7 ms·
> Of course, with open access and the ability to run the model on a widely available GPU, the opportunity for abuse increases dramatically. > “A percentage of
by PoignardAzur 4y ago
> Of course, with open access and the ability to run the model on a widely available GPU, the opportunity for abuse increases dramatically.
> “A percentage of people are simply unpleasant and weird, but that’s humanity,” Mostaque said. “Indeed, it is our belief this technology will be prevalent, and the paternalistic and somewhat condescending attitude of many AI aficionados is misguided in not trusting society.”
Holy shit.
On the one hand, I'm super excited by this technology, and the novel applications that will become possible with these open-source models (stuff that would never be usable if Google and OpenAI had a monopoly on image generation).
On the other hand, I really really really hope Bostrom's urn[0] has no black ball in it, because we as a society seem to be rushing to extract as many balls as possible over increasingly short timescales.
[0] https://nickbostrom.com/papers/vulnerable.pdf https://nickbostrom.com/papers/vulnerable.pdf
- sinenomine 4y agoWhat if the black ball was a red herring all along and the usual suspect tech-CEO's hand(s) rushing to control the said crystal ball are the real hazard?
- jackblemming 4y agoYes, I feel much safer if OpenAI and Google are the sole keepers of such technology. They have my and the publics best interest at heart.
- geraldwhen 4y agoIs this satire?
- pizza 4y agoYes
- ljlolel 4y agoGooglers and OpenAI legitimately believe this
- dkjaudyeqooe 4y agoSarcasm.
- PoignardAzur 4y agoLet me put it this way: it's not great to live in a world where the immense majority of nukes are controlled by Donald Trump and Vladimir Putin. But it's arguably better than living in a world where every single citizen has a nuke. (Though the potential for harm of diffusion models is far below nuke; it's not "kill millions of people", it's "produce cheap disinformation and very convincing fake evidence to ruin someone's life")
- nightski 4y agoI think that would be a tough argument to make (in regards to image generation). The same could be said of just about any computing technology. The problem is we lose out on a lot of potential good. Either way it doesn't matter, you can't control bits like you can enriched uranium. It's just a matter of time. In the grand scheme of things Open AI will be irrelevant.
- manquer 4y agoWouldn’t nuclear weapons or even plastics be a black ball already? Humanity is not homogeneous, we will always react to new inventions or tools differently , many will use it positively some won’t . Short of weapons of mass destruction I am not sure anything else will destroy civilization itself .
- ALittleLight 4y agoNo, the black ball is a technology that, once invented, humanity cannot survive. Nuclear weapons have been invented and humanity is surviving. Same with plastics. A black ball would be like - suppose nuclear weapons ignited the atmosphere. We test the first nuke, it ignites the atmosphere, a global fire storm consumes all breathable oxygen, kills all plants and everyone on the surface and everything else suffocates shortly after. Plastics aren't even close to this level of harm.
- neuronic 4y ago> Same with plastics Plastics are playing the long game. They have to turn into micro- and nanoplastics first and may then enact undesired, unforeseen biological functions, just like BPA [1]. Not even talking about weaponizing this stuff... [1] https://pubmed.ncbi.nlm.nih.gov/21605673/ https://pubmed.ncbi.nlm.nih.gov/21605673/
- ALittleLight 4y agoIt seems implausible to me that plastics are going to kill all humanity. If the paper you linked makes that case, then I will read it, but I didn't get that from skimming the abstract.
- michaelt 4y agoI guess people are getting confused because in terms of risk-of-destroying-humanity, nuclear weapons seem higher risk than DALL-E.
- PoignardAzur 4y ago
- krono 4y agoEither we equalise chaos or we reduce chaos, there exist no other options for entropy incarnate.
- mortenjorck 4y agoThe length of the democratization cycle we're seeing – months to weeks between a breakthrough model and a competent open-source alternative that runs on commodity hardware – really highlights the genie-stuffing posture of Google and OpenAI. All the thoughtful, if highly paternalistic guardrails they build in amount to little more than fig leaves over the possible applications they intend to close off. I'm personally in the "AI risk is overstated" camp. But if I'm wrong, all the top-down AI safety in the world is going to be meaningless in the face of a global network of researchers, enthusiasts, and tinkerers.
- hedora 4y agoThey claim the guardrails are for public good, but they're pretty clearly using them to try to establish a competitive moat. It's similar to the "we don't sell personal information" claim. Sure, but that's because they make money renting malicious actors access to a black box that contains your personal information. Selling the contents of the box would reduce their overall revenue.
- TulliusCicero 4y agoTo me it seems like an obvious case of reputational risk being much larger for more prominent organizations than for smaller ones. It makes sense for Google to wait for some startup to "go first" in releasing a model largely without controls. That way, some random startup takes the initial heat of "people are using AI for bad things!!" headlines plastering tech blogs. Then Google can do basically the same thing a little bit later, and any attack pieces will sound old hat.
- narrator 4y agoI was using AI Dungeon with the full power GPT-3 model before they crippled it. That thing had a very uninhibited mind for erotica! Imagine what would happen when that power comes to image models!
- lodovic 4y agoIt's only a matter of time before someone trains a model on stills from the major adult sites. Actually surprised it hasn't been done / made public yet.
- CM30 4y agoI don't see why this is incorrect. It seems ever since DALL-E, Midjourney, caught on, it seems like we've got more and more people trying to 'filter out' incorrect uses of their software under the assumption people cannot be trusted to just use it for whatever they want. And it depresses me, because well... imagine if other pieces of tech were treated this way. If the internet or crypto or computers or whatever were heavily limited/restricted so the 'wrong people' couldn't use them for bad things. We'd consider it ridiculous, yet it's somehow accepted for these image generation systems.
- rcxdude 4y agogood cryptography was treated this way for a while, at least by the US government.
- atwood22 4y agoYes, and look how many amazing uses of cryptography have proliferated after it was made available, unrestricted, to the masses.
- dinosaurdynasty 4y agoNuclear tech is treated this way (much stricter even).
- CM30 4y agoI think there might be at least a small difference between nuclear tech and image generation, at least as far as the effects that could happen if it goes wrong.
- dkjaudyeqooe 4y agoGet back to me when you can vaporise a city with a AI generated image.
- Aeolun 4y agoYou just need the right AI generated image.
- Iv 4y agoPeople will generate creepy porn and fake pictures. Humanity will survive this.
- forty 4y agoExactly, and there are many pros for humanity to this too: people will be able to make funny pictures and things like that, so it's not like it's a bad deal.
- PoignardAzur 4y agoYou're not addressing my broader point, though, just the easy-to-snide-at version of my point. Yes, it's pretty obvious that Dall-E and similar models won't destroy humanity. My point isn't that Dall-E is a black ball. My point is we better hope a black ball doesn't exist at all, because the way this is going, if it exists, we are going to pick it, we clearly won't be able to stop ourselves. (For the sake of dicussion, we can imagine a black ball could be "a ML model running on a laptop that can tell you how to produce an undetectable ultra-transmissible deadly virus from easily purchased items")
- dkjaudyeqooe 4y agoSocial media is being used to undermine societies globally, arguably it has a fair chance of destroying humanity. So I think that horse may have bolted already.
- Iv 4y agoIf you know that image generators are not it, then why talk about it here? Do you get this kind of angst at every technological increment? Apart from nuclear scientists I don't know a field where participants are as conscious of the risks as AI research.
- krisoft 4y ago> Apart from nuclear scientists I don't know a field where participants are as conscious of the risks as AI research. Great. Now some of these researchers preceived some risk with this technology. Not human extinction level risk, but risks. So they attempted to control the technology. To be specific: OpenAI is worried about deepfakes so they engineered guard rails into their implementation. OpenAI was worried about misinformation so they did not release the bigger GPT models. Note: I’m not arguing either way if OpenAI was right, or honest about their motivations just observing that they expressed this opinion and acted on it to guard the risk. Got this so far? Keep this in mind because I’m going to use this information to answer your question: > If you know that image generators are not it, then why talk about it here? Because it is a technology which were deemed risky by some practitioners and they attempted to control it, and those attempts to control the spread of the technology failed. This does not bode well towards our ability to restrain ourselves from picking up a real black ball, if we ever come across one. And that is why it is worth talking about black balls in this context. Note it is unlikely that a black ball event will completely blindside us. It is unlikely that someone develops a clone of pacman with improved graphics and boom that alone leads to the inevitable death of humanity. It is much more likely that when the new and dangerous tech appears on our horizon there will be people talking about the potential dangers. What remains to be answered: what can we do then? This experience has shown us that if we ever encounter a black ball technology the steps taken by OpenAI doesn’t seem to be enough. This is why it is worth talking about black ball technologies here. I hope this answers your question?
- adwi 4y ago> [0] https://nickbostrom.com/papers/vulnerable.pdf https://nickbostrom.com/papers/vulnerable.pdf Mobile friendly: https://onlinelibrary.wiley.com/doi/10.1111/1758-5899.12718 https://onlinelibrary.wiley.com/doi/10.1111/1758-5899.12718