Ishad
Ahn'Qiraj Raider
- 4,784
- 4,697
Writers Guild Strike 2023
- Thread starter Goatface
- Start date
Gravel
Mr. Poopybutthole
- 36,317
- 115,198
I just wanted to bump this thread because this video was in my recommended. Not that this video is special in particular, but he does give good examples of where AI has come in a year.
Despite working in tech, we have people like Mist who just for some reason can't comprehend exponential growth. I pointed out that yeah, maybe the writing sucks now, but it was only like 6 months in at that point. In the below video, he gives examples of how a year ago we had shitty Will Smith AI generated videos and now where we're at. Give it another year and humans will be obsolete.
- 1
Tuco
I got Tuco'd!
- 45,417
- 73,486
I don't dispute that it's hard to predict where AI videos or writing will go, but a video on how AI-created videos have evolved with Sora isn't a very compelling argument against how LLMs will have diminishing returns on producing high quality writing.
What old nerds like us have seen repeatedly is that technology capability often follows a sigmoid function
Where some new innovation causes exponential growth in capability, followed by a logarithmic decrease as that capability reaches effective limits dictated by physics, mathematics, etc. The current approach with LLMs is *somewhere* in that exponential growth phase, and predicting when it'll start to taper off for a given approach (in this case, writing scripts or screenplays) is foolish.
Personally I think LLMs will be very well suited to creating good writing on a micro-level for a variety of reasons, including how organized screenplay data already is, how quickly humans can review an LLM-generated line and how quickly LLMs will be able to perform semi-random association, which is a core component to humor.
I still have fun now and then creating AI-art and AI-stories with my 9 year old. He's growing up with those tools and I think his writing will be improved by it as it evolves. I don't mind being the fool that says AI (via LLMs or another approach) will be an essential companion to writers as their relative importance and headcount in creating TV shows and movies diminishes.
Last edited:
Mist
Eeyore Enthusiast
- 30,394
- 22,167
There aren't going to be exponentially more GPUs, exponentially more wattage available to power then, nor exponentially larger training sets.
There's no intelligence trick at work here, this is all done with brute-force compute.
Like yes, I can see some odd scenarios where we end up turning the entire world into datacenter to build a few big brains just to remix all of our old videos into crappy new movies instead of just making the movies ourselves? But that's almost farcical.
Mist
Eeyore Enthusiast
- 30,394
- 22,167
The biggest question is: why do we need this and what problem does it solve?
A giant machine that turns out infinite crappy entertainment is what we already have: Hollywood and the video game industry.
Even in the 'ideal' scenario where we end up in a world in which everyone gets their own custom cultural artifacts algorithmically created just for them, that's actually dystopian. The value of art and culture is giving people shared context for understanding reality and each other.
So ultimately, I think this technology will use a LOT of energy and hardware to create something that is actually bad. I believe we're hurtling headlong towards an age of inhumanity and this is just the bleeding edge of it.
A giant machine that turns out infinite crappy entertainment is what we already have: Hollywood and the video game industry.
Even in the 'ideal' scenario where we end up in a world in which everyone gets their own custom cultural artifacts algorithmically created just for them, that's actually dystopian. The value of art and culture is giving people shared context for understanding reality and each other.
So ultimately, I think this technology will use a LOT of energy and hardware to create something that is actually bad. I believe we're hurtling headlong towards an age of inhumanity and this is just the bleeding edge of it.
- 1
Tuco
I got Tuco'd!
- 45,417
- 73,486
Eh, not really. That's like, the 10th biggest question here at best.
Biggest question is probably, "Can I make $$$ by selling this?". And the answer is yes.
#8 biggest question might be, "Does this decrease the barrier to entry for creating this kind of content?" and the answer is most definitely yes. The tip of this spear in this domain is detailed 2d art, where many people are able to participate in creating illustrations who would otherwise never dream of it.
- 1
Mist
Eeyore Enthusiast
- 30,394
- 22,167
Remains to be seen. Without VC cash, none of these things are remotely profitable. The compute is way too expensive. Compute, and the energy it requires, would have to get a LOT cheaper and Moore's Law is pretty fuckin' dead at this point.
Kriptini
Vyemm Raider
- 3,640
- 3,538
Sora is going to single-handedly kill the stock footage industry. I was on a Shark Week show last year and we had to pull a lot of stock footage of specific species of sharks doing specific things. There is not a lot of that kind of footage out there, and from what there is, it is extremely expensive. We spent thousands and thousands of dollars on stock footage for one show. Now imagine if we could generate any stock footage we need for any of our shows with a single Sora subscription. That's a game-changer.
- 2
Mist
Eeyore Enthusiast
- 30,394
- 22,167
Kriptini
Vyemm Raider
- 3,640
- 3,538
AP will still exist and they will still be sending photographers to cover stories and events for news articles. For all of the after-the-fact for-profit documentaries, I guarantee you that if they have the ability to save money by generating something that looks just enough like the real event (even if it wasn't the real event), they will.
CaughtCross
Vyemm Raider
- 2,705
- 4,223
This does sound like game changer. Kinda like how drones made getting aerial shots way cheaper vs having to use a helicopter. Soon you can get all kinds of bits of footage created with AI to fill for all sorts of shots.
- 2
Mist
Eeyore Enthusiast
- 30,394
- 22,167
Will spamming prompts end up being cheaper than the already-very-good CGI tools we already have?
A whole lot of prompt-based, generative AI is just a toy in search of a problem. Prompting isn't stable or predictable enough for enterprise applications, this is obvious to anyone trying to build with it in the contact center industry right now.
Like yes, there will likely end up being a lot of subscription tier-based "creative companion" features built into office and art apps. But I think people are underestimating how hard it is to integrate generative AI into typical tools, or how computationally expensive it is to just keep spamming prompts until you get something you like, which you'd have to do with video.
"It'll just keep getting better, magically, on its own" is a total myth based on the anthropomorphizing of these tools. They're not big brains that keep getting smarter, they're static models that slowly get fine-tuned for specific applications, at a high compute cost.
People have just lost all sense of skepticism when it comes to these things.
Kriptini
Vyemm Raider
- 3,640
- 3,538
Yes, spamming prompts is FAR cheaper than CGI tools. Nobody is using CGI for stock footage. That's extremely cost-prohibitive considering how expensive it is to higher the artists (you need a modeler, texturer, lighting artist, and compositor at minimum, and most of the time those people can't work together without a coordinator) vs. having a "junior archival producer" (intern) spending an afternoon throwing prompts into Sora.
The enterprise subscription could cost $10,000 a month and networks like Discovery would still save money on it compared to what they spend on stock footage.
- 1
spronk
FPS noob
- 22,595
- 25,646
i didn't realize the latest round of sora prompt videos looks so crazy good. its wild how much its advanced in just 1 year, i'm starting to believe the nvidia CEO when he says its pointless investing a ton of money into making chips better and better because in the next 10 years AI itself will start designing chips better than humans can.
Sora: Creating video from text
openai.com
- 1
Mist
Eeyore Enthusiast
- 30,394
- 22,167
i didn't realize the latest round of sora prompt videos looks so crazy good. its wild how much its advanced in just 1 year, i'm starting to believe the nvidia CEO when he says its pointless investing a ton of money into making chips better and better because in the next 10 years AI itself will start designing chips better than humans can.
Sora: Creating video from text
openai.com
The closer you look the weirder it gets.
This one is just weird as hell:
Gravel
Mr. Poopybutthole
- 36,317
- 115,198
Don't worry,i didn't realize the latest round of sora prompt videos looks so crazy good. its wild how much its advanced in just 1 year, i'm starting to believe the nvidia CEO when he says its pointless investing a ton of money into making chips better and better because in the next 10 years AI itself will start designing chips better than humans can.
Sora: Creating video from text
openai.com
Mist
has assured us this is all a giant waste of time.He's really doubling down on this. It's going to be like Robert Reich and his "internet is just a passing fad" quote.
Mist
Eeyore Enthusiast
- 30,394
- 22,167
It's mostly just copying and splicing shit dude. It's all a grift.Don't worry,
He's really doubling down on this. It's going to be like Robert Reich and his "internet is just a passing fad" quote.
Will parts of this tech slowly make its way into other products? Sure. Will we all be prompt engineering everything 10 years from now? Fucking no.
Spicy autocomplete is a not a panacea. But yes, marketing companies will abuse the shit out of this stuff for a while, because even when it looks weird, looking for the weirdness captures your attention, at least until you get bored of it.
Mist
Eeyore Enthusiast
- 30,394
- 22,167
Video generation models as world simulators
We explore large-scale training of generative models on video data. Specifically, we train text-conditional diffusion models jointly on videos and images of variable durations, resolutions and aspect ratios. We leverage a transformer architecture that operates on spacetime patches of video and...
openai.com
Here's the full "paper." Really, marketing material.
A lot of this is very impressive. It doesn't say how many city blocks worth of electricity it took to generate some of these, but I can definitely see this tech being built into existing video editing tools with a very pricey license and a lot of costly cloud compute.
People are going to make some awesome ads with this stuff, but I block all those so who cares?
LiquidDeath
Magnus Deadlift the Fucktiger
- 4,889
- 11,292
Jesus Christ, your mewling about power consumption is so fucking retarded.Video generation models as world simulators
We explore large-scale training of generative models on video data. Specifically, we train text-conditional diffusion models jointly on videos and images of variable durations, resolutions and aspect ratios. We leverage a transformer architecture that operates on spacetime patches of video and...openai.com
Here's the full "paper." Really, marketing material.
A lot of this is very impressive. It doesn't say how many city blocks worth of electricity it took to generate some of these, but I can definitely see this tech being built into existing video editing tools with a very pricey license and a lot of costly cloud compute.
People are going to make some awesome ads with this stuff, but I block all those so who cares?
My genuine hope is that these endeavors suck so much power out of the grid that it causes rolling blackouts (in shithole states like California) and forces the hands of everyone involved to the decades-late realization that we never had a power generation crisis, only a crisis of faggots unwilling to embrace nuclear power as the actual, literal answer to all our energy problems.
- 1