Since downgrading wasn't possible either, I made my first attempt at using Claude Code to write software that would extract all my Nexus packages and restore them onto the original open source version.
https://battlepenguin.com/tech/fixing-yum-indexing-on-sonatype-nexus-and-the-non-opensource-upgrade/
@djsumdog @Suiseiseki He's going to scold you for using the term "open-source" instead of "free software".
I doubt "Sonatype" was ever actually free software, considering that the Eclipse Public License version 2 writes of imaginary property and is quite weak copyleft that only requires source code (but with weak terms that potentially could allow providing partial source code?).
"open source" isn't a useful term, as there is plenty of proprietary software that qualifies as "open source" and all free software happens to qualify for "open source" (it's clearly a grave insult to lump free software with proprietary software.
"closed source" isn't a useful term either, as there are many free software projects that aren't publicly developed.
To get the freedom straight, one should write in terms of free software and proprietary software.
>and absolutely useless AI-generated answers.
Such answers were not generated by artificial intelligence, those were combined via LLMs.
>closed proprietary commercial software
I don't understand why you wouldn't just write; "proprietary software".
Commercial software isn't the problem - proprietary software is.
>The entire way Sonatype pushed the update is very shady. People in corporate environments were suddenly struck with this change, pushing them towards an expensive pro license they might not be able to aff
The problem is not the expense - the problem is the lack of freedom, even after paying huge sums.
>I've started to experiment with newer models.
Cringe and proprietary pilled.
Do NOT use LLMs to copy software - those remove the copyright information and make the resulting software completely proprietary.
If you want to copy free software, copy it directly, keeping the copyright notices intact and making sure the licenses are compatible.
Maybe you should make a little web game, a quiz with some tricky questions to teach people all the terminology: Free (as in beer), Open Source, Free and Open Source, Open Source (but not Free as in Speech), etc. Something fun and funny like the JsDate quiz: https://jsdate.wtf/ (I hope you can run that. The code is MIT: samwho/jsdate.wtf is the github repo).
Would seriously be helpful.
Cringe and proprietary pilled.
Do you work in the software industry? I have a feeling you don't because you literally can't if you hold to your free software extremist views. For all us normies out there who make daily compromises (and I feel like I make fewer compromises than most, the LLM stuff isn't going away.
I was inherently skeptical of a lot of it too, because most of the weighted random token guessing machines are mostly wrong or annoying. The fact is, it has gotten better though. Early stuff could only regurgitate solved problems. When trying to develop an RTSP server in Kotlin/Java, I remember one of these tools, after several prompts/iterations, simply give me a skeleton and "Add your RTSP implementation here." It could not do anything novel. At my last job before layoffs, we had a CoPilot license and the code it generated wasn't great and had a lot of issues.
Recently using the IntelliJ $10 plan, I was trying to adapt my MediaHug app (AGPLv3) to work on Wayland. To do that, I needed to use the existing python-mpv library to render to OpenGL in Qt6. There were examples out there, but only for QQuick and my application uses traditional QWidgets. There were literally no examples I could find anywhere, and I looked at a LOT of QtOpenGLWidget examples. My forum question went unanswered. The generated example I got from the IntelliJ chat wasn't perfect, but it did work and I eventually massaged it into this implementation:
https://gitlab.com/djsumdog/mediahug/-/blob/master/mediahug/gui/mpv/gl_player.py?ref_type=heads
This is a novel implementation that doesn't exist in any off the official python-mpv documentation or any other examples I could find.
Do NOT use LLMs to copy software
The trouble is, that's probably not what LLMs do. A lot of people quote Carmack's fast inverse square root example of the early CoPilot while it was in beta (where it duplicated his code including comments). I suspect at this time Microsoft was probably still using a lot of Vector Search combined with RAG. I don't think a lot of model tools like Claude/GPT4 do that anymore for liability reasons (and all the enterprise licenses contain clauses where they'll accept liability for lawsuits. I really want to see one of these go to court). The models themselves cannot store actual code. They're massive weight/parameter mappings. That's why so much of it seems "made up." But at the same time, we can't really know what the big commercial models are doing behind the scenes. I doubt they're straight copying from RAG or Vector search anymore though.
The trouble in industry is we're facing massive layoffs. The AI stuff has massive speedup advantages, and it's getting weirdly better in some respects. I know several people who've told me they wouldn't hire any senior engineers who refuse to use it. There's no really turning back the clock on this stuff now, if you want to get/stay employed.
I refuse to run that, as arbitrary JavaScript execution is always proprietary software no matter the license, as I don't have control over the software.
>Maybe you should make a little web game, a quiz with some tricky questions to teach people all the terminology: Free (as in beer), Open Source, Free and Open Source, Open Source (but not Free as in Speech), etc.
I'd rather compose a total proprietary death presentation really.
>Do you work in the software industry?
No - I don't write proprietary malware.
>The generated example I got from the IntelliJ chat wasn't perfect, but it did work and I eventually massaged it into this implementation:
You won the gamble this time, but looks like slop.
>links to gitlab.
># License: AGPLv3
Is it AGPLv3-only, or AGPLv3-or-later?
(Although if a substantial amount was copied via a LLM, you are clearly not the copyright holder).
>The trouble is, that's probably not what LLMs do.
So tell me, where do LLM's get the software from? It doesn't come out of thin air does it?
>I suspect at this time Microsoft was probably still using a lot of Vector Search combined with RAG.
microsoft appears to have adjusted co-pilot to check the outputs against the entirety of github and discard outputs that were direct copies - combinations (which LLMs do) seem to get past such filter.
>The models themselves cannot store actual code. They're massive weight/parameter mappings.
Such weight/parameter mappings appears to work out as an undocumented form of lossy compression where the code is stored.
>I know several people who've told me they wouldn't hire any senior engineers who refuse to use it. There's no really turning back the clock on this stuff now, if you want to get/stay employed.
The NEET will inherent the Earth.
If you want to churn out slop fast, yes a LLM will be great - but good luck if you want reliable software that isn't full of bugs (for example, I've seen an LLM output that instructed to check for overflow with an if check (which gcc will optimize out), rather than the GCC overflow builtin (the actual fix)).
You won the gamble this time, but looks like slop.
The only generated parts where the OpenGL/MpvRender context handling and I've heavily modified it/rewritten it. So that's my code you're calling slop asshole.
links to gitlab.
Where do you hose your free and non-proprietary code? You have anything up and published anywhere? or do you just complain a lot without writing any software?
Such weight/parameter mappings appears to work out as an undocumented form of lossy compression where the code is stored.
There's a whole field of research dedicated to how information or facts are stored in neural networks. "lossy compression" isn't a good analogy. Token relations are stored in a very high dimensional space. 3Blue1Brown does some great explanations of this. Not sure if using yt-dlp goes against your ethos:
I refuse to run that, as arbitrary JavaScript execution is always proprietary software no matter the license, as I don't have control over the software.
Do you drive a car? Was it made after the 1970s? Does it have an ECM? Did you rewrite the software on the ECM? Maybe you only take trains. Is the software for the ticketing system or for the signal controls non-proprietary? Maybe you just bicycle. Did you make sure the CAD software used by all the component designers was non-proprietary? Have you ever flown on an airplane?
At this point, I can't see you as anything other than a troll. There's no way you could be this way in real life. You're obviously using the Internet and not a hermit in the woods. You're literally using dozens if not hundreds or proprietary routers and hardware to get and receive messages on the fediverse instance you use.
@djsumdog @Suiseiseki
>You're literally using dozens if not hundreds or proprietary routers and hardware to get and receive messages on the fediverse instance you use.
Stallman has openly said that networked services don't count as software and that nonfree software on routers, servers, and switches only affects the people running said devices.
https://www.gnu.org/philosophy/network-services-arent-free-or-nonfree.html
