I think you should have been more clear: the most alarming use case of this tech would be some sick pedo taking pictures of your child then using that source imagery to generate fake porn and pleasuring himself all over it.
(Warning: I'm having a very hard time determining if you are trolling or are for real..)
That's not the most alarming use case of this tech. By far. (IMHO)
Also, I find this reasoning very off-putting. Putting child porn into a discussion kills it. All participants are (mostly) willing and basically required to agree and "let's not talk about this further".
The fundamental technology that underpins these achievements is more than capable of destroying civilization if things start to go south - which I believe they will, sooner or later. I find that to be more worthy of discussion than moral jousting about things people do in their private lives that I will - hopefully - never know about.
Let's all use our imagination and see where these kinds of models, both diffusion and transformers can take us. Sure they can generate plausible visual information, but that's not all they can do. Some days ago someone posted about ACT-1, a transformer for actions. People can and will hook up these things in all sorts of complicated pipelines and boy, generating some insensitive imagery is way, way down on the list of things to worry about.
So you've thoroughly defended against the point about talking about porn, but you give no examples of you say we should "truly worry about". Can you at least explain further? Sounds too hand wavy
Good point. I am being handwavey, sorry about that.
First, I see "AGI" as a real problem we'll have to face at some point. I believe we will be too late by the time we recognize it as a problem, so let's ignore that "threat" for now.
The more pressing problem IMO is that, to use technical terms, a shitload of people will have to face the reality that a software system is outperforming them on just about anything they are capable of doing professionally. I believe this will happen sooner than later and I am totally not seeing society being ready for that. Already I am seeing these models outperforming me - and my collegues - on quite a few important axes, which worries me and also the fact they almost universally dismiss it because it's not "perfect". I know it's hot these days to either under- or overestimate AI, but I do feel we have crossed a certain line. I don't see this genie going back into its bottle.
Perhaps I'm still handwavey. I guess I am a handwavey person and I'm sorry about that, but when I see GPT3 finishing texts with such grace I can't help but see a transformer also being capable of finishing "motor movements" or something else entirely like "chemical compounds", "electrical schematics" or even "legal judgements". I just found out about computational law BTW, might interest someone. Even just the "common sense" aspect of GPT3 is (IMO) amazing. Stuff like: we make eye contact during conversation, but we don't when driving. Why not? But also stuff like detecting in which room of the house we are based on which objects we see. That sort of stuff is amazing and it's a very general model too. Not trained on anything specific.
I guess the core of what I'm saying is that "predicting the next token" and getting it right often enough is frightenly close to what makes a large percentage of the human populace productive in a capitalist sense. I know I'm not connecting a lot of dots here, but I clearly lack the space, time and perhaps more importantly, the intelligence to actually do that. I fear I might be a handwavey individual - in fact easily replaced by GPT#. Do you now see why am I so worried? :)
How about you express your views as an addition to the conversation instead of as a criticism for other people not expressing the particular variety of concern that you have...?
