Doesn't make sense to me why OpenAI has kept DALL-E closed source for so long. I can only guess either safety from misuse or leveraging it for money. At this rate though, Stable Diffusion is going to dwarf it
I still don't understand what this would mean. Where are all of the terrible things that were supposed to happen, now that Stable Diffusion is available?
We've been able to create completely photorealistic fiction for decades now. See any movie with CGI for an example of whole worlds, and people, that don't exist. The bar has gradually been lowering (see the amazing CGI that YouTubers do these days), and now maybe there is a bit of a step function down, but being able to make things that aren't real isn't remotely new. I don't understand the fear.
The danger for society is in not already knowing that is the case, since it's relatively trivial, without AI, to make fake content.
We are way past that. If anything, trivialising and democratising the generation of fake content, will educate the masses. Perhaps, bring back the trust for science, which we need a lot.
Obligatory XKCD:https://xkcd.com/2650/
I'm pretty sure one of the primary arguments for Dall-E and Stable Diffusion existing is that there are lots of people who can't draw anything.
On one side there's no problem to like CP if it's AI generated and no real kids are harmed.
On the other side it may desensitize the public to this type of content.
> I still don't understand what this would mean. Where are all of the terrible things that were supposed to happen, now that Stable Diffusion is available?
Mainly people making porn (e.g. stuff like deepnudes). It seems like a lot of work has gone into into preventing that (e.g. filtering porn out of training data, having porn-detection models to block porny output). There's also been a lot of talk about political fakes, etc, but I'm not sure how likely that is to actually happen at this point. I think one of the "selling points" of limiting access to DALL*E was that they could revoke access to people who they deemed to be misusing it.
I wonder is it possible to distribute the task for volunteer's idle GPU (like BOINC)
Also you’re probably not seeing all the pedo content that people are already generating for themselves.
I think that's a good thing. A slow creep is dangerous, because people may not notice, and would be tricked. An explosion of insanity makes everyone realize they can't believe everything they see on the internet.
We'll get more Q-Anon than Wikipedia.
Any favorite examples?
Paternalistic moralizing as a method to discriminate who gets access to models. Everyone else gets these cloud-service table scraps. That's why Stable Diffusion is so awesome -- YOU have the model!
It seems more like they were trying to accomplish the opposite.
https://www.technologyreview.com/2020/09/23/1008729/openai-i...
And being so open they first lock the model up and charge a fee, so anyone can pay. Just spreading democracy through paid API calls. /s
I was a bit mean, they did kick the field in the butt and pushed us ahead even with all the stubbornness and secrecy. But now they are just holding us back.
You might think that the nuclear bomb is a good analogy to use here, but it is not, because once the field has advanced to the point in which one group can develop AGI, it is now possible for other groups to develop it with relative ease, unless you actively take over the world first and deny those other groups the compute resources necessary to train/run AGI.
The point is, once these algorithms are upon us, you must be willing to accept what impacts they will have, even if it destroys entire industries. The alternative being that you destroy the industry slowly rather than quickly, while simultaneously widening the gap between the elites and everyone else.
The mistake is thinking that people can't adapt to the times, which is only true if you are actively holding them back.
If someone developed AGI today, the best thing to do would be to instantly throw up a torrent of it and spread it as fast as possible, because if a sole entity is able to get it first and kick the ladder away, we are most likely screwed.
Of course maybe it's just because text is less visually impressive than images.
BLOOM incorporates some new ideas like ALiBi which might make it better in a more general sense. They haven’t released official evaluation numbers yet though so we’ll have to see.
The former is being used as justification for the latter.
and it shouldn't be difficult to pick which of these is actually true.
>leveraging it for money
Funny thing is that people didn't stand still and invented diffusion and other CLIP guided image synthesis methods, and DALL-E 2 copied the method, completely changing from the first architecture.
Their arrogance is that they think they can ride the dragon. They want to be the ones to discover, advance it, and control it. But everyone else doesn't have time for that shit.