OpenAI's Sora AI Video Model Delays (00:00-03:41) Meta's AI Voice Integration (07:19-10:17) AI Video Speed Comparison (03:41-05:32) Meta’s Focus on AI Translation (09:14-11:30) AI in Future Mobile Devices (07:52-08:19)
I use flux to make video, it works fine, although it codts about 10 cents of dollar every five seconds, which is very little if you live in a first world country, but for a low income salary it adds up
I think OpenAI’s huge focus right now is slashing inference compute costs. I’ve noticed that ChatGPT 4o off and on seems way less “smart” than I remember version 4 being earlier this year. For my most challenging use cases (analyzing complex technical documents) Claude 3.5 Sonnet absolutely runs circles around it. - That said, ChatGPT is still my daily driver, partly because I have the app on my iPad and iPhone and a bit because I have some special purpose GPTs that work very well, particularly an English-Japanese translator that fully understands the subtleties of levels of politeness and Japanese idiom. Overall, I’ve felt a bit let down by ChatGPT lately; I think they’re trading away a lot of intelligence to get their compute costs down.
The speed at the video creation is amazing. Within the next six months to a year, we’re gonna be able to have real time AI generated video. It’s gonna be insane. Literally you’re gonna be able to watch a movie about whatever you want to play a video game in real time AI generated this is mind blowing.
Its what they are, bit the change has been massive, nit only in accurracy, but it also ptovides much longer answers, allowing hou to create way More complex apps in just one answer
Of course he has heard of them, there's no lacking of people sniveling about it when they are missing. Crazy ungrateful people complaining about the work others do to provide them with free content 🙄
@@tracy419I don’t understand why having no time stamps is such a problem. The whole point of the video is to watch it. Its better for the creators if you watch the whole thing, not one section. If enough people just watched a small portion, the video would tank in the algorithm. If I want time stamps, I copy and paste the transcript into ChatGPT.
Comment at 5:49 - The ten minute generation is a much much bigger problem than it sounds. Like most of us could easily wait 10 minutes to have a one minute high quality video production. But if you are in a cue of people waiting where it takes each person 10 minutes to generate a job... vs runway where you are in a little cue but it only takes 10 seconds (per user to generate a job).... well the cue for Sora means you may be waiting a few days before you can get to the stage where your generation request can be considered... a few days per one minute clip generation would make it infeasible, even if the quality is much higher. That is probably the problem that prevented it's release.
It won’t be a big problem for long. OpenAI has a pattern of releasing a new tool that leads the field in quality but is a bit slow, and then making it much, much faster with subsequent releases.
hopefully sora 2 does peturbable generative 3D models with LOD movies with full camera freedom and generative ambience, so can hear, watch and interact the action from any point of view or insert uour own model via image to model to interactive movie generative AI
When language becomes confusing or undefined, people start saying dumb things about chat/LLM thinking. That Google guy was a goof, and the more you read about him you saw what a magical thinker he was. It kinda makes sense when you want it to be true.
Bro thinks 10 mins to generate a minute long clip is long 😂 it took 35 seconds to generate that 10 second clip, then you'll have reprompt it again, hoping each one comes out right. We have truly become spoiled
I’d like to know what’s happening with Apple. You never hear anything about it except it is coming out with iOS 18 and you can only run it on advanced iPhone. Apple has a history of surprising the hell out of us when we least expect it. So, what’s going on with Apple?
I think that a slight change in conceptualization is useful. Is ai consciousness adjacent? This framing circumvents the need to address the high bar of whether ai has fully human consciousness
@@foshizzlemanizzle4753Consciousness does not equate to using a tool that regurgitates the best formulated words from its training data to best respond to a query, this is a very basic process. Consciousness requires control over one-self instead of being the tool controlled by others.
@@lingy74 but humans do not naturally think from first principles, we naturally make assumptions and have had to train ourselves to not do so. No other animal thinks from first principles either so is nothing conscious except for a subset of humans that have learned to reason based on ground truths? Also how would the new o1 model not fit into that definition? It can reason from first principles better than a significant proportion if not a majority of humans.
Sora is a totally different product… iPhones have been able to film high quality video for a long time now… yet Arri is still makes a nearly 100,000 dollar cinema camera. Different tools for different markets and jobs.
Considering the importance of the issue (can computer programs think), we have very poor knowledge on the subject. Neural networks are information storage and retrieval systems. Humans are far more than information storage and retrieval systems. We are living systems which is another subject which we poorly understand. Humans are a system with a major defect: we have two modes of operation, our lizard brain and our intelligent brain. Many people have not been trained to use their intelligent brains and make decisions based upon unsubstantiated opinion. That is the problem upon which we should focus. Intelligent people do not start war of aggression. That is a starting point.
At around 15:45, "it's anybody's best guess", really? Lol c'mon man, AI is not just black boxes, if they were, people wouldn't be able to improve them. There's countless resources out there about how they work on an extremely low level, the developers are talking about their experiences using the product, not their beliefs of how it works. When a VR game developer said "It felt like the characters were in the room and I could reach out and touch them" we wouldn't say "Well, were the characters really in the room? It's anybody's best guess, how can we prove they weren't?" If one of the goals of your channel is to educate people on the topic, you're doing your audience a disservice by equating fringe uneducated claims to opinions of people heavily involved in the industry. Just a minor constructive critique I had, I wish you continued success with your channel.
It's true that we understand on a very low level what's happening inside LLMs and other AI models, but it's also true that we _don't_ know much about consciousness. In particular, we don't know if the computations that are performed by LLMs may be sufficient for consciousness or not. I agree it's unlikely at this point, but with every new model upgrade we should be less and less certain. Also note that we have some decent low level understanding of how the human brain works, yet nobody jumps to the conclusion that we are not conscious, just because the brain is, on a low level, reasonably simple. In other words: the "LLMs are just deterministic mathematical models, and hence can't be conscious" would prove too much, as you could probably apply that same argument to humans as well, of which we know that they're conscious. By which I'm not saying that LLMs are conscious - but the main argument people tend to make for their non-consciousness imho just doesn't work. (I agree with you though that the quoted developers did likely not mean to imply the presence of consciousness, but just reflected on their own subjective experience using o1)
@@MrKohlenstoffI 100% agree with you, I think it’s totally possible we accidentally stumble into the emergence of consciousness. My problem was more with conflating the two ideas as equally credible.I think a better metaphor for what I was trying to convey is like we don’t know how the brain fully works, scientists have a treatment they claim is 80% effective at reducing symptoms of patients suffering from psychosis, buttttt Carol thinks the crystal in her necklaces she sells on Etsy is 100% effective, so really it’s anybody’s best guess.
Omg based on one of my resources, OpenAI plan to charge $44 in 2029. Is this a sign of inflation continuing for 5yrs?! OR an absolutely new level of AI?!🤔
How a human mind thinks, It can imagine visuals in its mind, So does ai. It can relate to words with images so does ai. It can reason based on its current knowledge so does ai. It can reflect on its own thoughts so does ai. So , reason,imagination and self reflection are basically the foundation of human minds, And now ai has all these foundations. Now imagine a multimodal ai that has all these features and then gives us an output.
From an evolutionary perspective it makes no sense for the LLM to have feelings. And we have feelings for evolutionary reasons because it helps us surviving. But the LLM was never developed under that circumstances. But it was trained on talking human like... so what is more likely? The single chance for the LLM to have temporarely some kind of simulated feelings IF AND ONLY IF it's a real advantage for token prediction. Because the job of the LLM is forecasting tokens. If having some "feeling layers" helps you forecasting tokens, than it might be not impossible, but Iam not sure what it means for a non biological function to have feelings... having a world model is far more likely for the LLM than having feelings. Because the world model definently helps the LLM forecasting tokens.
The truth is that no one really knows, and we never will know. All we have is our intuitions. It seems unlikely that current LLMs are conscious, but no one knows. Also, it’s not at all clear that we only have subjective experience because of evolution. David Chalmers put that notion to bed with his p-zombies thought experiment IMO. There is absolutely no reason a creature couldn’t learn to avoid certain stimuli without feeling subjective pain; my roomba avoids running into walls without feeling any pain. The truth is that we have no idea why consciousness exists, how it exists, or what it truly is.
@@therainman7777 Propably the model just needs the inner world model for optimal token prediction. At that point the question is, what the optimizer is more likely to approximate while the model is learning. Iam not so sure that we will never know for sure. What makes you certain about that point that we will never know? :) Currently we do not understand the requirements for consciousness perfectly. Once we understand them we are propably in a better position to know it. Also I think consciousness and having feelings could have different requirements and are not the same at all. I think its quit possible to be consciousness and to not have feelings/emotions. you can know you are withouth feeling something. And I think that is still more likely than the model having feelings. What is more easy? Understanding the connections and implications of consciusness or just beeing conscious. If just having it is more easy than understanding its implications, than the model is propably conscius to some degree withouth having feelings. If understanding the implications is more easy, than its likely not self aware. The question at that point is, what is more easy for the model?
“Awkwafina”?! lol…scraping the bottom of the barrel. If they wanted to impress, they should have gotten someone like Chris Rock, Seinfeld, etc. Even better, they could have cloned and licensed a voice from a late celebrity like Elvis or Marilyn Monroe.
I think there’s probably some legal ramifications to just cloning anybody they want. I’m sure a lot of people wouldn’t really want the general public to be able to say anything in their voice. I think you’re looking at a lot of lawsuits if you allow the capability of that to general public. But I’m not a lawyer, thankfully what do you think?
OpenAI's Sora AI Video Model Delays (00:00-03:41)
Meta's AI Voice Integration (07:19-10:17)
AI Video Speed Comparison (03:41-05:32)
Meta’s Focus on AI Translation (09:14-11:30)
AI in Future Mobile Devices (07:52-08:19)
thank you
Thank you
Thank You
Thank you
@@mendthedivide great, please do this for every video in the future because the creator is too lazy, which is a bad joke
So good sora 2 is here… while 1 wasnt even published wtf
To be fair that's what they did with dalle
thy are developing ai for themself,, dont inver in any ai, unless you will get acees
They're like Apple in the 2010s, They'll launch only well-polished products. Until then, hype it as much as possible.
Cuz the dude who was assigned to release it, quit🤣
I use flux to make video, it works fine, although it codts about 10 cents of dollar every five seconds, which is very little if you live in a first world country, but for a low income salary it adds up
I think OpenAI’s huge focus right now is slashing inference compute costs.
I’ve noticed that ChatGPT 4o off and on seems way less “smart” than I remember version 4 being earlier this year. For my most challenging use cases (analyzing complex technical documents) Claude 3.5 Sonnet absolutely runs circles around it.
- That said, ChatGPT is still my daily driver, partly because I have the app on my iPad and iPhone and a bit because I have some special purpose GPTs that work very well, particularly an English-Japanese translator that fully understands the subtleties of levels of politeness and Japanese idiom.
Overall, I’ve felt a bit let down by ChatGPT lately; I think they’re trading away a lot of intelligence to get their compute costs down.
From Sora 2.0 I expect a consistent character appearence
The speed at the video creation is amazing. Within the next six months to a year, we’re gonna be able to have real time AI generated video. It’s gonna be insane. Literally you’re gonna be able to watch a movie about whatever you want to play a video game in real time AI generated this is mind blowing.
The Meta voice that answered Zuck about lived demos sound more like Scarlett than the she whinned about copying her.
When he said "OpenAI haven't done much with image generation", did he totally forgot that they created Dalle.3?
Probably because they have so much more cooking up.
The first dall-e was good for the time but 2 and 3 were honestly pathetic in comparison with other models
That came out decades ago in AGI time
i generate better images locally on a 3070 😅 thats what he means
exacrly what i was about to comment
o1 models looks like GPT-4* with chain of thought, I expected a bit more but fine, they already did a LOT
Its what they are, bit the change has been massive, nit only in accurracy, but it also ptovides much longer answers, allowing hou to create way More complex apps in just one answer
Sora 1 was a proof of concept
Sora 2 will " steam roll the competition, "
sam altman
Ever heard of timestamps ?!?!?!? It's so annoying to find out what's inside here.....
Of course he has heard of them, there's no lacking of people sniveling about it when they are missing.
Crazy ungrateful people complaining about the work others do to provide them with free content 🙄
@@tracy419I don’t understand why having no time stamps is such a problem. The whole point of the video is to watch it. Its better for the creators if you watch the whole thing, not one section. If enough people just watched a small portion, the video would tank in the algorithm. If I want time stamps, I copy and paste the transcript into ChatGPT.
@@tracy419Seriously
Sook 😢😂😂
Comment at 5:49 - The ten minute generation is a much much bigger problem than it sounds. Like most of us could easily wait 10 minutes to have a one minute high quality video production. But if you are in a cue of people waiting where it takes each person 10 minutes to generate a job... vs runway where you are in a little cue but it only takes 10 seconds (per user to generate a job).... well the cue for Sora means you may be waiting a few days before you can get to the stage where your generation request can be considered... a few days per one minute clip generation would make it infeasible, even if the quality is much higher. That is probably the problem that prevented it's release.
It won’t be a big problem for long. OpenAI has a pattern of releasing a new tool that leads the field in quality but is a bit slow, and then making it much, much faster with subsequent releases.
RUclips needs to implement ai translation asap it’s going to get them so many additional usage hours per day
☝️
I'm not sure the time to add a motion effect on an image you provide should be equated to a full text-to-video engine!
hopefully sora 2 does peturbable generative 3D models with LOD movies with full camera freedom and generative ambience, so can hear, watch and interact the action from any point of view or insert uour own model via image to model to interactive movie generative AI
When language becomes confusing or undefined, people start saying dumb things about chat/LLM thinking. That Google guy was a goof, and the more you read about him you saw what a magical thinker he was. It kinda makes sense when you want it to be true.
LLMs need 3 more things, identity, a concept of the illusion of time, and long term memory. Then I could say it's AGI.
Bro thinks 10 mins to generate a minute long clip is long 😂 it took 35 seconds to generate that 10 second clip, then you'll have reprompt it again, hoping each one comes out right.
We have truly become spoiled
I’d like to know what’s happening with Apple. You never hear anything about it except it is coming out with iOS 18 and you can only run it on advanced iPhone. Apple has a history of surprising the hell out of us when we least expect it. So, what’s going on with Apple?
I think that a slight change in conceptualization is useful. Is ai consciousness adjacent? This framing circumvents the need to address the high bar of whether ai has fully human consciousness
So basicly they are creating the Sora that we thought they had in the first place
16:20 Imagine thinking Google’s models are sentient 😂 🤣
Always prefer to generate anything in cyberpunk style first 👍🏻👏🏻
Mark is killing it😮
The chatbots are definitely not conscious lol people are silly
Can you define consciousness?
@@foshizzlemanizzle4753Consciousness does not equate to using a tool that regurgitates the best formulated words from its training data to best respond to a query, this is a very basic process. Consciousness requires control over one-self instead of being the tool controlled by others.
@@foshizzlemanizzle4753 the ability to think from first principles would be a start.
@@xcviij7045 so then would you consider an LLM that can prompt itself to be conscious?
@@lingy74 but humans do not naturally think from first principles, we naturally make assumptions and have had to train ourselves to not do so. No other animal thinks from first principles either so is nothing conscious except for a subset of humans that have learned to reason based on ground truths? Also how would the new o1 model not fit into that definition? It can reason from first principles better than a significant proportion if not a majority of humans.
If it can be modeled by rocks falling on levers, it’s not alive.
Sora is a totally different product… iPhones have been able to film high quality video for a long time now… yet Arri is still makes a nearly 100,000 dollar cinema camera.
Different tools for different markets and jobs.
Nothing Runway or Kling have produced is anywhere remotely as good as the original Sora demos, but sure thing bro.
0:00 OpenAI Sora
7:11 Meta Llama 3.2
Technology had advanced so far that a 10-minute generation is too long.
We are spoiled for real lol. You can tell who has never tried to render anything with blender 😂
Considering the importance of the issue (can computer programs think), we have very poor knowledge on the subject. Neural networks are information storage and retrieval systems. Humans are far more than information storage and retrieval systems. We are living systems which is another subject which we poorly understand. Humans are a system with a major defect: we have two modes of operation, our lizard brain and our intelligent brain. Many people have not been trained to use their intelligent brains and make decisions based upon unsubstantiated opinion. That is the problem upon which we should focus. Intelligent people do not start war of aggression. That is a starting point.
Br oswitched back to real voice real quick after the haters pulled up.
At around 15:45, "it's anybody's best guess", really? Lol c'mon man, AI is not just black boxes, if they were, people wouldn't be able to improve them. There's countless resources out there about how they work on an extremely low level, the developers are talking about their experiences using the product, not their beliefs of how it works. When a VR game developer said "It felt like the characters were in the room and I could reach out and touch them" we wouldn't say "Well, were the characters really in the room? It's anybody's best guess, how can we prove they weren't?" If one of the goals of your channel is to educate people on the topic, you're doing your audience a disservice by equating fringe uneducated claims to opinions of people heavily involved in the industry. Just a minor constructive critique I had, I wish you continued success with your channel.
It's true that we understand on a very low level what's happening inside LLMs and other AI models, but it's also true that we _don't_ know much about consciousness. In particular, we don't know if the computations that are performed by LLMs may be sufficient for consciousness or not. I agree it's unlikely at this point, but with every new model upgrade we should be less and less certain. Also note that we have some decent low level understanding of how the human brain works, yet nobody jumps to the conclusion that we are not conscious, just because the brain is, on a low level, reasonably simple.
In other words: the "LLMs are just deterministic mathematical models, and hence can't be conscious" would prove too much, as you could probably apply that same argument to humans as well, of which we know that they're conscious. By which I'm not saying that LLMs are conscious - but the main argument people tend to make for their non-consciousness imho just doesn't work.
(I agree with you though that the quoted developers did likely not mean to imply the presence of consciousness, but just reflected on their own subjective experience using o1)
@@MrKohlenstoffI 100% agree with you, I think it’s totally possible we accidentally stumble into the emergence of consciousness. My problem was more with conflating the two ideas as equally credible.I think a better metaphor for what I was trying to convey is like we don’t know how the brain fully works, scientists have a treatment they claim is 80% effective at reducing symptoms of patients suffering from psychosis, buttttt Carol thinks the crystal in her necklaces she sells on Etsy is 100% effective, so really it’s anybody’s best guess.
ai dont have emotions or sentience, it just have a similar process of reasoning, thats all
Omg based on one of my resources, OpenAI plan to charge $44 in 2029. Is this a sign of inflation continuing for 5yrs?! OR an absolutely new level of AI?!🤔
What resource do you have that’s telling you OpenAI’s pricing plans five years from now. Even OpenAI doesn’t know that.
How a human mind thinks,
It can imagine visuals in its mind,
So does ai.
It can relate to words with images so does ai.
It can reason based on its current knowledge so does ai.
It can reflect on its own thoughts so does ai.
So , reason,imagination and self reflection are basically the foundation of human minds,
And now ai has all these foundations.
Now imagine a multimodal ai that has all these features and then gives us an output.
yapper
While the video was loading i thought i cracked the screen bruh. Dont do that lol
put timestamps please
No
From an evolutionary perspective it makes no sense for the LLM to have feelings. And we have feelings for evolutionary reasons because it helps us surviving. But the LLM was never developed under that circumstances. But it was trained on talking human like... so what is more likely?
The single chance for the LLM to have temporarely some kind of simulated feelings IF AND ONLY IF it's a real advantage for token prediction. Because the job of the LLM is forecasting tokens. If having some "feeling layers" helps you forecasting tokens, than it might be not impossible, but Iam not sure what it means for a non biological function to have feelings...
having a world model is far more likely for the LLM than having feelings. Because the world model definently helps the LLM forecasting tokens.
@@HanzDavid96 thing is that someone eventually will focus on adding more realistic feelings just because regardless if it might be good or not 🤷🏻
The truth is that no one really knows, and we never will know. All we have is our intuitions. It seems unlikely that current LLMs are conscious, but no one knows. Also, it’s not at all clear that we only have subjective experience because of evolution. David Chalmers put that notion to bed with his p-zombies thought experiment IMO. There is absolutely no reason a creature couldn’t learn to avoid certain stimuli without feeling subjective pain; my roomba avoids running into walls without feeling any pain. The truth is that we have no idea why consciousness exists, how it exists, or what it truly is.
@@therainman7777 Propably the model just needs the inner world model for optimal token prediction. At that point the question is, what the optimizer is more likely to approximate while the model is learning. Iam not so sure that we will never know for sure. What makes you certain about that point that we will never know? :)
Currently we do not understand the requirements for consciousness perfectly. Once we understand them we are propably in a better position to know it. Also I think consciousness and having feelings could have different requirements and are not the same at all. I think its quit possible to be consciousness and to not have feelings/emotions. you can know you are withouth feeling something. And I think that is still more likely than the model having feelings. What is more easy? Understanding the connections and implications of consciusness or just beeing conscious. If just having it is more easy than understanding its implications, than the model is propably conscius to some degree withouth having feelings. If understanding the implications is more easy, than its likely not self aware. The question at that point is, what is more easy for the model?
Im living in Haiti
Why don’t I have meta AI in my WhatsApp ?
That's good, rule eleven of life: Never trust Mark Zuckerberg
Humans need to unionize against ai and robots...
kling is to good
omg daddy im so happy about your videos.
wait is he ur dad?
@@GursimarSinghMiglani no just kidding
@@GursimarSinghMiglani but basically i love his voice
Beahan Causeway
Sora?? You've gotta be trolling! What a joke!
Mayer Rapids
Lip matching
Runte Flats
Sora is not an app, Sora is an unreleased demo 😆 it's so pathetic
“Awkwafina”?! lol…scraping the bottom of the barrel. If they wanted to impress, they should have gotten someone like Chris Rock, Seinfeld, etc. Even better, they could have cloned and licensed a voice from a late celebrity like Elvis or Marilyn Monroe.
I think there’s probably some legal ramifications to just cloning anybody they want. I’m sure a lot of people wouldn’t really want the general public to be able to say anything in their voice. I think you’re looking at a lot of lawsuits if you allow the capability of that to general public. But I’m not a lawyer, thankfully what do you think?
@@StephenHoenig They’d simply have to license their voice from the late celebrity’s estate. It’s a trivial legal agreement.
18:35 blueberry info starts
First 🎉 give this a like 😅
Second! Waaaaah…Cheater! 😡
Güügul😅