Hacker News Reader

Comment by patates | Hacker News Reader

Half of the announcement talks about safety. The next step will be these control mechanisms being built into all sorts of software I suppose.

It's "safe" for them, not for the users, at least they should make that clear.

View on HN

What's equally interesting is that while they spend a lot of words on safety, they don't actually say anything. The only hint what they even mean by safety is that they took "reasonable steps" to "prevent misuse by bad actors". But it's hard to be more vague than that. I still have no idea what they did and why they did it, or what the threat model is.

Maybe that will be part of future papers or the teased technical report. But I find it strange to put so much emphasis on safety and then leave it all up to the reader's imagination.

Remember when AI safety meant the computers weren’t going to kill us?

Now people spend a lot of time making them worse to ensure we don’t see boobs.

I truly wonder what "unsafe" scenarios an image generator could be used for? Don't we already have software that can do pretty much anything if a professional human is using it?

I would say the barrier to entry is stopping a lot of ‘candid’ unsafe behaviour. I think you allude to it yourself in implying currently it requires a professional to achieve the same results.

But giving that ability to _everyone_ will lead to a huge increase in undesirable and targeted/local behaviour.

Presumably it enables any creep to generate what they want by virtue of being able to imagine it and type it, rather than learn a niche skill set or employ someone to do it (who is then also complicit in the act)

"undesirable local behavior"

Why don't you just say you believe thought crime should be punishable?

[Edited: I'm realizing the person I'm responding to is kinda unhinged, so I'm retracting out of the convo.]

I imagine they might talk about things like students making nudes of their classmates and distributing them.

Or maybe not. It's hard to tell when nobody seems to want to spell out what behaviors we want to prevent.

Students already share nudes every day.

Where are the Americans asking about Snapchat? If I were a developer at Scnapchat I could prolly open a few Blob Storage accounts and feed a darknet account big enough to live off of. You people are so manipulatable.

Students don’t share photorealistic renders of nude classmates getting gangbanged though

Would it be illegal for a student who is good at drawing to paint a nude picture of an unknowing classmate and distribute it?

If yes, why doesn't the same law apply to AI? If no, why are we only concerned about it when AI is involved?

Because AI lowers the barrier to entry; using your example, few people have the drawing skills (or the patience to learn them) or take the effort to make a picture like that, but the barrier is much lower when it takes five seconds of typing out a prompt.

Second, the tool will become available to anyone, anywhere, not just a localised school. If generating naughty nudes is frowned upon in one place, another will have no qualms about it. And that's just things that are about decency, then there's the discussion about legality.

Finally, when person A draws a picture, they are responsible for it - they produced it. Not the party that made the pencil or the paper. But when AI is used to generate it, is all of the responsibility still with the person that entered the prompt? I'm sure the T's and C's say so, but there may still be lawsuits.

Right, these are the same arguments against uncontrolled empowerment that I imagine mass literacy and the printing press faced. I would prefer to live in a society where individual freedom, at least in the cognitive domain, is protected by a more robust principle than "we have reviewed the pros and cons of giving you the freedom to do this, and determined the former to outweigh the latter for the time being".

You seem to be very confused about civil versus criminal penalties....

Feel free to make an AI model that does almost anything, though I'd probably suggest that it doesn't make porn of minors as that is criminal in most jurisdiction, short of that it's probably not a criminal offense.

Most companies are only very slightly worried about criminal offenses, they are far more concerned about civil trials. There is a far lower requirement for evidence. AI creator in email "Hmm, this could be dangerous". That's all you need to lose a civil trial.

Why do you figure I would be confused? Whether any liability for drawing porn of classmates is civil or criminal is orthogonal to the AI comparison. The question is if we would hold manufacturers of drawing tools or software, or purveyors of drawing knowledge (such as learn-to-draw books), liable, because they are playing the same role as the generative AI does here.

Because you seem to be very confused on civil liabilities in most products. Manufactures are commonly held liable for the users use of products, for example look at any number of products that have caused injury.

Surely those are typically when the manufacturer was taken to have made an implicit promise of safety to the user and their surroundings, and the user got injured. If your fridge topples onto you and you get injured, the manufacturer might be liable; if you set up a trap where you topple your fridge onto a hapless passer-by, the manufacturer will probably not be liable towards them. Likewise with the classic McDonalds coffee spill liability story - I've yet to hear of a case of a coffee vendor being held liable over a deliberate attack where someone splashed someone else with hot coffee.

> You seem to be very confused about civil versus criminal penalties....

Nah, I think it's a disagreement over whether a tool's maker gets blamed for evil use or the tool's user.

It's a similar argument over whether or not gun manufacturers should have any liability for their products being used for murder.

>It's a similar argument over whether or not gun manufacturers

This is really only a debate in the US and only because it's directly written in the constitution. Pretty much no other product works that way.

Are we on the same HN that bashes Facebook/Twitter/X/TikTok/ads because they manipulate people, spread fake news or destroyed attention span?

Can you point to other crimes that are based on skill or effort?

Photoshop also lowers that barrier of entry compared to pen and pencil. Paper also lowers the barrier compared to oil canvas.

Affordable drawing classes and YouTube drawing tutorials lower the barrier of entry as well.

Why on earth would manufacturers of pencils, papers, drawing classes, and drawing software feel responsible for censoring the result of combining their tool with the brain of their customer?

A sharp kitchen knife significantly lowers the barrier of entry to murder someone. Many murders are committed everyday using a kitchen knife. Should kitchen knife manufacturers blog about this every week?

I agree with your point, but I would be willing to bet that if knives were invented today rather than having been around awhile, they would absolutely be regulated and restricted to law enforcement if not military use. Hell, even printers, maybe not if invented today but perhaps in a couple years if we stay on the same trajectory, would probably require some sort of ML to refuse to print or "reproduce" unsafe content.

I guess my point is that I don't think we're as inconsistent as a society as it seems when considering things like knives. It's not even strictly limited to thought crimes/information crimes. If alcohol were discovered today , I have no doubt that it would be banned and made schedule I

> Hell, even printers, maybe not if invented today but perhaps in a couple years if we stay on the same trajectory, would probably require some sort of ML to refuse to print or "reproduce" unsafe content.

Fun fact: Many scanners and photocopiers will detect that you're trying to scan/copy a banknote and will refuse to complete the scan. One of the ways is detecting the EURion Constellation.

https://en.wikipedia.org/wiki/EURion_constellation

IANAL but that sounds like harrassment, I assume the legality of that depends on the context (did the artist previously date the subject? lots of states have laws against harassment and revenge porn that seem applicable here [1]. are you coworkers? etc), but I don't see why such laws wouldn't apply to AI generated art as well. It's the distribution that's really the issue in most cases. If you paint secret nudes and keep them in your bedroom and never show them to anyone it's creepy, but I imagine not illegal.

I'd guess that stability is concerned with their legal liability, also perhaps they are decent humans who don't want to make a product that is primarily used for harassment (whether they are decent humans or not, I imagine it would affect the bottom line eventually if they develop a really bad rep, or a bunch of politicians and rich people are targeted by deepfake harassment).

[1] https://www.cagoldberglaw.com/states-with-revenge-porn-laws/...

^ a lot of, but not all of those laws seem pretty specific to photographs/videos that were shared with the expectation of privacy and I'm not sure how they would apply to a painting/drawing, and I certainly don't know how the courts would handle deepfakes that are indistinguishable from genuine photographs. I imagine juries might tend to side with the harassed rather than a bully who says "it's not illegal cause it's actually a deepfake but yeah i obviously intended to harass the victim"

That's not even necessarily a bad thing (as a whole - individually it can be). Now, any leaked nudes can be claimed to be AI. That'll probably save far more grief than it causes.

Such activity is legal per Ashcroft v Free Speech Coalition (2002). Artwork cannot be criminalized because of the contents of it.

Artwork is currently criminalized because of its contents. You cannot paint nude children engaged in sex acts.

The case I literally just referenced allows you to paint nude children engaged in sex acts.

> The Ninth Circuit reversed, reasoning that the government could not prohibit speech merely because of its tendency to persuade its viewers to engage in illegal activity.[6] It ruled that the CPPA was substantially overbroad because it prohibited material that was neither obscene nor produced by exploiting real children, as Ferber prohibited.[6] The court declined to reconsider the case en banc.[7] The government asked the Supreme Court to review the case, and it agreed, noting that the Ninth Circuit's decision conflicted with the decisions of four other circuit courts of appeals. Ultimately, the Supreme Court agreed with the Ninth Circuit.

I appreciate you taking the time to lay that out, I was under the opposite impression for US law.

Stability is not an American company. The US is not the only country in the world.

You’re very welcome to ask for clarification - I kept it abstract because there is a lot of grey area and it’s something we need to understand and discuss as technology and society evolves.

To spell out one such instance: I would like to live in a world where it is not trivial to depict and misrepresent me (or anyone) in a way that is photorealistic to the point that it can be used to mislead others.

Whether that means we need to outright prevent it, or have some kind of authenticity mechanism, or some other yet-to-be-discovered solution? I do not know, but you now have my goalposts.

What do you mean should be... it 100% is.

In a large number of countries if you create an image that represents a minor in a sexual situation you will find yourself on the receiving side of the long arm of the law.

If you are the maker of an AI model that allows this, you will find yourself on the receiving side of the long arm of the law.

Moreso, many of these companies operate in countries where thought crime is illegal. Now, you can argue that said companies should not operate in those countries, but companies will follow money every time.

I think it's pretty important to specify that you have to willingly seek and share all of these illegal items. That's why this is so sketch. These things are being baked with moral codes that'll _share_ the information, incriminating everyone. Like why? Why not just let it work and leave it up to the criminal to share their crimes? People are such authoritarian shit-stains, and acting like their existence is enough to justify their stance is disgusting.

>I think it's pretty important to specify that you have to willingly seek and share all of these illegal items.

This is not obvious at all when it comes to AI models.

>People are such authoritarian shit-stains

Yes, but this is a different conversation altogether.

Once it is outside your mind and in a physical form, is it still just a thought sir?

In my country there is legal precedent setting that private, unshared documents are tantamount to thought.

Eh, a professional human could easily lockpick the majority of front doors out there. Nevertheless I don't think we're going to give up on locking our doors any time soon.

Similar to why Google's latest image generator refuses to produce a correct image of a 'Realistic, historically accurate, Medieval English King'. They have guard rails and system prompts set up to force the output of the generator with the company's values, or else someone would produce Nazi propaganda or worse. It (for some reason) would be attributed to Google and their AI, rather than the user who found the magic prompt words.

Yeah this is probably the most realistic reason

For some scenarios, it's not the image itself but the associations that the model might possibly make from being fed a diet of 4chan and Stormfront's unofficial YouTube channel. The worry is over horrible racist shit, like if you ask it for a picture of a black person, and it outputs a picture of a gorilla. Or if you ask it for a picture of a bad driver, and it only manages to output pictures of Asian women. I'm sure you can think up other horrible stereotypes that would result in a PR disaster.

The major risky use cases for image generators are (a) sexual imagery of kids and (b) public personalities in various contexts usable for propaganda.

thanks, i hadn't fully realized that 'safety' means 'safe to offer' and not 'safe for users'. i won't forget it

This is the world we live in. CYA is necessary. Politicians, media organizations, activists and the parochial masses will not brook a laissez faire attitude towards the generation of graphic violence and illegal porn.

Not even legal porn, unfortunately. Or even the display of a single female nipple…

looking at the manual censorship of the big channels on youtube, you don't even need to display anything, just suggesting it is enough to get a strike.

(of course unless you are into yoga, then everything is permitted)

> (of course unless you are into yoga, then everything is permitted)

...or children's gymnastics.

> This is the world we live in.

Great talk about slavery and religious-persecution, Jim! Wait, what were we talking about? Fucking American fascists trying to control our thoughts and actions, right right.

PSA: There are now calls to embed phone-home / remote kill switch mechanisms into hardware because “AI safety”.

examples? seems like it would be easier to instead communicate with ISPs.

Hmm, all computers have remote kill switches in them unless you have a generator at home.

They rather talk about "reasonable steps" to safety. Sounds like "just the minimum so we don't end up in legal trouble" to me...

Any large publicly available model has no choice but to do this. Otherwise, they're petrified of a PR nightmare.

Models with a large user base will have an inverse relationship with usability. That's why it's important to have options to train your own with open source.

I think this AI safety thing is great. These models will be used by people to make boring art. The exciting art will be left for people to make.

This idea of AI doing the boring stuff is good. Nothing prevents you from making exciting, dangerous, or 'unsafe' art on your own.

My feeling is that most people who are upset about AI safety really just mean they want it to generate porn. And because it doesn't, they are upset. But they hide it under the umbrella of user freedom. You want to create porn in your bedroom? Then go ahead and make some yourself. Nothing stopping you, the person, from doing that.

There is some truth in what you say, just like saying you're a "free speech absolutist" sounds good at first blush. But the real world is more complicated, and the provider adds safety features because they have to operate in the real world and not just make superficial arguments about how things should work.

Yes, they are protecting themselves from lawsuits, but they are also protecting other people. Preventing people asking for specific celebrities (or children) having sex is for their benefit too.

I really wish that every discussion about a new model didn’t rapidly become a boring and shallow discussion about AI safety.

AI is not an engineered system; it's emergent behavior from a system we can vaguely direct but do not fundamentally understand. So it's natural that the boundaries of system behavior would be a topic of conversation pretty much all the time.

EDIT: Boring and shallow are, unfortunately, the Internet's fault. Don't know what to do about those.

At least in some latest controversies (e.g. Gemini generation of people) all of the criticized behavior was not emergent from ML training, but explicitly intentionally engineered manually.

But that's the thing, prompt formulation is not engineering in the sense I'm talking about. We know why a plane flies, we know why an engine turns, we know how a CPU works - mostly. We don't know how GenAI gets from the prompt to the result with any specificity at all. Almost all the informational entropy of the output is hidden from us.

This announcement only mentions safety. What else do you expect to talk about?

It's also "safety" in the sense that you can deploy it as part of your own application without human review and not have to worry that it's gonna generate anything that will get you in hot water.

I agree with you, but when companies don't implement these things, they get absolutely trashed in the press & social media, which I'm sure affects their business.

What would you have them do? Commit corporate suicide?

This is a good question. I think it would be best for them to give some sort of signal, which would mean "We're doing this because we have to. We are willing to change if you offer us an alternative." If enough companies/people did this, at some point change would become possible.

I get a slightly uncomfortable feeling with this talk about AI safety. Not in the sense that there is anything wrong with that (may be or may be not), but in the sense I don't understand what people are talking about when they talk about safety in this context. Could someone explain like I have Asperger (ELIA?) whats this about? What are the "bad actors" possibly going to do? Generate (child) porn/ images with violence etc. and sell them? Pollute the training data so that the racist images pops up when someone wants to get an image of a white pussycat? Or produce images that contain vulnerabilities so that when you open that in your browser you get compromised? Or what?

> Could someone explain like I have Asperger (ELIA?)

Excuse me?

You sound offended. My apologies. I had no intention whatsoever to offend anyone. Even if I am not diagnosed, I think I am at least borderline somewhere in the spectrum, and thought that would be a good way to ask people explain without assuming I can read between the lines.

Let's just stick with the widely understood "Explain Like I'm 5" (ELI5). Nobody knows you personally, so this comes off quite poorly.

I think ELI5 means that you simplify a complex issue so that even a small kid understands it. In this case there is no need to simplify anything, just explain what a term actually means without assuming reader understanding nuances of terms used. And I still do not quite get how ELIA can be considered hostile, but given the feedback, maybe I avoid it in the future.

Saying "explain like I have <specific disability>" is blatantly inappropriate. As a gauge: Would you say this to your coworkers? Giving a presentation? Would you say this in front of (a caretaker for) someone with Autism? Especially since Asperger's hasn't even been used in practice for, what, over a decade?

> In this case there is no need to simplify anything

Then just ask the question itself.

AI isn't a coworker, not a human so it's not as awkward to talk about one's disability.

I don't see how this is a response to anything I've said. They're speaking to other humans and the original use of their modified idiom isn't framed as if one were talking about their own, personal disability.

Just as an example:

https://arstechnica.com/information-technology/2024/02/deepf...

The bad actor might be the model itself, e.g., returning unwanted pornography or violence. Do you have a problem with Google’s SafeSearch?

I'm not part of Stability AI but I can take a stab at this:

> explain like I have ~~Asperger (ELIA?)~~ limited understanding of how the world really works.

The AI is being limited so that it cannot produce any "offensive" content which could end up on the news or go viral and bring negative publicity to Stability AI.

Viral posts containing generated content that brings negative publicity to Stability AI are fine as long as they're not "offensive". For example, wrong number of fingers is fine.

There is not a comprehensive, definitive list of things that are "offensive". Many of them we are aware of - e.g. nudity, child porn, depictions of Muhammad. But for many things it cannot be known a priori whether the current zeitgeist will find it offensive or not (e.g. certain depictions of current political figures, like Trump).

Perhaps they will use AI to help decide what might be offensive if it does not explicitly appear on the blocklist. They will definitely keep updating the "AI Safety" to cover additional offensive edge cases.

It's important to note that "AI Safety", as defined above (cannot produce any "offensive" content which could end up on the news or go viral and bring negative publicity to Stability AI) is not just about facially offensive content, but also about offensive uses for milquetoast content. Stability AI won't want news articles detailing how they're used by fraudsters, for example. So there will be some guards on generating things that look like scans of official documents, etc.

So it's just fancy words for safety (legal/reputational) for Stability AI, not users?

Yes*. At least for the purposes of understanding what the implementations of "AI safety" are most likely to entail. I think that's a very good cognitive model which will lead to high fidelity predictions.

*But to be slightly more charitable, I genuinely think Stability AI / OpenAI / Meta / Google / MidJourney believe that there is significant overlap in the set of protections which are safe for the company, safe for users, and safe for society in a broad sense. But I don't think any released/deployed AI product focuses on the latter two, just the first one.

Examples include:

Society + Company: Depictions of Muhammad could result in small but historically significant moments of civil strife/discord.

Individual + Company: Accidentally generating NSFW content at work could be harmful to a user. Sometimes your prompt won't seem like it would generate NSFW content, but could be adjacent enough: e.g. "I need some art in the style of a 2000's R&B album cover" (See: Sade - Love Deluxe, Monica - Makings of Me, Rihanna - Unapologetic, Janet Jackson - Damita Jo)

Society + Company: Preventing the product from being used for fraud. e.g. CAPTCHA solving, fraudulent documentation, etc.

Individual + Company: Preventing generation of child porn. In the USA, this would likely be illegal both for the user and for the company.

Their enterprise customers care even more than Stability does.

I think this isn’t software as much as a service. When viewed through this lens the guard rails make more sense.

From George Hotz on Twitter (https://twitter.com/realGeorgeHotz/status/176060391883954211...)

"It's not the models they want to align, it's you."

What specific cases are being prevented by safety controls that you think should be allowed?

Not specifically SD, but DallE: I wanted to get an image of a pure white British shorthair cat on the arm of a brunette middle-aged woman by the balcony door, both looking outside.

It wasn‘t important, just something I saw in the moment and wanted to see what DallE makes of it.

Generation denied. No explanation given, I can only imagine that it triggered some detector of sexual request?

(It wasn‘t the phrase "pure white", as far as I can tell, because I have lots of generated pics of my cat in other contexts)

Well for starters, ChatGPT shouldn't balk at creating something "in Tim Burton's style" just because Tim Burton complained about AI. I guess its fair use unless a select rich person who owns the data complains. Seems like it isn't fair use at all then, just theft from those who cannot legally defend themselves.

Fair use is an exception to copyright. The issue here is that it's not fair use, because copyright simply does not apply. Copyright explicitly does not, has never, and will never protect style.

Didn't Tom Waits successfully sue Frito Lay when the company found an artist that could closely replicate his style and signature voice, who sang a song for a commercial that sounded very Tom Waits-y?

Yes, though explicitly not for copyright infringement. Quoting the court's opinion, "A voice is not copyrightable. The sounds are not 'fixed'." The case was won under the theory of "voice misappropriation", which California case law (Midler v Ford Motor Co) establishes as a violation of the common law right of publicity.

Yes but that was not a copyright or trademark violation. This article explained it to me:

https://grr.com/publications/hey-thats-my-voice-can-i-sue-th...

That makes it even more ridiculous, as that means they are giving rights to rich complaining people that no one has.

Examples: Can you great an image of a cat in Tim Burton's style? Oops! Try another prompt Looks like there are some words that may be automatically blocked at this time. Sometimes even safe content can be blocked by mistake. Check our content policy to see how you can improve your prompt.

Can you create an image of a cat in Wes Anderson's style? Certainly! Wes Anderson’s distinctive style is characterized by meticulous attention to detail, symmetrical compositions, pastel color palettes, and whimsical storytelling. Let’s imagine a feline friend in the world of Wes Anderson...

…in the US. Other countries don't have fair use.

Tell me what they mean by "safety controls" first. It's very vaguely worded.

DALL-E, for example, wrongly denied serveral request of mine.

I don’t feel like it truly matters since they’ll release it and people will happily fine-tune/train all that safety right back out.

It sounds like a reputation/ethics thing to me. You probably don’t want to be known as the company that freely released a model that gleefully provides images of dismembered bodies (or worse).

You are using someone elses propietary technology, you have to deal with their limitations. If you don't like there are endless alternatives.

"Wrongly denied" in this case depends on your point of view, clearly DALL-E didn't want this combination of words created, but you have no right for creation of these prompts.

I'm the last one defending large monolithic corps, but if you go to one and want to be free to do whatever you want you are already starting from a very warped expectation.

As far as Stable Diffusion goes - when the released SD 2.1/XL/Stable Cascade, you couldn't even make a (woman's) nipple.

I don't use them for porn like a lot of people seem too, but it seems weird to me that something that's kind of made to generate art can't generate one of the most common subjects in all of art history - nude humans.

For some reason its training thinks they are decorative, I guess it’s a pretty funny elucidation of how it works.

I have seen a lot of “pasties” that look like Sorry! game pieces, coat buttons, and especially hell-forged cybernetic plumbuses. Did they train it at an alien strip club?

The LoRAs and VAEs work (see civit.ai), but do you really want something named NSFWonly in your pipeline just for nipples? Haha

I’m not sure if they updated them to rectify those “bugs” but you certainly can now.

I seem to have the opposite problem a lot of the time. I tried using Meta's image gen tool, and had such a time trying to get it to make art that was not "kind of" sexual. It felt like Facebook's entire learning chain must have been built on people's sexy images of their girlfriend that's all now hidden in the art.

These were examples that were not super blatant, like a tree landscape that just happens to have a human figure and cave in their crotch. Examples:

https://i.imgur.com/RlH4NNy.jpg - Art is very focused on the monster's crotch

https://i.imgur.com/0M8RZYN.jpg - The comparison should hopefully be obvious

Not meant in a rude way, but please consider that your brain is making these up and you might need to see a therapist. I can see absolutely nothing "kind of sexual" in those two pictures.

Not taken as rude. If its not an issue, then that's actually a positive for you. It means less time taken reloading trying to get it to not look like a human that happens to be made out of mountains.

I have in fact gotten a nude out of Stable Cascade. And that's just with text prompting, the proper way to use these is with multimodal prompting. I'm sure it can do it with an example image.

Parody and pastiche

Generating images of nazis

https://www.theverge.com/2024/2/21/24079371/google-ai-gemini...

Oh the big one would be models weights being released for anyone to use or fine tune themselves.

Sure, the safety people lost that battle for Stable diffusion and LLama. And because they lost, entire industries were created by startups that could now use models themselves, without it being locked behind someone else's AI.

But it wasn't guaranteed to go that way. Maybe the safetyists could have won.

I don't we'd be having our current AI revolution if facebook or SD weren't the first to release models, for anyone to use.

No, it's the cacophony of zealous point scores on X they want to avoid.

We detached this subthread from https://news.ycombinator.com/item?id=39466910.

BTW Nvidia and AMD are baking safety mechanisms into the fucking video drivers

No where is safe

Do you have a reference on this?

Reply on news.ycombinator.com