OpenAI collapses media reality with Sora, a photorealistic AI video generator

Hello, cultural singularity—soon, every video you see online could be completely fake.

See full article...

 

[marsilies]

I had some fun yesterday pointing out all the flaws in the videos, and pretty much every video has some sort of flaw in it. It's really bad at handling crowds, for example, so if you follow a particular person/animal/item in a scene, they may disappear, or morph, or duplicate, etc.

 

[marsilies]

It must have an impact on the stock footage industry. I'd MUCH rather tell an AI what I want than spend time searching through video archives.

Too bad they didn't develop an AI that could better find applicable existing stock footage from a text prompt instead.

 

[marsilies]

Pornhub is heavily investing. "What's your fetish? What's your kink? We don't judge, not at our $9.99/mo tier we don't." Not to mention no more concerns around consent or revenge porn posts.

Revenge porn is still going to be an issue, and probable worse, if the prompt can be specific enough to create a doppelganger of the victim. Like with deepfakes, it will no longer be limited to actual acts the victim performed, but unlike deepfakes, it doesn't even have to be an act anyone has performed before, and there won't be a reference video to locate and expose the manipulation.

 

[marsilies]

It searches for videos that contain the same keywords as the text.
It does not generate video from text. It is an offensively easy to prove lie.

No, check the videos, there's always some weirdness in them that wouldn't be in a video shot in reality, or even CGI rendered.

No idea if it's actually as easy to create as using the example prompts they give, or if there's a lot more specificity and tweaking they're not showing publicly, but this definitely looks like diffusion-generated video.

 

[marsilies]

Your argument is as old as fiction. The same arguments were discussed by critics of the first novels who claimed that normal people couldn't be trusted to tell the difference between a textual retellings of real history and fictional history.

Talking about making things up, I'm going to need to see some evidence that this was a serious concern about the first novels, and not completely made up nonsense.

This doesn't mention anything like what you suggest:
https://victorianweb.org/genre/skilton.html
Also, what could be considered the first novel is far older than you may think, like 1st Century Greece:
https://bookriot.com/the-worlds-first-novel/

 

[marsilies]

The Victorians (1820-1914) had already figured it out. I was thinking of Don Quixote (1605) which is itself a satirical response to critics of fiction. There's also Dante's Inferno (1314 and not a novel), where there are reports that some of villagers in Dante's town believed the story of his descent into hell was autobiographical.

That's not support for your argument. I need citations, i.e. links to sources, and not apocryphal stories.

There's nothing in the WIkipedia for either work that backs up your claims.

https://en.wikipedia.org/wiki/Inferno_(Dante)https://en.wikipedia.org/wiki/Don_Quixote

 

[marsilies]

How about Plato's critiques of poetry as mimesis in the Republic? Does that work for you?

No. Have you completely forgotten what your argument was? Because that doesn't support it at all.

Wikipedia is a weak primary source kid.

And yet you've supplied fuck all as a source for your fake claims.

 

[marsilies]

When you're reaching for Plato critiquing poetry to try and explain why it's no big deal that people can create video of anything with a prompt I think you're beyond reaching and should probably just let it go.

This isn't a poem, or a novel, or a hoe, or desktop publishing. There's just nothing of real value to be found in trying to compare AI to old tools, there is no precedent in human history.

I prefer to just discuss it for what it is.

The thing is, people always knew that spoken and written words could be untrue. We've always made stuff up, as a species. It can sometimes be true, sure, but there's nothing inherent about a person saying or writing something that meant it was true. It's the same with music we make, or visual art we make, whether sculpture or painting. It's inherently a creation of a person.

The issue is when it gets to the modern age, and the concept of recording something by an objective device, whether light, sound, or both. An audio recording is presumed to be a recording of a real thing, something people actually said or performed. A photograph is presumed to a recording of a real scene, something that actually happened in front of a camera, even if staged. And motion film/video is thought to be both. It's not just someone writing down in a book, or painting something after the fact (if in fact at all), but a recording of the actual moment. And it looks/sounds real enough to us to be analogous to actually being there ourselves.

Of course, recordings can be manipulated. But at the least, that meant some real image had to initially exist to be manipulated, and/or effort had to be performed to do the manipulation. But now, the AI has gotten to the point where it can generate whole cloth out of random noise images and motion that, at least at first glance, look as realistic as something real captured on video. And we don't have the history of understanding something that looks so real being completely fake in that way. So it's a paradigm shift in a way that never occurred with the written word, because we always knew words could be completely fake, but many of us have lived for decades in a world where photos and video couldn't be so easily and convincingly faked.

 

[marsilies]

Movies and video games can have the originals tracked down and disseminated to debunk the narrative it's being misused for. Different locations can be detected by video analysis to confirm the location and/or tracking down the original photo or video. This technology is different. If it's not already possible, it won't be long until you can fabricate video of completely manufactured events in very real locations and even match up time-related differences in the scene to make fakes that can only be debunked by interviewing the people in the area at the specific time the video is claimed to be from. And that's often going to be a difficult to impossible task, and might not do any good anyway depending on the trustworthiness (or existence) of those interviewed.

And why should I trust your debunking video anyway? Maybe the original was real, and you generated a fake video to "debunk" it, with fake interviews of fake people.

 

[marsilies]

AI can't even do math, something computers have been good at since computers first existed. Let's not get ahead of ourselves with how smart this shit is.

Ahem...

I think Aurich's statement should be revised to: "LLM's can't even do basic math reliably and consistently, something computers have been good at since computers first existed."

Note that the second link states that the AI did "at almost the same level as a human gold medalist." Humans are fallible and not as consistent and reliable as a simple calculator, and the AI did worse than that.

As for the other article, the advantage with the LLM was that it was able to produce attempted solutions quickly, not that it was inherently that good at it. They needed another program to evaluate the solutions the LLM produced.
https://www.scientificamerican.com/article/ai-beats-humans-on-unsolved-math-problem/

FunSearch automatically creates requests for a specially trained LLM, asking it to write short computer programs that can generate solutions to a particular mathematical problem. The system then checks quickly to see whether those solutions are better than known ones. If not, it provides feedback to the LLM so that it can improve at the next round.

“The way we use the LLM is as a creativity engine,” says DeepMind computer scientist Bernardino Romera-Paredes. Not all programs that the LLM generates are useful, and some are so incorrect that they wouldn’t even be able to run, he says. But another program can quickly toss the incorrect ones away and test the output of the correct ones.

 

[marsilies]

LLM's aren't the only DL models.

They were the only models used in the articles you cited. Do you have another article?

Competing at a level comparable to a Human Gold Medalist, is way beyond a simple calculator. Here is a sample problem from that competition:

Will your calculator solve that for you?

I don't really pretend to be good at math, but then I'm not touting an LLM pretending to be good at math either.

 

[marsilies]

I think the cynical-but-true take is that AGI is as much a marketing term as it is a goal for true believers, and there's a lot of pressure to move the goal posts to be able to announce that you've achieved it because you can turn on a money faucet with the news.

It's basically the new Turing Test, since that's now achievable. Before the latest LLMs, there'd always be someone claiming their AI "passed" the Turing Test because it fooled some non-scientists into thinking it was a barely literate 13-year-old.
https://www.theguardian.com/technol...-simulates-13-year-old-boy-passes-turing-test