|
Top-level
You know how complicated is to train a pile of matrix multiplications to make it output what you want? The amount of choices you have available? Data augmentation, saddle points and optimizers, residual connections and architectural modifications, losses, normalization schemes, etc... I understand you are mad at AI grifters like Sam Altman because they are fucking up everything and making people unemployed, but please, that does not mean that training AI systems is trivial. 29 comments
I wish I could reproduce the original toot but it seems I have been blocked. If I'm not wrong it did insinuated that procedural art is more complex than training AI system. So my reply arguing otherwise was appropriate. Now, if we move into whether LLMs can be fair or not compared to procedural algorithms I must concede. But again, this was not the original topic. And ofc there is people working in AI reflect a lot on that and that is why there is a lot of people working in fairness and efficiency. Others only care about money, but that also happens in other fields. Finally, I'd like extend this concern for fairness and the consequences of automation to other fields, such as yours for instance. I see you hold a PhD in computer science. Have you wondered if the impact of all automatization carried thanks to computers? Did anyone gave a fuck about all the people who lost their jobs because an industrial robot replaced them in a factory? Probably not. The only difference is that instead of a poor fucker being the one to lose their job, now it is the turn of intellectual workers who (mistakenly) thought that they were free from being threatened by automatization. And the sheer scale of the problem. But that's all.  Rather more that it's an entirely different class, if they were on a Venn diagram they'd be separate circles. Individuals not understanding either getting hung up on the term generative. You don't have to jump to the defence of LLMs. Sadly, I cannot check it because I have been blocked. If it is not I'll apologize, but my interpretation is that "oh, procedural art is much harder than throwing heaps of data to a GPU".  @remenca @glowl "seeing pushback against procedural generation as a result of alleged 'AI' and that's really fucking sad You response is inaccurate. You can use CC0'd data to train your model, you can do it using your own gaming GPU, without incurring in more energy cost than gaming which is a socially accepted practice. And then, if you manage to train the network then there is all the difficulty involved with prompting until the network finally spits an imagine you are satisfied with. There is also the philosophical argument about what is creativity other just mixing already seen inputs, and Hume and Kant and bla bla bla. Imho creativity is precisely this, combining things that exist in novel ways, because ex nihilo nihil fit. After all, the human illustrator has learned by copying others too. In the end I do not see such a difference. @remenca You're missing the point. You can use CC0 data, but you don't because it's just not enough training data. And I don't care about what you think about creativity. That wasn't the argument in the first place. This is about the human cost, the human effort. No matter how "hard" it is to train your model with unethically obtained data. YOU are not the one being creative here. And it shows. I think that you are the wrong here. Please tell me what's unethical about using this model. @remenca The fact that I'm exposed to such horrid "art". As I already said, it's just not enough data. It gives terrible results. @remenca I answered it in the first sentence. I obviously ran on the assumption that you didn't use CC0 data. But if you do that's fine. It doesn't contradict my previous arguments. Or OP's argument for that matter. Regular artists (including the ones creating good procedural generation) obtain much better results than with any generative model, CC0 or not. And then, when we only consider CC0 data, the results are so much worse that I don't even know why you're using it as an argument in the first place. I am using it because you stated that there is no way to train a network on CC0'd data and some other bullshit, and I wanted to prove you wrong. Now that we have agreed on that and we can dismiss the "its unethical" argument, we can move to I never claimed that AI generated art is better than procedurally created art, I just said it was art, and now you are trying to play an strawman on me. I do not think that this conversation is in good faith, so I would like to conclude here. @remenca I never claimed that you can't. I said that you don't (as in: people don't). It's unethical because the ethically obtained results are shit, plain and simple. If you come in good faith you shouldn't have replied to a thread with an irrelevant argument in the first place. OP is not talking about that. I just wanted to explain _why_ it is irrelevant and why you got blocked by Eniko. I replied like I replied because it makes me mad to see so many people shitting on AI, in a manner similar to Eniko. Not all AI is Sam Altman. I fault people for putting the blame on AI when, as you are admitting, the blame is with capitalism. But I don't know. Maybe you are right. I started researching AI years ago thinking that we were working to make a better world where people wouldn't need to work because machine would do it for us, and all I got is this exploitative genAI bullshit that makes climate change worse and makes people lose their jobs and I'm very mad at it. I think I'll change careers. |
Gregory's Server
@remenca @eniko No one here is calling the latest generations of AI "trivial", but whether the models and algorithms made using any kind of AI contribute to fairness and sustainability is something each and everyone working with them needs to reflect upon seriously. It is substantially harder for an LLM to be fair and sustainable than a crafted procedure-based generation algorithm, so the latter seemingly receiving backlash as of late is unreasonable.