saw someone on HN say they will never contribute to Nura again since we ban LLM generated contributions (they had receipts too, they had contributed in the past)

their argument was that having done device porting manually before, doing it with an LLM was something like 10x faster

having seen the various LLM driven ports going around, we can't deny that LLMs are capable of doing kernel bringup, finding a functional implementation, and making many more attempts to trial and error than we humans are.

of course, this doesn't come without a cost, as with most technical advancements this cost is mostly offloaded onto those without the resources to prevent it -- the under educated working class and the global south, those who will be most affected by droughts and worsening climate change

this is reason enough to disengage from this technology, at least the implementations of it that cause most of this harm, but there is another more devious issue that i haven't seen mentioned a whole lot

in the same way that huge retail chains and global shipping extract wealth from small towns, LLMs extract knowledge from communities

the Nura wiki and other infrastructure was unusable for days at a time due to web scraping, and now imposes a proof of work on visitors (which also prevents access from old devices)

now that this data is in accessible models, people who aren't active in our community can use an LLM to port a device, to generate their own private documentation, they can enable devices in the mainline kernel or contribute to other upstreams if their LLM patches are accepted. People that may normally have become active members of the Nura community instead engage with a chatbot operated by big tech, which of course uses their chat history for training new models

As maintainers, for us to allow LLMs into our community would in my opinion be the beginning of the end, as our community knowledge base would be gutted, developers would become burnt out and disillusioned, and our culture would be erased.

as far as prescriptions go: i think we should continue to be vigilant about protecting FOSS communities, we should keep being educated about how this technology works and what it's capable of.

We should advocate for a whole lot of sociological research into this, both LLMs generally as well as their impact on FOSS communities

i don't doubt there is a possibility - maybe even a future - where local models can be useful tools without burning the planet or breaking your brain, heck i'd love to optimise some of my day to day work. But i'll be damned if i lose the curiosity and drive that makes me spend days reading codebases and experimenting to find a bug

@cas This. So. Much.

Even if there's a work that's drudgery, I don't want to hand it over to an LLM. Because it rids me of the possibility to learn and gain knowledge. Of course, drudgery is still something I'd like to avoid, but in my experience, doing a deep dive into this annoying task often allows me to build a better understanding of the problem space and thus allow me to optimize and get rid of the drudgery in a neat and insightful way, that's possibly even reusable or has impact beyond the optimization.

An LLM rids me of this ability. Not only monopolizes knowledge, but it's not an optimization in the way the task is solved but rather a brute force that offloads tasks without understanding. I find this depressing.

There are tasks that can never be optimized through this insight and remain drudgery, perhaps. But in my experience these tasks often only exist because of certain "business requirements", effectively "bullshit jobs". An LLM could help here, but I won't use one for the other reasons you stated.

My whole drive is about understanding, not about optimizing for some abstract goal. Humanity is old and there's no race, there's no point in "gotta go fast" (except when tackling sth like climate crisis, but here, going slow might be better even so). I don't care about "efficient" workflows per se, I care for the careful process of understanding and the resulting optimization. I don't care for outputs but the work done. LLMs take that away from me. I don't want a world where I can "just generate", it honestly makes me sick.

@cas wrote about ethics here (need to update): https://logicalerzor.codeberg.page/blog/ai-kernel-development/

re climate change: its muddy rn. ive seen https://youtu.be/DNkZRt-X4RM and otoh ive seen ur position. hard to see the data given people’s incentives

re culture: cultures change. noobs today over engineer where pros know where to cut

ethical AIs in the works:
https://fosstodon.org/@Logical_Error/117303253315458075
maybe more

is it better to:
- prevent ewaste and help people switch to Nura by using ai
- not use ai and delay reducing ewaste
🤷

@Logical_Error in practise i don't think the ewaste argument holds much weight today, it's completely theoretical and i haven't seen anyway directly put this into practice, just used as a shield to push for us to adopt LLMs

@cas ewaste numbers: https://www.bbc.com/news/science-environment-63245150

global south is worst affected by ewaste because its bought from global north. iirc there are individuals who scavenge for scraps and reuse parts too

and then there is https://en.wikipedia.org/wiki/Electronic_colonialism

i have mixed feelings about ai tbh

the speed at which i was able to mainline my device is insane: https://fosstodon.org/@Logical_Error/117325107817396215 getting Nura in more hands is better for privacy

but i hate ai companies, their greed, and their games they play

@cas I've been feeling more and more likely to contribute to Nura lately, especially helped by your hard no-LLM stance. Please keep it. I'm pretty busy now, but might have time do something about FP5 come winter.

@cas
I understand the concern about knowledge extraction, but I see it differently. I recently used an LLM to write a camera driver for an abandoned phone line in about 15 min, and devices that would otherwise end up as electronic waste are now running Linux again. These tools are available and affordable right now, and they can help small communities keep hardware alive long after vendors stop supporting it.The condition is that everything we build goes upstream and gets documented in the wiki

@sv1sjp in the same vein, a small town of almost exclusively local businesses can allow a coffee shop franchise to move in without much harm being done.

on an individual level, if you use an LLM to do something like write a camera driver that you otherwise wouldn't bother with and it made a big difference, sure, you can moralise it a bunch of ways

the problems arise when all the local businesses get replaced with franchises, when everyone becomes reliant on LLMs

we should be careful not to conflate individual and systemic critique

@Logical_Error @cas as someone who absolutely hates ewaste: these are very good points but we shouldn't assume LLMs are actually as useful as they seem. For one-off experiments you can indeed get lucky, but long term maintainability is looking incredibly grim. And that's something fundamentally unfixable (an impenetrable barrier from a theory POV) as long as we stay in the LLM space. And then there's the social & individual corrosion aspect.

@dngrs @cas yeah… its just a sad shitty situation all around

personally, i would not encourage people to buy a subscription, but i would like those who generate useless art to instead do something useful (eg. mainline their devices)

id want this at Nura: extremely discourage ai, but accept ai patches (with clear indication). maintainers can deprioritize ai submissions if needed

re maintainability: perfect is the enemy of good

It is a battle like the one between the sophists and Plato, who argued that truth and the sharing of that truth was more important than "successful outcomes". The sophists also charged high sums for their "knowledge". So once again we have a battle between those who want knowledge and truth to be free, and those who wish to capitalise upon it.

@cas
Where can I read about this decision?

@b3nis https://docs.nura.eco/policies-and-processes/development/ai-policy.html

there are some wording improvements in an open mr on the docs git repo too

@cas assume LLMs become error-free and powered by clean energy, and content scraping is minimized

using LLMs for FOSS projects would *still* be bad then, because they can only be trained - and thus controlled - by huge corporations

these corporations *will* exploit this control. we all know there's no such thing as a free dependency

I don't want to use projects that use LLMs because... I don't want to be controlled by corporations

is that ethics? it feels like pragmatism

@cas in my ideal future,
the LLMs go bankrupt.
all the investors, ruling class, who bet against humanity by replacing labour, art, skill, knowledge with an extension of capital, gets a level of punishment that they deserve (to be decided by the victims of their actions).
small language models under 100M for converting natural language to structured things like tool calls, code, JSON data, etc. these don't do any thinking, just language processing.

we are very close to the last thing.
see Cactus 14M, and Falcon 100M.
in my tests, falcon scores zero on logic tests: following a deduction chain, A => B, B => C, so if A, then C?, but in sentence form.
but it's starting to be good at doing needle in haystack: read a manual page and ask for a flag's description.
both cactus and Falcon are very good at tool calling: turn "lights off" to {home_assistant: { thing: "lights", action: "off" }}

I hope people realise that the cheap capable models serve most of our needs at negligible cost, and find no reason to pay the LLM overlords anymore. if that happens 1 will follow, and I hope 2 will also follow.

I also fear that the doom and gloom around the AI discourse make people depressed and hopeless, and feel that there's no point in investing in alternative tech to LLM, and they just surrender their lives, livelihood by renting them out from the capitalists. maybe that's precisely why they use the language they use when talking about AI.

@farseen @cas

The larger language models don't do any thinking either. :(

@resuna @cas yes, but it appears to follow logic and people seem to mistake that for intelligence. if we build a language model, which is just a program that can parse natural language, and that's it, then hopefully people will stop treating it as sentient.

@farseen @resuna @cas it only appears though, there is no logic in LLMs

@sv1sjp @cas you wrote nothing. You used fashtech to plagiarise and helped burn down our planet more while doing so.

@sv1sjp @cas oh wait, “blockchain” in the profile. You were already burning down the planet. The block chain you’re going on is mine. Well, block list.

@cas HN has had quite hot takes these past few, but so far I've seen this one a lot outside HN. Glad you're bringing it up and replying to it in a comprehensive toot ^^

Fair distinction, and I agree that mass dependence on LLMs is a real systemic concern. My case was individual: Xiaomi Mi Max 2 (oxygen, MSM8953) was dead, and my limited knowledge+ AI help got camera, torch and SIM working, so calls & SMS work on Nura. It extended a 10-year-old phone's life for even more, and thats the beauty of open source. Only the loudspeaker is left. All research and patches are public for anyone to reuse or to write manually with no AI for mainline: https://github.com/sv1sjp/mi-max2-oxygen-msm8953-nura-drivers

@sv1sjp I have checked your AI "work" and it is wrong in many ways. The audio part is completely unmaintainable and wrong direction clearly a slop.

For the camera part if it is working it still can cause legal headaches, anyway no one stops you to send those drivers to LKML if you do that and it will be accepted sooner or later it will be available in kernels used in nura.

SMS, Calls and modem stuff was working already no change was needed.

@cas I doubt it is 10x faster, if a SoCs support already in the kernel porting can be made very quickly. If someone learning the tricks how to make a port a lot of time can be reduced during the kernel bring up.

For the Nura part of the work making a device package is not too hard, packaging the kernel if it is not packaged yet also not too hard there are a bunch example in our repo what can help in these without any LLM.

@barni2000 i probably should have put it in quotes, the implication was that they may be exaggerating... that being said, at this point i've seen multiple cases where someone claims to throw claude at the problem for a day and get rather comprehensive results. it wouldn't surprise me if even a local model could do most of the boilerplate work and test loop faster than i could by hand, at least on average over time...

@barni2000
Hi, audio jut got redesigned and finally it works.
I can understand your concerns, but at the end of the date, we are talking about a 10 years old, end of life device.
Sure, not the best way to maintain it, but it has to be maintained for more than 8 years, so... It got revived

@sv1sjp If you think the code driver is in good shape, send it to LKML.