@mnl This sounds perfectly sane, but you will get bunched in with people who act naively, and the political situation around us makes it very hard to dinstinguish your behavior from a stan, because most people are less competent and experienced about these things than you are?

You are a privileged early adopter. Why does it surprise you, that this is not an equal position?

@promovicz that’s exactly my point lol. The discussion we are having is the ones I want to have. It’s been 3 years now, everybody knows what ChatGPT is, there’s front page articles about Claude code in the nytimes. Yet the discourse here is often still “AI is useless” because people assume that we’re still in the gpt-3.5 times or something. That’s _on them_, honestly. I don’t have some super privileged starting position besides being someone working as an engineer and able to afford a $20 sub. I’m just a dude with a computer. I barely have friends in tech, I’m not in a tech hub, i don’t have a degree, I don’t have a secret network of insiders.

Because taking a knee jerk anti ai stance makes it impossible to discern signal from noise, and thus take meaningful action. All that’s left is diffuse fear.

@promovicz I wrote this 3 years ago: https://the.scapegoat.dev/llms-a-paradigm-shift-for-the-pragmatic-programmer/

I was called delusional (still am) and told I have no clue about software engineering. I could more easily forgive it then because models were really difficult to wield. That’s not the case anymore, anybody using even opensource models running on my local i7 (I don’t even own a gpu!) is miles ahead of where things were. And yet the discourse on mastodon doesnt seem to have progressed an inch…

@mnl I don't really disagree. Locally run models would be cool, but you don't have the training data - right?

I also think that text-based LLMs may get beaten by token/construct-based approaches? It's not like this whole subject has no interesting angles... :)

@mnl I see no parallel between fimware and training data. The point is, that your local model is not independent. Most of them are *both* stolen and based on unclear sourcing. And that's a problem, in any case. It's not about copyright/left either - but about powerful entities dictating policy through involuntary means. As long as we carry that, or profit from it, this problem won't go away.

Who pays me, for my contribution to the LLMs? This is not optional. It is a right.

@promovicz the firmware being a black box also means it’s outside of one’s control. I personally dgaf about copyright, and while I think companies should be required to at least provide the training corpus metadata to the public and respect “do not train” licenses, in order to actually contribute back to the commons, and it also putting a stick right where it hurts (properly labeled training data is hard in the same that deleting data is hard, vs just harvesting stuff), I am glad to be in the training corpus and I never wrote opensource for compensation or recognition. It might be different for other people, and it’s a valid point. Not one I personally agree with, actually out of moral principles: I’m privileged enough to have the skills and time to write opensource, and everybody should benefit from it, even if they don’t mention my name in some obscure file no one ever reads (speaking of MIT here, which I’ve used since the nineties).

The whole copyright thing is a double edged sword for big tech anyway, as we’ve seen with the Claude code leak _immediately_ being laundered. Now let’s do it for the wifi firmware as well (to circle back 🤣).

@mnl but wi-fi firmware can be reversed and reconstructed. OSS has done that repeatedly, in various forms.

you do not have to care about copyright, to appreciate that other people will not allow having their rights violated. this is where you come off as snobbish and privileged. you may not care - but what about other people’s livelihood? the state has a duty to maintain a minimum balance of rights, and LLMs would tip towards corporate rights. I don’t want that tendency.

@promovicz please read what I wrote. I said I want companies to be regulated when it comes to the training data. I said wanting to have owns copyright respected is valid. I don’t know how to put it in clearer terms.

And models excel at reverse engineering, which is why despite having the skills I have reversed more in the last year than in my entire career, because it made it so much more accessible.

@mnl Let me put it differently:

As a scene hacker and pirate (myself) I can accept your local compromise, but I can not make it non-marginal to the mainstream debate. You are acting ahead of the law and its perceived justice. A public-acting politician cannot do that. They have to argue yesterday's justice, or take their followers along for a transition in zeitgeist - and they can't easily do both at the same time. AI marketing tactics are the cause of this blockage, through disruption.