How Is This Possible??
ฝัง
- เผยแพร่เมื่อ 20 ก.พ. 2024
- OpenAI shows off Sora, it’s new text-to-video model - a massive jump forward for machine-generated video.
Watch the full WAN Show: • 80% of Influencers Hav...
► GET MERCH: lttstore.com
► GET EXCLUSIVE CONTENT ON FLOATPLANE: lmg.gg/lttfloatplane
► SPONSORS, AFFILIATES, AND PARTNERS: lmg.gg/partners
► OUR WAN PODCAST GEAR: lmg.gg/wanset
FOLLOW US ON SOCIAL
---------------------------------------------------
Twitter: / linustech
Facebook: / linustech
Instagram: / linustech
TikTok: / linustech
TikTok (LMG Clips): www.tiktok.com/@_lmgclips_
Twitch: / linustech - วิทยาศาสตร์และเทคโนโลยี
What’s funny is we humans will have to go back to old-school methods. Anytime something important happens, we will have to meet in person. An important business decision needs made? Money needs to go somewhere? Got to meet in person to make sure that there’s no funny business going on.
I hope so! Maybe this all will push more real human interaction. Until cyborgs are convincingly human..
@boscorner then we will sign everything in blood, until the cyborgs get good synthetic blood.
I'm glad I have my typewriter
@@boscorner Please no. I hate people.
@@personzorzhonestly, yes. There is a very real possibility of typewritten notes being mailed to complete transactions.
The scary thing is you can just drop the resolution of these videos and I bet they would look very real to your average person on facebook...
Most of the scam ads that use deepfakes are also (mostly) in low-mid resolution. OpenAI should implement stricter guidelines to prevent any misuse for bad things
@sihamhamda47 How though? Anyone can make a video look super low quality.
to be fair, you dont need a lot to fool the average facebook user
@hmilles I think the thing I'm most concerned about is that for now. On no one will trust any video. Whether it's real or not, you can just call it a deep fake or an AI video and move on.
You are very generous to yourself when assuming that you couldn't be reliably fooled by this
"This is the worst that this tech will ever be"
Assuming no large scale war
Dude. Y’all were looking. Give someone that video who doesn’t know it’s AI and they won’t think twice
yeah the cat one fooled me for a sec even with me knowing but when you know you really quickly start to pick it apart.
this is a tired phrase at this point but, this is the worst it'll ever be. at the pace ai is progressing at this point, i won't be surprised if we get indistinguishable realism within the next year
They should have Sora doing "Will Smith eating spaghetti"
Wait no one suggested Will Smith eating spaghetti when they were taking requests for the new model?!
Probably avoiding recreations of specific likenesses.
Will Smith took it upon himself to recreate it..yesterday
i would go for a slap montage
no. but will smith did a tiktok/yt short eating actual spaghetti. who ever he hired for his meme PR knows what they are doing.
True
The progression of this will make video evidence in criminal prosecutions a lot less powerful
and a lot easier to incriminate the innocent. Prove that it's NOT Ai, and how do you prove something IS Ai?
i love the idea of eating spaghetti to verify that you're real
The problem is that I can show you footage from real movies with these small issues, like the character changing clothes in shot changes or even waves looking off even though they are real.
Show it then
But they don't do it as part of a continous shot, nor do groups puppies merge into each other.
Word calculators can't understand logic, and that will forever be a weakness
@@anivicuno9473 I don’t think you got the idea here. The AI is already producing results better than 90% of FX artists.
and it does that with a single sentence prompt! It’s also advancing in a god like speed. These merging things and issues are some sharp edges that need to be smoothed out. that’s it. in a few year there will be fully AI generated movies and animations. With no real actors. The OpenAI entered public attention only last year. And just look how different everything is from then. Most people won’t realize these video clips are AI generated even right now.
@@henryvasquez4109 Just lookup the Elsa clip of her hair clipping lol
And Frozen was a $1 bil box office movie
@@anivicuno9473 so because there's some teething issues right now it's never going to get any better? Most daft take on this video that shows you generated footage from less than a year ago to what Sora can do.
It's like saying the fact the internet requires me to not use the landline, while the internet is hooked up I can't make or receive calls. That will forever be a weakness. We are in the dial-up age of AI. This is the worst it's ever gonna be.
I've spotted so many lil issues with visual FX made by human artists and their creativity. To shirts being folded in different spots, drinks being full/empty, fuel gauges changing between shots, hair styles, facial hair being shorter or thicker and more. To be honest AI video generation would easily surpass that at the rate it's developing in months. Not to mention that the raw output is easily editable. It will take minimal effort to "regenerate" certain aspects that didn't come out quite right the first time. Don't need to regenerare the entire thing, just highlight the areas that were wonky and with the right prompting or tools, will certainly be fixable with less effort than it takes to film a real shot.
We don't get cities on the moon
We don't get flying cars
We don't get abundant energy
We get this shit.
Idea: CAPTCHA to prove you're human and not an AI robot on video calls:
- eat a bowl of spaghetti
- put your jacket out of frame for at least 1 minute and then put it back on...
Any situation AT ALL that can be automated and assessed via CAPTCHA can itself become training data for a model that can perfectly eat spaghetti and remember its jacket. That's basically how these models train. MODEL one tries to fool model two and model two tries to detect errors in model one.
Remember this is year 2. I work in the visual effects industry and to be honest this is ground breaking tech that is accelerating so fast, that next year could be a the year where we can't tell the difference.
I suspect this will have a "long tail". It won't take long to get it 99% right but fixing the last 1% will be significantly harder.
And the year after that you'll be out of a job sadly.
@@woodytobiasjr.2663 hahaha, not until you get full control over the camera and it can take agency notes. Stock footage will be out of business as well production companies that go out to grab that footage.
Not year 2, but yeah, it's going damn fast.
10:10 It's so hard to not laugh at "his face looks pretty okay", even when you know he's right in context!
I just can't deal with Boy Wonder Linus - can they maybe do the reverse Cavill and use AI to put a digital beard on his face?
Is this the final nail in the coffin for stock footage? It very well might be
most likely is, was there really ever that much money to be made anyways? i mean it seems there was SOO much of it that doing it would never really net you any real income.
@@zachmoyer1849 I think it is/was an industry of very low margins at a very large scale. Perhaps it's not a bad thing to see it go, but it does make me sad that moving forward, most if not all of the footage we see in TH-cam essays or even professionally produced documentaries will probably be fake. Ah well, we live in a mostly fake, plastic world already I suppose
OpenAI announced their partnership with Shutterstock last year
@@zachmoyer1849some people do that as a career, so yes
Years? You mean months?
Im surprised they didn't notice the bun magically appearing in the Tokyo clip
Yeah that was the first thing that popped out to me
This is crazy good for ai generated. The cat one has some plants passing through the cat but the flaws are getting so minor, it gets it right more than it misses on some of these.
The cartoony one at 8:29 had a clipping issue with the candle flame. Doesn't *quite* look right.
@@BikeHelmetMk2 If you told me that animation was done by an animation student, I'd assume the flame was because they're still figuring it out or that it's a working draft and i wouldn't even consider that it was AI generated 😮
14:30 This transition made me realize something.
At some point people will be able to simulate an entire live stream and have it naturally transition into some "investment advice".
It'll be like the stolen account Elon livestream scam on steroids. Maybe even have it interact with chat in a convincing enough way.
It's already happened (kinda). There was a scandal in China where a deepfaked BOARD OF EXECUTIVES convinced an employee to wire a few hundred million in money via a zoom call. Granted, that heist was no doubt very coordinated not just technologically but in terms of social engineering to be able to compromise communication to that degree.
10 years from now, you'll be able to generate your own movies, just by copy-pasting a book into it...
From a book generated by the same AI based on a one sentence prompt. Crazy
Add B-roll content creators to the list of jobs AI will take.
Companies will charge the same, they just won't have to pay so much to produce it. It's up to the consumer now to refuse AI content.
@@noperception5073 consumers would just go directly to sora themselves bypassing the middle man , those companies are in trouble too
your issue is not with technology it is with capitalism
@@jackieAZ Actually neither technology nor capitalism are inherently problematic, it is how people decide to use them.
@@sofaking4779yeah, the problem is the fundamental nature of men, for they will always misuse both
This is crazy and I'm scared
Give it less than 5 years and you'll be able to create a digital version of yourself for family to chat to via a program that functions like a video call when you're gone. It'll be trained on the photos you feed it, and an audio recorded you feed it, and all of the social media posts you've ever made for it to understand your thought processes and speaking style.
Don't forget these are the commercial versions of AI. This is not what they actually have.
i think it is there is so much competition in the space that they are gonna show what they can do so they can garner investment quicker than their rivals.
@@zachmoyer1849definitely. this is an arms race by private companies whose main goal is to secure more investment. They have nothing to hide.
bet the government have a much refined version of this by now if the public are just hearing about it
For this kind of tech? I doubt it. It has been a generally accepted truth that military tech is a few years ahead of civilian tech, but this is the kind of bleeding edge tech that the military might not have an actual advantage on.
Maybe not. Sora isn't publicly available but that might be the commercial version they're showing off. Wouldn't they want the version shown to the public to be the best version currently available in order to entice corporate customers?
It has to be powered by some crazy pipeline that generates 3D objects with textures, puts them into Unreal Engine 5 and estimates the animation.
It’s called training on synthetic data and it’s a huge field yea
The scary thing is, if you use this combined with instagram or facebook, many scoll so fast and don't really look. Add a bit "bad quality" shader over it, and many won't notice shit. This is already bad.
how is that bad? because it ruins people ability to make money?
@@zachmoyer1849 i have seen many of videos of famous people AI videos trying to sell a scam or promote a product, not being able to see who is real or fake will be a mess
@@zachmoyer1849 It makes it easier to spread misinformation, for one.
@@zachmoyer1849 it's bad because conspiracy nutters can create their own reality and ppl will get even worse. we don't need this, it solves nothing
OT: Just looked it up, IPv6 is supported (experimentally) in Linux since 1996, that means it's supported in Linux for longer than I had access to the internet (something like 1999, maybe 1998, under Windows 98).
IMO, the problem with v6 is NAT. NAT was deprecated in favor of address blocks (which is how v4 USED to work), but no carriers actually want to allocate you a full /8 or /16. Mobile networks usually only get /2s and I've even seen this with residential broadband connections. Support for cross-tunneling each protocol is basically non-existent, which hurts adoption as well.
If commercial version is this powerful imagine the actual power of it backend.
imagine what governments will do with this tech.
It would be absolutely hilarious if AI has trouble with people eating and that because of that EVERYONE will be eating when shooting a video.
This is impressive, but keep in mind that rapid advances in tech rarely continue at the same breakneck pace for long. Self-driving cars, quantum computing, nuclear fusion, all of these have at one point seemed like only a few years away from completely revolutionising the world. Don't get me wrong, really interesting work is being done in all these fields. But usually a lot of effort is put in to drumming up hype and making the tech feel imminent when the companies involved are trying to generate investment interest.
It's amazing how fast text to video is improving!!
"Enhance!" made me startle my cat with how loud I laughed.
And 15 years ago they called me crazy and totally out of the times for demanding that our critical ciphers and countersign codebook to be stored offline only (on paper).
"In the future everything will be digital and authentication will be quick, easy and foolproof" .. ..
These days they are looking through the books i wrote at the time, to see what they can implement. Because the digital methods they advocated and deployed could become a security issue.
Most of the digital stuff was created in the name of convenience. Convenience and security are at odds.
I want a version of this that stays in this weird state, It's both impressive and funny at the same time!
Yeah there's something about AI video that I like lol
Couldn’t impress these dudes even if you showed them real aliens. They’d be like “yeah, but the ones in the movies look cooler, so I’m not really that impressed”
In the future film continuity errors will mean points where the ai made a mistake
Shaved Linus is clearly AI
Will Smith actually ate spaghetti weirdly on camera and pretended it was AI generated.
the woman walking through the city really reminded me of the actor who plays Avasarala in The Expanse tbh
With all the little problems you can find with text to video ai videos, it is improving so fast.
The clips of animals were plenty real enough to fool my dog while it was playing on my TV.
IPv6 makes total sense for a multiplanetary civilization where each devices have their own IP on the space Internet that covers the whole galaxy. Until we're there, that tech should be shelved.
This is just scary. This apocalyptic-level scary. A big portion of video making process, film making process will just go in the trash : set designers, costume designers, grip, SFX teams, actors, film crew, story-boards, entire professions are at risks. Yeah, it's revolutionnary, but in a bad way revolutionnary.
Mark my word on this : in the next months, we will see more and more TH-cams channels popping out, with video made out ot nothingness, soulless videos, because it would cost nothing at all to enter prompts a leave Sora do the work. The future is grimm. Very grimm.
Only way we can tell if the president actually said it is if they happen to be eating spaghetti in the press conference
we will have to do what should have been done a long time ago the government handles streams of its own so basically if its not from whatever.gov its not real/been edited/messed with. Because news outlets do that already and its bullshit.
Even the red dress is changed at the end
Does any of you understand how powerful this AI can be for graphics in games? Like, you just render the 3D world's and objects' geometry and then you just ask the AI to "paint" the light on them, meaning you do not need as much as power for rasterization and you lean more into NPUs which also clearly, is more realistic, this is the feature, like it or not.
The woman's hair in the Tokyo video also changes. It's completely down at first, and when you see it again she has some kind of bun going on
it might also be rendering in layers to help with consistency
Text+video to video is also crazy good
Just waiting for someone to sue the fuck out of them for having unlicensed data
The Japanese woman also has a bun appear on her head when the new jacket appears.
AI has taught me that will smith really loves his spagetti
Fun fact: Apart from collabs and tech upgrades there hasn't been a mainline LTT video that wasn't hosted by someone else other than Linus for over a year since last years CES videos...
I cant wait for next year when we get text to box office hit movies "make me a new iron man movie that incorporates the transformers series and godzilla"
ah great so here's gonna be the technology I can't tell from reality that my kids will go "what you can't tell it's AI?"
they have a example of a man eating a burger. so they have already tried consuming food test
Well shit, time to move to the monastery to escape this for as long as possible
the not so subtle reference to Jane from Firefly with the knitted helmet dude.
I think the bird one towards the end and the golden retriever puppies could both fool me if I didn’t have the pretext. I can find little bits to nitpick if I try really hard but if nobody told me before I watched it then I would scroll right by those on instagram
Sora doesn’t “render” more than it shows you rather it tracks concepts in space. So encoded into this idea of a woman wearing a jacket was that it had a certain type of straps & a certain kind of flaps of a certain shape. These are all just concepts as weights on a graph for-each object though so it can recreate what it thinks should be there when it comes back to into frame. It takes a similar path through this graph to arrive at rendering the second leather jacket.
Part of the reason they are training it on examples of physics interactions instead of having it do the actual calcs. The chameleon should be able to break the branch off screen changing the map of concepts associated with both those objects in Sora 2.0
I'm guessing the reason it's so good at holding static textures, but sometimes fucks up with moving ones is that it artificially generates 3d scenes in a rendering engine, like unreal, probably similarly to Dall-E, then picks a perspective to view the scene from, the same way a video game would. if I'm right, that element of development was probably comparatively trivial. ensuring the aesthetic consistency of every frame of a dynamic texture, on the other hand?
I know this is wrong, but I think Will Smith Spaghetti had to come from someone with a sense of humour
It’s gonna be crazy when people can run this on ther own PC
4:35.Her hair in the second shot changed too to be fair, but it's nothing you could not correct with an inpaint equivalent.
Soon you will give it Star Wars Episode 1 - 3 with a text list of changes and it will output a new Episode 1 - 3 but perfect.
Only a matter of time until you can give this video to AI and tell it to replace Linus with Will Smith eating Spaghetti and it's gonna look convincing. This is scary.
I make small web apps sometimes for private use and I almost always only use ipv4. Sometimes I use an ipv6 too since almost all hosts give you an ipv6 too but most of the time I don’t even bother. I probably should
There's a railway one with reflection that is scary realistic
For some reason I miss configured my router once and disabled ipv4 a few days ago. Half of the websites were not opening. And also, connecting to a server with its ipv6 and port always goes to Google search.
It's funny how far we've come that tech guys are like "ahh yeah here is an issue and there it doesn't look real" when this leap is HUUUGE in such a short time. I mean AI wasn't even a thing (you know what I mean) not too long ago and now we're at this. Imagine what it will look like in 2025/26 and especially in 2030!!! Everything moves so fast that nothing really feels too crazy anymore.
This stuff is deeply terrifying.
I think we will finally reach the age of UFO videos no longer being potato quality. We’re going to finally be seeing some good quality ones
Dan is broken by spaghetti.
No one's talking about using this AI to render video games LIVE in the future and the combination of this with VR or Nuralink
Give it a month and an equivalent will be available in Stable Diffusion where people have far more control over it and no limits enforced by OpenAI dor ethical reasons.
As another side-note, this could potentially do massive damage to stock footage companies. I dont know how much AI Imagery has harmed shutterstock, so i could be wrong, bht the ability to generate a photoreal insert shot for your advertisement or whatever without spending any money could be very tempting.
Open AI did a deal with Shutter stock I believe
Luke, the chameleon isn't "okay", it's completely inaccurate to how any of the species looks or moves, these things are nightmare fuel for biologists because, in addition to fighting general human stupidity, they now have to worry about realistic videos of nonexistent species online.
Look man, if you have to be a biologist to notice it's wrong, that means it's pretty good
The video it produces is slightly dream-like with some of the changes it makes and such.
Every time new ai models come out we all go “it’s good but it has problems” then those problems are completely or mostly fixed within a year.
I’m not doubting anything anymore. I have no idea how good this thing will get in just 6 months.
You see how quick Linus moves on when he realizes he's wrong on IPV4 prices having a downward trend. Chilling insight into working with him...
I mean, the downtrend in recent years is not even that huge and that level of volatility is completely normal in investment world (just like russell 2000). The price literally more than doubled in a few years when it is a commodity and you can even lease the address to get more return.
The jacked turned into a trench coat. Gained almost 1 meter in length making it possible for her to place her hands in the pockets.
Moore’s law goes brr
Really wished they saw the video from a train where their reflections show up
I wonder what this all will mean for evidence in lawsuites. From my perspective any kind of digital evidence will become completely worthless.
That poor hungr guy. Look what AI did to his eyebrows
So great to see a beardless host.
Linus, take some psychedelics and you mind understand the science behind Sora and pretty much all these generators. From a data scientist to a layman. It is seriously eye opening 😂
05:39 Were you paying attention Neo? Or were you looking at the woman in the red dress?
The final test will be making the same Will Smith video but with Sora 😅
cant wait to generate my own LMG clips... get it yet?
With 1 minute clips, I want to see someone put together like an hour or two movie. Just leave all the object and concept permanence issues there, people are going to have a wild time laughing at them. 😂
Shaven Linus?
Shinus! Shines bright, like his future!
Finally
Oh boy
Holly cou! Linus looks as if he had no beard or mustache! This AI is impressive!!
Being obsessed with the sign going behind her head is a red hearing. Just got to make the model treat background and the object separately. So it parses the location out of your prompt then processes out what's going on. Renders the background then puts the women over it. This is also seen in the example of the couple walking in a street market. Its looks like people walking in front of a blue screen with the background being done in post. At one point they are practically walking on the roof and the scale and camera movements are synced. This is how these system improve they come up with work arounds like separating the background and the subject. Like they aren't gonna fix hands by giving it more data to learn what a hand is they are just going to develop a system to determine when a hand is being drawn and have it work down a logically separate path to make sure hands are drawn right.
What about the dog on two beaches and the bird that disappears? The way you described it is definitely a way it could be done but I don't know if it's what they did here
Even with all its current flaws if I was able to go "Give me Invincible season 3 now!" I'd be happy as can be. And it's brand new!
It’s also far more important for people (who aren’t making a living off of content creation) to consider whether it’s worth the potential consequences of uploading their voice and likeness to the internet
I own a /40 IPv6 block and love it, every one of my clients gets a /64 included with their services.
I would love to own a v4 block but at the cost of $60+/IP, its too expensive even when transferring from other regions.
Yeah. I'm looking forward to the exact use case everyone is thinking about without spelling it out
The will smith face is Sid the sloth from ice age
About ipv4 and IPv6, ipv4 is easier to use on a lot of plane but the main reason is because of the limitations of ipv4 they changed a bunch of things about how it works so it ends up being easier on a lot of planes but some get harder but most people are now used to how things work for ipv4 now, so IPv6 works the way ipv4 worked before but people don't know anymore so until IPv6 cannot be avoided, people won't switch to it