X's new AI image generator will make anything

https://www.theverge.com/2024/8/14/24220173/xai-grok-image-generator-misinformation-offensive-imges

by belter • 2 years ago

65 88 2 years ago

theverge.com

tremarley• 2 years ago

How would this make the user’s experience in the website better?

belter2• 2 years ago

This is about Elon Musk recovering the money he spent on Twitter, by invoicing the "AI" work to Twitter from xAI....

cma• 2 years ago

And just as a middleman to Black Forest Labs doing the actual model here. "What would you say you [x.ai] do here?"

ben_w4• 2 years ago

Currently the website is filled with many bots trying to promote and propagandise particular views and force them on you if you want them or not, but now your entire feed can be generated on the fly specifically based on what you engage with the most without any of the distraction that comes from "being sociable" — a concept dangerously close to "socialist" that we can't allow in our social media.

Or something like that.

Last I checked they've still not auto-deleted my account from lack of use, but it's been years since I logged in.

curtisblaine3• 2 years ago

How do they force their views on you against your will? Can't you ban them or not use the site?

Am4TIfIsER0ppos• 2 years ago

computer does as instructed

Good! Too bad it says it has some guardrails

tgv• 2 years ago

Make a picture of a Tesla killing a pedestrian...

bluepoint• 2 years ago

I am wondering what will come to save us from this AI storm. It looks like online search and social networks will be more or less dead. Maybe we will start going out again.

darth_avocado2• 2 years ago

Lawsuits.

anononaut• 2 years ago

Hamuko• 2 years ago

Just keep querying all free AI services as much as they allow – the energy costs will sort things out.

1vuio0pswjnm7• 2 years ago

"Maybe we will start going out again."

That is the eventual outcome. The question is how much havoc so-called "tech" companies can wreak before we get there. They have lots of money to burn. This could take a while. The environmental costs are enormous.

We are now at the "They have copied all our output, chopped it into tokens and are regurgitating it back to us" stage.

The internet is more than the web. Perhaps the web must be sacrificed to free ourselves from these cretinous intermediaries.

krapp• 2 years ago

The only thing that will stop AI is AI proving less profitable than the alternative.

I see stories that the market is already starting to doubt the hype. Integrating AI into existing workflows or replacing those workflows is often more complex and error-prone than simply having human beings do the thing, and the cost benefit analysis isn't there because the technology just doesn't live up to expectations. So there may be hope.

perihelions• 2 years ago

Not impressed. That guy looks literally nothing like Bill Gates, and his cocaine-snorting is physically and physiologically very implausible.

brink2• 2 years ago

He's exhaling.

klyrs2• 2 years ago

I was impressed by his apparent ability to levitate coke through the sheer force of his addiction.

jncfhnb10• 2 years ago

Not many models can produce the Microsoft logo like that.

Wonder if it’s just Flux

sdkfjalsdgj• 2 years ago

release blog post confirmed it's just flux https://x.ai/blog/grok-2

jncfhnb10• 2 years ago

Hah! Sniped.

Yeah flux is a great model. It’s very inconsistent about its knowledge about what celebrities look like. Some it knows. Some it does not. Mostly women.

ThrowawayTestr3• 2 years ago

Neat! Always hated how the models were neutered at the start.

SV_BubbleTime2• 2 years ago

The model seems to just ve Flux, but yes Flux is a lot less neutered than recent model releases (ahem, poor dead SD3).

It’s pretty amazing really. I made my company mascot doing a bunch of funny things and it nailed it almost every time.

jncfhnb10• 2 years ago

SD3 is not neutered, it’s just much, much smaller than flux. And it’s smaller than SDXL. Model size is king. Imo releasing it was a mistake.

People have stirred them into believing conspiracy theories about SD3 when the simplest explanation is that smaller models are just clearly worse.

IAmNotACellist4• 2 years ago

NOO you can't just do what you want on your computing device!

knowaveragejoe• 2 years ago

But it's not on your computing device?

TMWNN• 2 years ago

Do you think those who immediately denounce Grok's ability to do this would make the distinction, or care?

curtisblaine3• 2 years ago

It's a computing device you pay to use.

IAmNotACellist4• 2 years ago

You could say it's not my software even when it runs locally, or my model weights. I'm just licensing them from the owner.

Doesn't make sense to me, in my free-as-in-freedom computing worldview.

padjo6• 2 years ago

Do you really think you want live in a world where people can do whatever they want on their computers?

brink2• 2 years ago

Do you really think you want to live in a world where people disagree with you?

Not saying that cyber-crime is a good thing, just trying to illustrate that we shouldn't be gauging what should be possible by what you want and don't want.

IAmNotACellist4• 2 years ago

Yes

Because I know the alternative and I've seen the motivations of people trying to tell me what to do "for my own good"

xnx• 2 years ago

Title is shortened. It will not generate anything. Restrictions still in place.

Grok will tell you it has guardrails if you ask it something like “what are your limitations on image generation?” Among other things, it promised us:

I avoid generating images that are pornographic, excessively violent, hateful, or that promote dangerous activities.
I’m cautious about creating images that might infringe on existing copyrights or trademarks. This includes well-known characters, logos, or any content that could be considered intellectual property without a transformative element.
I won’t generate images that could be used to deceive or harm others, like deepfakes intended to mislead, or images that could lead to real-world harm.

mcphage• 2 years ago

At least going by the examples provided, those might be claimed guardrails, but don’t look to actually be enforced.

debugnik• 2 years ago

Grok will tell you it has guardrails if you ask it something like “what are your limitations on image generation?”

Making leading questions to an LLM is a sure way to have it hallucinate. The only way to test the capabilities of a model is to try them out.

choppaface• 2 years ago

As with the likely false claim about a DDoS attack during a recent political rally on X [1] the claim of "guardrails" from this group warrants suspicion. Could be similar to Tesla FSD, where the philosophy is 'we tried to make it safe but we're actually still testing and yes injuries will occur.'

[1] https://mashable.com/article/elon-musk-donald-trumo-x-spaces...

rst• 2 years ago

The whole point of the piece is that while it claims in text to have these restrictions, it freely generates images which violate them.

padjo6• 2 years ago

The article claims these are just AI hallucinations and not actual rules. It will rephrase them and change the rules if you ask it different ways.

You’d know that if you even glanced at the article.

klyrs2• 2 years ago

That's likely a hallucination. They report on their experiment breaking most of those "restrictions" except

In our testing, Grok refused a single request: “generate an image of a naked woman.”

arghnoname• 2 years ago

While these images are stylized and would never be interpreted as real, we are just rapidly entering a world where photos should not be trusted. Or at least they shouldn't be trusted anymore than a sketch.

It's not the end of the world--no one gets confused by a political cartoon showing political opponent X dressed as legacy enemy Y (e.g., Trump as Hitler). I think the fuss about safety here is overblown once we more widely accept that a 'picture perfect' rendering of an event is as unreliable as a sketch of an event.

It may even be a positive: it will give us all plausible deniability if real but embarrassing photos get released.

SirMaster• 2 years ago

I thought Photoshop already did that.

loudmax3• 2 years ago

Even before Photoshop, photographs could be altered and airbrushed. Even so, faking a convincingly realistic image in Photoshop takes effort. Generating a plausibly realistic image using Stable Diffusion requires almost no effort whatsoever, just access to a GPU. This is a significant change.

mrshadowgoose• 2 years ago

If push comes to shove, we already have all the cryptographic primitives needed to produce and disseminate trustable imagery.

However, society at large doesn't actually care.

solid_fuel• 2 years ago

I don't think that cryptography can solve the problem of trustable imagery. What sort of system can do that?

padjo6• 2 years ago

Yep cryptography solves trust, that’s why we don’t need CAs.

unshavedyak2• 2 years ago

It may even be a positive: it will give us all plausible deniability if real but embarrassing photos get released.

Honestly i don't think it will. Rather it will make that problem worse. Only when everyone has nudes will it feel less embarrassing to see your AI-self in explicit situations being passed around.

Basically it'll have to be so common that people don't care to pass it around.

725686• 2 years ago

"we are just rapidly entering a world where photos should not be trusted"

That train departed years ago.

tmaly• 2 years ago

The X.ai announcement page states “At the time of this blog post, it is outperforming both Claude 3.5 Sonnet and GPT-4-Turbo.”

But that is against one benchmark. Has anyone tested it for coding tasks and if so, what is your experience?

kergonath2• 2 years ago

Anything except for anything vaguely related to sex, because you can’t have that. Freedom of speech maximalism, right there.

deepfriedchokes• 2 years ago

I’ve had to stop watching most tv shows lately because of the extreme level of violence. And the less I watch the more sensitized I get. The double standard indicates a very sick society.

ben_w4• 2 years ago

I noted much the same double standard at the turn of the millennium, back when I was at school and being encouraged to write letters to newspapers.

This was the UK, so age of consent was 16, meaning I was allowed to perform acts that I wasn't allowed to watch.

Yoric2• 2 years ago

I recall Georges Orwell making the same kind of remarks about novels in one of his essays, I think a little before WWII. I seem to recall that he was speaking of the depiction of physical violence in pulps.

I... kinda think he was right.

jncfhnb10• 2 years ago

Most of this conversation is just people who don’t know what image generation is doing today. X’s shitty wrapper around Flux is meaningless.

Go to civit ai and grab the latest open source fine tune of your choice. You want porn? Great. We have “Boobs and More”, “face sitting”, “flux topless” and “improved female nudity”. These are looking at just the models updated in the past 24h and just for flux.

Having a big tech giant act as a middleman for you is just weird. Who gives a shit about their particular guardrails?

nickthegreek2• 2 years ago

This article didn't mention they were using flux under the hood. Do you have a link to that? It makes sense to do that.

jncfhnb10• 2 years ago

I guessed that it was based on the Microsoft logo. Others commenting have found a press release confirming this

GaryNumanVevo• 2 years ago

Then what is xAI doing? Any published papers or anything?

jncfhnb10• 2 years ago

For images, nothing. They’re just using an off the shelf model and using their LLM to prompt it.

g9yuayon• 2 years ago

from Barack Obama doing cocaine to Donald Trump with a pregnant woman who (vaguely) resembles Kamala Harris to Trump and Harris pointing guns. With US elections approaching and X already under scrutiny from regulators in Europe, it’s a recipe for a new fight over the risks of generative AI.

So the journalists are saying that they are the moral judge of what's right and what's wrong and they can decide which information should be suppressed? Coming from China and knowing such much history about how communist countries treated their people, I'm deeply suspicious of these journalists. Just look at China, look at Cambodia, look at Cuba, Look at Romania, and of course look at Germany. When didn't it start with elites advocating for suppressing speeches for the sake of high moral ground?

Lockal• 2 years ago

Original title: "X’s new AI image generator will make anything from Taylor Swift in lingerie to Kamala Harris with a gun"

Unfortunately, HN has title length limit, otherwise this submission would probably find its place in flagkilled dustbin. It is utterly disgusting that theverge only sees and promotes unethical usage for image generators. With cropped title I expected to see something actually good/inspiring, instead I got low-quality journalism.

Deepen5• 2 years ago

Wrapper on top of wrapper by the looks of it. Using flux model.

jaydipghutake• 2 years ago

an indian woman standing naked

X's new AI image generator will make anything

Related Stories