- reply

So you're claiming something without being able to back it up with ... anything?
@npub1aj4…cdjs

2026-04-13 11:59:32 UTC

- reply

Can you point to a case where a coding assistant has done so that would constitute copyright infringement?
@npub1aj4…cdjs

2026-04-13 11:58:52 UTC

- reply

I don't understand your reply. My comment was about your assumption that they perform copying of training data into the output that would constitute copyright infringement.

2026-04-13 11:01:27 UTC

- reply

Why do you think coding aids output copies of training data that would fall under copyright protection?

2026-04-12 16:01:06 UTC

- reply

fina-effing-lly
Full workflow from lazyvim on workstation building docker image of the app, pushing to remote registry where it's then pulled and deployed on my nextcloud instance where the app manages to register and run.
There's a huge lesson to be had in not using Nextcloud's "simple skeleton python app" example and instead looking up the examples at https://github.com/cloud-py-api/nc_py_api/
#Nextcloud
now what was it I wanted to do??

2026-04-12 14:07:47 UTC

- reply

ffs even trying to get my own dev app up in Nextcloud is proving to be insanely difficult/complicated or something. The examples/guides seem to all have been written from a theoretical perspective instead of from what it actually would look like when someone's out to do it.
Now juggling between a local docker install daemon, which I'm not certain can work in a rootless environment, and the manual install daemon. But hey let's have a local unix socket somehow fighting listening to a port on a docker service binding to localhost so not discoverable in the docker network ...
I'll get there, but I would've liked spending time on coding rather than figuring out all this.