"As an ai, i cannot create something that may infringe on copyrighted works of others, as it is unlawful and unethical. With that being said, we'll decide the deadliest strike location of a generic missile defense weapons system, not Lockheed martin's ER GMLRS...."
With AI trained on copyrighted material. They block every prompt with intellectual property because we would figure out that they had some things in the training dataset they shouldn't have.
@@XxXnonameAsDXxX I think you mean they *should* have. AI should be trained on everything, and that doesn't even violate copyright law. AI training is transformative. If youtubers like AI Explained can use content they don't own or have permission to, and that's considered fair use, then processing content into pattern recognition data is fair use too.
AI safety fearmongering is reserved for the people. When it's the most effective and destructive force in the world, the US military, then all their concerns about safety go out the door!
the goalposts shift every two years :> whenever somebody actually achieves it, the hype will die down as will the funding. so they just always stay 'a year away from agi' to keep the dosh inbound
I bet it's a marketing ploy. Keeping people in the belief that they have access to agi but deliberately hide it might incentivise more investment and misplaced confidence in their tech
That + military use is insane and honestly only makes sense if you're just trying to get as much investor cash as possible and don't think you can't actually get to AGI/ASI in the "thousands of days" range
@@zoeherriot haha so true...id love it if companies used our newfound ai overlords to just boost the productivity of existing employees, but I think we all know what would probably happen instead
yea the render quality at the pixel level is incredible good off the charts , however the floaty transforming inconsistencies make it unacceptable . i think in general there needs to be a more robust actual "world model" which can act as a rendering engine assembling generated components in a physics simulation , this intermediary format of modular components , like a blender scene with gaussian splats , would be ideal for any edits and consistency , and allow for refinement of each modular element .
yeah, because most people are hating on AI, and i dont even know whats wrong with AI! they dont think AI is good enough for military complex, majority is shilling on AI
This was always inevitable. It WILL happen and there is nothing you can do to stop it. Also your adversaries (China, Russia, Iran, North Korea, ...) are going to do it and they have absolutely no inhibitions or regulations to stop them. You will not survive if you don't do it (develop AI for military purposes). And what will happen then... nobody knows.
Investing in military AI is a question of survival. The free world is in hybrid wars against bad actors whether we want to or not. If we don't stay at the forefront or at least keep up, we are cooked. I'm more worried about intelligence services using AI to control public sentiment and some bad president/leader using that against their people and against other democracies. Public sentiment is already controlled by foreign powers to a larger extent than most people are aware of. But it is our own leaders that we need to fear the most. Increasing AI advancements need to be balanced by having genuinely good people in power and creating governing structures that prevent misuse. Yes, it seems we are screwed, but we should not accept that fate.
Because the idea that militaries are going to be prevented from getting ahold of this is laughable. No amount of concern, that is realistically achievable, will prevent that..
One thing I think AI explained didn't make clear: the Pro plan allows unlimited generations of any quality on relaxed mode, where your requests are sent to a queue that may take a while to complete. You get limits only if you use priority mode to generate them quicker. The Plus plan can only use priority mode, and has much tighter limits than the Pro plan, 1,000 vs. 10,000 credits.
Altman is clearly a super villain right now, he's pretty much Lex Luthor. He started an AI arms race, kicked out the senior leaders to get control, made it for profit to give himself a stake, changes the terms so they can make for profit weapons with AGI. He's literally the worst person to be leading this and a massive danger to the world. At least Hassabis is a nerd who likes puzzles.
If the AI is any good then there will be fewer innocent civilian deaths than if we send in the troops now. At the very least there will be no raping and looting. Also, What working people? The whole point is that there will be no working people, only robots. Whether that is good or not is another matter.
A few years ago, I had high hopes for OpenAI to achieve AGI safely... But this past year, they've been on a very clear path to become a greedy conpany like any other, driven by money with safety and moral standards getting less and less important. And for Microsoft, it has no loyalty. As soon as the deal gets inconvenient for them, they'll drop it. They are supporting other AI labs and have their own researchers working on models. It's a sad world wmfor those worried about AI safety.
Once you've met Humanity long enough you stop worrying. You simply say your prayers at the beginning of the day and keep going, as log as their greed and violence allow it.
It's a for profit corporation and all their competitors (Anthropic, Google, Meta, Amazon, Apple, to a degree Microsoft) are doing exactly the same. I understand that people judge that something called "OpenAI" is just as closed and commercial as everyone else. But I don't understand the hate beyond this point. All the other AI companies are doing exactly the same.
@@thornelderfin There is no need to go beyond that at all.. The fact that he corrupted OpenAI's mission and betrayed the people that put him there is enough to distrust him. That's why most of the technical leader left the organization. It is not hate. It's distrust.
sorry to be sarcastic... but who would have guessed! I think it is safe to say, that if AGI isn't going to be benevolent because it's the default for a higher intelligence. it will be the best turbo capitalist on the planet and we get better ready to be exploited on a whole new level... good times!
This is genuinely scary. It's gonna be a race to the bottom if things continue improving at the current rate. I miss good old days of quirky benign bots...
There will be nothing to exploit once artificial intelligence is better than us at everything. We'll be genocide'd on a whole new level. And I'm not joking.
The average movie is around 105 to 120 minutes long, which means it costs approximately 10,000 to 12,000 dollars to make. That’s a small amount compared to the huge budgets Hollywood movies have.
@ yes, I could see Hollywood being very excited. Most of a movies budget goes to VFX these days, AI has an opportunity to completely change the industry
@@omarnomadlol, nobody's gonna create good movies with Sora. What a nonsense comparison. Real movies have (mostly) real physics. You can create 10, 000Sora clips and none will have real physics and only a minority will have precisely followed your instruction.
its a joke, I don't know why people are getting so excited. If you use the hallouai model you get 3 days of unlimited generations. you have to wait in a queue but you can set a bunch of them going at once. I managed to make a 2 minute video telling a coherent story for my niece to cheer her up about penguins all setting off on a trip around the world to see her. Using pictures of her on holiday as a source to animate. And you just need an email address to get at that and do it all over again for another 3 days. At least until it goes under as they must be burning through money trying to attract a paying audience.
That GPTs get to moderate content generation and sharing is an aspect of all this that I think doesn't get enough attention. These companies create the moderators that are bound to become our "content police" whenever we interact with our machines. Take any provocative piece of art as an example, lets say Lolita by Nabokov...will you be able to write that novel on a laptop in 10 years? Or will your word processor refuse to cooperate? Or will you be too worried about your software licenses being blocked or even the police being called to even attempt the project? Would you be able to sell it via amazon?
After getting to the end of your video, I have to take your side Phil. Thought they might be doing this for awhile and it's good to know others are just as wise
To be fair, it doesn't take much for the US military to move a few billions from its trillion-dollar budget to develop AI the moment it sees it could have effective military applications. Let's be real; this isn't something that can be stopped in the long term, just like any big enough nation nowadays can build its own nuclear weapons if it really wants to.
Everything is a distraction. (This will get worse) Assume you are captured by it and remove yourself from the equation for a bit. Silence is golden Pure gold
@ “EGO” Agreed, this ls the powerful perspective (behind motivations) that dominates us all into assuming we are so precious and powerful to seek advice, correction, compassion or change. Ego is check point. It doesn’t go away for good reason
I've started to gain a sort of sixth sense for detecting when sponsorships are about to happen, and I'm glad to say it didn't give a false positive when you mentioned VPNs XD
Great coverage and content as always man. I am starting to get the impression that we’re in the prequel movie to a cyberpunk dystopian trilogy where the “leading evil corporation” is OpenAI and the greedy corpo hellscape must be stopped by a heroic net runner and his sidekick. _que intro music_ Looking forward to your next upload. Thanks again for no shameless clickbait headlines and exaggerations.
They are clearly encoding a positional element into multiple conceptual objects on screen 6:50 & this is how it relates to AGI. This is much cheaper to store than a 3D model. See jackets & other clothing maintaining the same style & some occluded objects still existing on the other side…clearly there is the ability to create & destroy objects that have their own constructs describing them in a loose DAG-like fashion. It is the interactions that can create new objects or destroy them…if something is off screen for long enough it gets forgotten but it sounds like in their papers that they want a way to cheaply store a conceptual model of everything in the world, not just what is in the next frame based on the last few frames as previous techniques have done…I think that is what they mean by AGI…if a robot butler could have the concept of a dishwasher in a room which itself has this concept of a handle, etc. chained together in a way where it could do new things without having to hardcode any functionality…that is the AGI-esque dream.
I know you've said you're using VPN oto access SORA, but at 0:39 you can clearly see a sentence: "Accessing or offering access to our services outside of the countries and territories listed below may result in your account being blocked or suspended." . so please be careful, this being a public video
@@user-sl6gn1ss8prealistic. Open Ai have been wrestling with these issues for the last 12 months. So it’s not going to be a surprise for them. Hoping to find a real business model for it. It’s not good enough for professional uses cases. They know that that is why it hasn’t launched. Hence bundle it in the middle of this PR extravaganza. Much more useful things to come in the next 9 days I think. But Sora is a toy as it stands.
Even if Sora was released to generate more revenue, new and increasingly greater streams of revenue will likely be required to fund the development of AGI. Similar to how SpaceX’s commercial efforts have been in service of the larger mission to make humanity a multi - planetary civilization.
With regards to the applicability of Sora to AGI, there’s a bit of a phenomenon in instruct models as we know them. They have a limited output distribution, meaning they will trend towards specific set phrases, and are much more limited in other ways in the content they can produce. “Base models” without instruction tuning can, in this way, function as a sort of “world model”, or a much more “true” distribution of natural data that an instruct model can clean up for use in making a final prediction, for instance. I think the idea of Sora is similar. It’s meant to be a sort of world model, or simulation of a world, so that the end AGI model can evaluate the impact of its decisions on the real world (no more water staying in an upside down glass of water problems).
If these employees really cared about getting assurances about the technology not being used for war, then they should quit immediately. It obviously will be. An assurances are lies
🇧🇷🇧🇷🇧🇷🇧🇷👏🏻 Thank you for sharing this incredible video! As always, your work never fails to impress! No real time video yet, hope they release it soon...
And, @aiexplained-official, I have to tell you that proper physics in videos involves many factors, such as background movement, color persistence, and even the evolving appearence of plants, and othe things, among other details. A lot of those elements are already handled almost perfectly by SORA. It’s really amazing what this technology can achieve from just a prompt.
What's crazy is through an application like Krea, You get flux image generation, nearly the same amount of video generation, a real-time editor, and an image enhancer/style changer for nearly the same price as the $200 membership for open AI. I have spoken endlessly to people about how it's horrific that Amazon offers a lower quality version of a movie for a cheaper price. Especially when it costs them pennies to stream the higher quality version one. It's like they're specifically dumbing it down to justify having a higher tier cost. And I feel like Sora is doing the same thing. They literally dumb down the quality to 720p and put a watermark on it if you aren't willing to pay $200 a month 🤦♂️ extremely far fall from what open AI seemed to promise in its inception...
@@aiexplained-official It'd be quite impressive if Gemini 2.0 Flash is actually the Experimental 1206 model. On twitter Logan did have the tweet referencing a "Flash" and I did DM one of the research engineers at deepmind about it (asking if it was flash or pro) but all they responded with was the emoji "🤫" lol. But with Sora, well first, I honestly kind of think AGI will be more like a Gato model, but obviously much more scaled up and expanded in some of the more recent research directions (i.e. agents and "reasoners"). But essentially a model trained on a massive number of tasks and modalities under a generalist policy, and with Sora I think a pretty big modality may well be video generation.
kinda... melancholic. was hoping AI was going to change the world for the better. it still could i guess, but alas history seems to repeat itself over and over and over where the powerful get more powerful while the weak stay weak. promises being broken, just like always, yippee!
I get you. I feel that after covid I have just been rugpulled like 5 times now. There is never solid ground to stand on, and it seems like whatever meager contribution you can make to the world is just that, meager.
yeah, the video sadly feels like a reality check. The pessimistic prophecies of some now ring in my ears, and looks like the cool stuff will be for the rich and powerful in the end, and we could all die because of military AI. Weeeeeeee
That is a you problem. Trusting billionaires and companies with their promises. You think Altman cares about you? Musk? Gates? Dude, their priority is not to make you work less and feel better, it is to make money and maintain power. Only trust those closest to you, and have the mentality of "everything is a manipulation of truth until proven differently" for others, especially the rich folks.
Sora is dissapointing, but I feel you missed an important detail about pricing and that is that you have unlimited generations in "relaxed" mode using the $200 pro plan, like you have with runway and hailuo. That being said I would wait to see how long these queues end up being.
Easily the best release so far from OpenAI this month is the folders (at least in the non-API web interface), which they call Projects. A+ idea, C- implementation. I assume you talk to John ChatGPT. Please tell him to add: 1) subfolders 2) display of more than 5 folders at a time (I have no idea why this is a hard limit) 3) drag-and-drop reordering of folders 4) icons other than the "folder icon" to be used for folders I don't care if it breaks the "Projects" functionality otherwise (couldn't care less about it); I just want my conversations organized.
As for how Sora could lead ro AGI: many tasks require vision. Training in simulation can be scaled much easier than training in the real world. So having a visual world model could eventually be used to faster train smarter models in simulation. The generative models aren't good enough yet, but some time soon(ish), they might be.
Thank you for being critical and making us aware of what's going on. I love how differentiated your reports are! I did not know how much of the original OpenAI sentiment already got watered down.
There's an ex Microsoft guy called Dave (he wrote the windows task manager) with a YT channel and he borrowed a 50 thousand dollar server that had some 500gb of memory. He loaded up GPT4 level LLM and ran it on the machine. It was slower than the commercial GPT4 even though it was running locally on a machine right in front of him. That really hit home for me how much compute is actually going on server side just for the big text models. So when you factor in images, diffusion models for video. That has to be some wild compute in use. Its no wonder even the $200 tier doesn't get you much Sora use.
2:26 should we remind you at each one of your video that you also sometime can hallucinate? And this one was easy to spot, so imagine all the other one we aren’t able to spot in your videos?
After the latest releases (O1, Sora), it's clear to me that proprietary AI will be confined to consumer use and evolve into exactly what we saw in old sci-fi movies. The API will likely be reserved for key stakeholders, such as governments and militaries. It's a shame, but what did we really expect? That OpenAI would democratize AGI out of the goodness of their hearts?
The thing I find so annoying about these media generation models is they will consistently ignore specific details, which is where the use really comes in. This has been the case for about 2 years now, the quality has improved a lot but the adherence to the prompt is still weak, hope this is focused on soon.
It's fascinating to see how Sora blends cutting-edge video generation with so many limitations-both technical and ethical. The physics-defying outputs are a clear reminder that generative AI, no matter how visually impressive, still struggles to ground itself in the real world. The pricing model feels steep, especially given how quickly credits can vanish, but it’s also telling about the monetization strategies tied to this technology. The broader implications of OpenAI's evolving policies are harder to ignore. The potential pivot away from the AGI-for-humanity ethos is a major red flag, especially with the reported flexibility around their Microsoft agreement. If AGI is as transformative as they claim, consolidating control under a profit-driven entity seems at odds with their original mission-and deeply concerning for the broader impact on society. The military collaborations are another point of contention. Defensive or not, these moves hint at a willingness to adapt to industrial demands, which could erode trust in OpenAI’s long-term vision. It’s a stark contrast to the utopian ideals that originally drove the company and raises questions about how much influence external stakeholders like governments and corporations now wield. Overall, while Sora showcases incredible technical achievements, it’s hard to separate the tool itself from the larger narrative of OpenAI’s shifting priorities and the ethical trade-offs involved. It feels like these product launches, impressive as they are, might be serving as a smokescreen for more contentious decisions being made behind the scenes.
"You can't ask for a video in the style of a _living_ artist." All artists with popular visual styles now more afraid for their lives than health insurance CEOs.
I think one might note that a realistic creation of illumination effects when a scene moves could be a demonstration of an advanced understanding of physics. I wonder if Sora has not digested optical physics knowledge in its training data, which is, at least in its comprehensive visualization, difficult to model or produce-even with modern ray tracing programs.
It clearly learned physics. Whether it can follow the prompt or call on the appropriate abstractions in the network or avoid statistical contradictions / noise is another matter. It may not have learned everything perfectly either, but again, completely different claim than saying it doesn't understand physics.
Note that this is not Sora, this is Sora Turbo. Also remember in the previous web page about Sora they explained that they got better physics and world understanding by adding compute. So I think that is what is missing, more compute. And actually video data along with transcripts and larger model size is going to be key to getting more robust cognition, common sense reasoning, etc.
This has been the enduring argument during the past couple years, and we have seen evidence that adding more compute does improve the situation, but I'm growing increasingly skeptical that compute alone would be enough.
I believe the “Sora -> AGI” pipeline will look like having a model “imagine” situations using video generation (like the classic marble/cup/microwave thing); but it seems like hallucinations in the present model mean it currently falls short.
More importantly, I think it serves as a good indicator of the model's current understanding of reality and its fundamental laws and principles. So long that it keeps doing weird stuff, it means it does not actually have understanding. Whether it ever will remains an open question.
But he doesn't endorse the idea of getting a VPN to use Sora because it's not worth it, so I'm glad he didn't. And if he DID say it was worth it and did a VPN ad, we'd probably be skeptical. The only winning move was to make the joke.
It'll work eventually, but right now the error rate even making single images is still high enough that I have to spam out large batches and cherry-pick the best ones for anything even remotely challenging. Expecting the same processes to stay on the rails for even five seconds seems pretty optimistic to me.
I wonder if a video model will at some point be kinda mashed with text to 3D and text to animated skeletons and all that, because that should (if not only relying on the neural net, but also on actual simulation within the generated video or something along those lines) significantly help to reduce most hallucinations. Basically mix classical and AI systems to get the best of both worlds... Or just make it 10000x bigger and pray it works. Anyway, it's gonna be fun to try and jailbreak Sora xD, allthough much less tries than with initial DallE3.
@ 2:43 - that is not quite true. You get unlimited generations in relaxed mode on sora in the pro mode AND 500 videos in fast mode. And 50 in chatgpt plus. So each video even in plus (20 dollar) mode should just be 2% of the allowance. I dont know how you got to 5% Per video. Thats not even true in Plus.
As far as video generation's link to AGI goes... If you consider "dreams" ... That kind of world simulation is likely quite useful. I imagine that the dreaming system is part of the human brain's equivalent of generating more examples for neural net training. Plus, OpenAI is also working on robots. So, they need a decent world simulation system... Hopefully, someone has built a more interesting base model, than the one that is being used here for video generation.
The only thing that could impress me with these 12 days of openAI is if they could show us a path forward concerning how to tackle hallucinations. I honestly don't care about videos or other stuff. I'm mostly just interested in te bigger picture wand and so far I have seen very little to show that we have not hit a wall. It seems like a core issue they can't seem to get passed
Its rather easy to see how sora is related to AGI. It will in time be able to generate test data. Test data will help train AGI. Imo such a lodel should be only to handle text as well as video and audio.
To be fair, we know well that denoising sucks at instruction-following due to its paradigm: It's pretty much a telephone game, so keeping consistency over the iterations is... not simple.
This often gets overlooked, but companies need a revenue stream to fund and sustain capital intensive projects with long time horizons. Musk literally views SpaceX’s business model as a money printing machine that exists to finance the colonization of Mars. The development of AGI will require similar streams of revenue and I have absolutely no problem with Sora being one of them.
Once agi is actually achieved I'm sure rival labs are going to be able to replicate the results, anthropic seems to be really close and sometimes even ahead of openAI
At least one person working at OpenAI has stated that o1 already qualifies as AGI. This is why they're ditching it at this time. Because some define AGI as a machine capable of adequately performing most computer tasks that humans do. We are very soon going to be getting into the weeds of what exactly the definition of AGI is. Both the anti-AI luddites and the capitalist investors of AI are going to be working together to ensure the definition is near impossible to meet.
AI adverts used in real time are horrendous but the next logical step. You get the companies assets (logo/etc) and then when keywords are hit, advert is created in the style of the website/theme.
Their path through latent space to get between prompts is insane and inhuman, but that’s moderately expectable. Still, so far they can’t go far coherently. They need to be xtrained on game engines.
Feels like basically a novelty. Not sure how film or other content makers can leverage tools like this. Seems like it has a long ways to go (if it ever gets there at all).
Jason Wei (AI Researcher at OAI) on twitter said on twitter "in a decade using a computer will just be interacting with a sora model" which is essentially what my own thoughts have been for the development of this video generation tech as well.
BS. You and Jason (and most people) have learned nothing from 2020, for that was just a teaser. 2.0 is almost here. "Futurists" are WAY off on what's coming for humanity.
@@FilmSpook I think the LLM OS (kind of like envisioned by Anrej Karpathy) is rather accelerationst, and I personally think we'll get this within a few years.
Btw now o1-full takes less time thinking than o1-mini their "faster at reasoning" model, and dispite that o1's limit is still the same, it feels like theyre just trying to maximize profit now
![Chris Langan - The Interview THEY Didn't Want You To See - CTMU [Full Version; Timestamps]](http://i.ytimg.com/vi/9miVG2xT5jY/mqdefault.jpg)
"No, I'm not going to do an awkward ad for VPN's." lol So good.
Instantly earned a like from me lol
This is how he intended for y'all to react
I was just about to skip ahead 30s
it wouldve actually been quite smooth so i wouldnt complain if he did
Think how much engagement it would've created if he continued with a VPN sponsorship directly, without even acknowledging the contradiction.
/jk
"Given the potential for abuse, we're not *initially* making that available for *all* users"
So reassuring
Just those that pay far more.
The pro version even removes the watermark... The purpose is definitely monetary, not safety
Rich people are more trustworthy in Sama's view
@@tellesu commies who game the system and liberal vrmn will love it
Lol. How many ppl are gonna be pised u cant upload celebs or characters with IP.
$200 to create stock footage. 😂😂😂
"As an ai, i cannot create something that may infringe on copyrighted works of others, as it is unlawful and unethical. With that being said, we'll decide the deadliest strike location of a generic missile defense weapons system, not Lockheed martin's ER GMLRS...."
With AI trained on copyrighted material. They block every prompt with intellectual property because we would figure out that they had some things in the training dataset they shouldn't have.
@@XxXnonameAsDXxX I think you mean they *should* have. AI should be trained on everything, and that doesn't even violate copyright law. AI training is transformative. If youtubers like AI Explained can use content they don't own or have permission to, and that's considered fair use, then processing content into pattern recognition data is fair use too.
AI safety fearmongering is reserved for the people. When it's the most effective and destructive force in the world, the US military, then all their concerns about safety go out the door!
That AGI clause being reworked should be huge news. Shady shady shady.
the goalposts shift every two years :> whenever somebody actually achieves it, the hype will die down as will the funding. so they just always stay 'a year away from agi' to keep the dosh inbound
I bet it's a marketing ploy. Keeping people in the belief that they have access to agi but deliberately hide it might incentivise more investment and misplaced confidence in their tech
That + military use is insane and honestly only makes sense if you're just trying to get as much investor cash as possible and don't think you can't actually get to AGI/ASI in the "thousands of days" range
@@WoolyCow If they got AGI, they wouldn't need investment - every company in the world would be clamoring to buy a license.
@@zoeherriot haha so true...id love it if companies used our newfound ai overlords to just boost the productivity of existing employees, but I think we all know what would probably happen instead
5:12 “not bad at all” - bro the ship is parked on the highway instead of in the water 😂
Haha I didn't catch it first time
lmaaaao
It's quite funny how this whole A.I. video industry manages to be very impressive and dogshit at the same time😅
Some people have absurdly high standards and aren't impressed with anything.
Exactly my feelings. It's both remarkable that it's possible and completely useless.
It's incredible how much time you can spend making something that is almost but not quite entirely unusable.
yea the render quality at the pixel level is incredible good off the charts , however the floaty transforming inconsistencies make it unacceptable . i think in general there needs to be a more robust actual "world model" which can act as a rendering engine assembling generated components in a physics simulation , this intermediary format of modular components , like a blender scene with gaussian splats , would be ideal for any edits and consistency , and allow for refinement of each modular element .
It's physics!
Scanning the comments, it looks as if about 1% express concern about military+AI. And these are people who follow AI news! We are so screwed.
yeah, because most people are hating on AI, and i dont even know whats wrong with AI!
they dont think AI is good enough for military complex, majority is shilling on AI
This was always inevitable. It WILL happen and there is nothing you can do to stop it. Also your adversaries (China, Russia, Iran, North Korea, ...) are going to do it and they have absolutely no inhibitions or regulations to stop them. You will not survive if you don't do it (develop AI for military purposes).
And what will happen then... nobody knows.
Investing in military AI is a question of survival. The free world is in hybrid wars against bad actors whether we want to or not. If we don't stay at the forefront or at least keep up, we are cooked.
I'm more worried about intelligence services using AI to control public sentiment and some bad president/leader using that against their people and against other democracies. Public sentiment is already controlled by foreign powers to a larger extent than most people are aware of. But it is our own leaders that we need to fear the most. Increasing AI advancements need to be balanced by having genuinely good people in power and creating governing structures that prevent misuse. Yes, it seems we are screwed, but we should not accept that fate.
Because the idea that militaries are going to be prevented from getting ahold of this is laughable. No amount of concern, that is realistically achievable, will prevent that..
@@jopearson6321 We might not prevent it but at least to slow it down. Or limit its impact
"Sight for sora eyes" is a good one.
One thing I think AI explained didn't make clear: the Pro plan allows unlimited generations of any quality on relaxed mode, where your requests are sent to a queue that may take a while to complete. You get limits only if you use priority mode to generate them quicker.
The Plus plan can only use priority mode, and has much tighter limits than the Pro plan, 1,000 vs. 10,000 credits.
Next Christmas announcement by OpenAI: At long last, we have created the Torment Nexus from classic sci-fi novel "Don't Create The Torment Nexus"
The only difference is that it also steals your money.
Tired: Sora is out!
Wired: new AI Explained video is out! 🔥
Altman is clearly a super villain right now, he's pretty much Lex Luthor.
He started an AI arms race, kicked out the senior leaders to get control, made it for profit to give himself a stake, changes the terms so they can make for profit weapons with AGI. He's literally the worst person to be leading this and a massive danger to the world. At least Hassabis is a nerd who likes puzzles.
But they have no moat
business as usual for a smallhat
He is also a jew
@@larion2336 nah they just have working brains, but I can see why that'd be hard for you to grasp
@@CarlSnub my brain works just fine, it's why I'm not a spineless dog like you. Either that or you're one of them, which is worse.
"by the way, this isn't just rumor, this is according to _multiple people familiar with the conversations_ " 😂
What's the time stamp on that? I missed it
@@Progpm11:15
Do you not understand the concept of corroboration?
that one def raised my eyebrow as well
@@justtiredthings ah, so no corroboration in rumors? shut up doofus
Imagine working as an AI dev and your legacy being innocent civilian deaths and worsening the class divide by undermining working people.
That's not the legacy, that's a side effect.
If the AI is any good then there will be fewer innocent civilian deaths than if we send in the troops now. At the very least there will be no raping and looting.
Also, What working people? The whole point is that there will be no working people, only robots. Whether that is good or not is another matter.
@@ronilevarez901it's the literal point
A few years ago, I had high hopes for OpenAI to achieve AGI safely...
But this past year, they've been on a very clear path to become a greedy conpany like any other, driven by money with safety and moral standards getting less and less important.
And for Microsoft, it has no loyalty. As soon as the deal gets inconvenient for them, they'll drop it. They are supporting other AI labs and have their own researchers working on models.
It's a sad world wmfor those worried about AI safety.
Have you read the system card for o1 released by OpenAI regarding the "safety" research done by Apollo Research? Rather terrifying.
Once you've met Humanity long enough you stop worrying.
You simply say your prayers at the beginning of the day and keep going, as log as their greed and violence allow it.
is it any real surprise? for profit incentives end up ruining everything. it's just the inevitable consequence of capitalism
This is like a caveman clearly seeing fire's path to destruction.
That Altman returned increasingly appears to be a huge mistake. He seems to have no solid principles, apart from money, money, money.
yeah. jesus. fuck. they really put him back in there????????? what the FUCK
It's a for profit corporation and all their competitors (Anthropic, Google, Meta, Amazon, Apple, to a degree Microsoft) are doing exactly the same. I understand that people judge that something called "OpenAI" is just as closed and commercial as everyone else. But I don't understand the hate beyond this point. All the other AI companies are doing exactly the same.
@@thornelderfin we always hope for someone to be better than the rest and become the saviors of Humanity.
@@thornelderfin it's still called "Open" AI for a reason, the backstabbed their supporters
@@thornelderfin There is no need to go beyond that at all.. The fact that he corrupted OpenAI's mission and betrayed the people that put him there is enough to distrust him. That's why most of the technical leader left the organization. It is not hate. It's distrust.
I regret that I have but one like to give. THANK YOU for reporting on their shady ethical trajectory!
sorry to be sarcastic... but who would have guessed! I think it is safe to say, that if AGI isn't going to be benevolent because it's the default for a higher intelligence. it will be the best turbo capitalist on the planet and we get better ready to be exploited on a whole new level... good times!
I think it will quickly overcome capitalism and find something "better" right away.
Universal Paperclips here we come! We used to laugh at Clippy, now look who's in charge...
This is genuinely scary. It's gonna be a race to the bottom if things continue improving at the current rate. I miss good old days of quirky benign bots...
There will be nothing to exploit once artificial intelligence is better than us at everything. We'll be genocide'd on a whole new level. And I'm not joking.
$10 per video is a very hard sell considering the variability In output quality.
It’s pretty good if you consider unlimited 4o and o1 prompts, o1 pro, and ChatGPT 4.5 which is rumoured to cone out in these 12 days of releases
The average movie is around 105 to 120 minutes long, which means it costs approximately 10,000 to 12,000 dollars to make. That’s a small amount compared to the huge budgets Hollywood movies have.
@ yes, I could see Hollywood being very excited. Most of a movies budget goes to VFX these days, AI has an opportunity to completely change the industry
@@omarnomadlol, nobody's gonna create good movies with Sora. What a nonsense comparison. Real movies have (mostly) real physics. You can create 10, 000Sora clips and none will have real physics and only a minority will have precisely followed your instruction.
its a joke, I don't know why people are getting so excited. If you use the hallouai model you get 3 days of unlimited generations. you have to wait in a queue but you can set a bunch of them going at once. I managed to make a 2 minute video telling a coherent story for my niece to cheer her up about penguins all setting off on a trip around the world to see her. Using pictures of her on holiday as a source to animate. And you just need an email address to get at that and do it all over again for another 3 days. At least until it goes under as they must be burning through money trying to attract a paying audience.
That GPTs get to moderate content generation and sharing is an aspect of all this that I think doesn't get enough attention. These companies create the moderators that are bound to become our "content police" whenever we interact with our machines. Take any provocative piece of art as an example, lets say Lolita by Nabokov...will you be able to write that novel on a laptop in 10 years? Or will your word processor refuse to cooperate? Or will you be too worried about your software licenses being blocked or even the police being called to even attempt the project? Would you be able to sell it via amazon?
we need simplebench detailed test on o1 pro
he only does it once the API is available, which it isn't
You get unlimited generations on the $200 tier. If you turn on Relaxed mode it will take longer but gives you unlimited generations.
Oh cool. Do you have it?
After getting to the end of your video, I have to take your side Phil. Thought they might be doing this for awhile and it's good to know others are just as wise
To be fair, it doesn't take much for the US military to move a few billions from its trillion-dollar budget to develop AI the moment it sees it could have effective military applications. Let's be real; this isn't something that can be stopped in the long term, just like any big enough nation nowadays can build its own nuclear weapons if it really wants to.
I guess the video -> agi connection is for making an imagined plan for an embodied agent before attempting to execute that plan
Also I assume their figure agi remote worker will be able to be a video avatar
Everything is a distraction. (This will get worse)
Assume you are
captured by it and remove yourself from the equation for a bit.
Silence is golden
Pure gold
Golden order. Distraction is how ego takes back control.
@
“EGO”
Agreed,
this ls the powerful perspective (behind motivations)
that dominates us all into assuming we are so precious and powerful to seek advice, correction, compassion or change.
Ego is check point.
It doesn’t go away for good reason
I've started to gain a sort of sixth sense for detecting when sponsorships are about to happen, and I'm glad to say it didn't give a false positive when you mentioned VPNs XD
I have that skill too alan, nice
Great coverage and content as always man.
I am starting to get the impression that we’re in the prequel movie to a cyberpunk dystopian trilogy where the “leading evil corporation” is OpenAI and the greedy corpo hellscape must be stopped by a heroic net runner and his sidekick. _que intro music_
Looking forward to your next upload. Thanks again for no shameless clickbait headlines and exaggerations.
I was “sora- ly” disappointed lol. It refused do a video generation that was animation because it said there were people in it.
They are clearly encoding a positional element into multiple conceptual objects on screen 6:50 & this is how it relates to AGI. This is much cheaper to store than a 3D model. See jackets & other clothing maintaining the same style & some occluded objects still existing on the other side…clearly there is the ability to create & destroy objects that have their own constructs describing them in a loose DAG-like fashion. It is the interactions that can create new objects or destroy them…if something is off screen for long enough it gets forgotten but it sounds like in their papers that they want a way to cheaply store a conceptual model of everything in the world, not just what is in the next frame based on the last few frames as previous techniques have done…I think that is what they mean by AGI…if a robot butler could have the concept of a dishwasher in a room which itself has this concept of a handle, etc. chained together in a way where it could do new things without having to hardcode any functionality…that is the AGI-esque dream.
🤓☝️
(very interesting comment, I didn't think of it that way, but yeah, that aligns a lot with the idea of wanting to create a model of the world)
Like object permanence? Thats toddler's level
@@XinDseal A toddler's level is infinitely better than a fetus's level.
Yes, AGI needs relationships between objects through time.
Great video, the speed in which you drop them so they are current is incredible!! Thank you
i wouldn't really imagine that AGI has the potential to be contained while changing the world
12 days of releases might be 12 days of disappointment. OpenAI can’t break capitalism while leaning on it. I don’t have high hopes.
I know you've said you're using VPN oto access SORA, but at 0:39 you can clearly see a sentence: "Accessing or offering access to our services outside of the countries and territories listed below may result in your account being blocked or suspended." . so please be careful, this being a public video
I sacrifice for the video
especially considering the critical tone
@@user-sl6gn1ss8prealistic. Open Ai have been wrestling with these issues for the last 12 months. So it’s not going to be a surprise for them. Hoping to find a real business model for it. It’s not good enough for professional uses cases. They know that that is why it hasn’t launched. Hence bundle it in the middle of this PR extravaganza. Much more useful things to come in the next 9 days I think. But Sora is a toy as it stands.
@@aiexplained-official legend
Even if Sora was released to generate more revenue, new and increasingly greater streams of revenue will likely be required to fund the development of AGI. Similar to how SpaceX’s commercial efforts have been in service of the larger mission to make humanity a multi - planetary civilization.
People apparently think you can make money out of thin air to achieve a moonshot
Lightning fast as always, thanks Philip!
With regards to the applicability of Sora to AGI, there’s a bit of a phenomenon in instruct models as we know them. They have a limited output distribution, meaning they will trend towards specific set phrases, and are much more limited in other ways in the content they can produce.
“Base models” without instruction tuning can, in this way, function as a sort of “world model”, or a much more “true” distribution of natural data that an instruct model can clean up for use in making a final prediction, for instance.
I think the idea of Sora is similar. It’s meant to be a sort of world model, or simulation of a world, so that the end AGI model can evaluate the impact of its decisions on the real world (no more water staying in an upside down glass of water problems).
God I hope Ilya gets to AGI first
I'm starting to think that human greed dictates that we would be better off never reaching agi
@bitterheat3415 Pandora's Box is opened though. Don't think we will stop until we do
Based in Israel. You might want to take a look at the track record of the tech sector there.
how
Agi is so close I doubt it. I bet it will be Google and open ai within months of each other. Anthropic as well
If these employees really cared about getting assurances about the technology not being used for war, then they should quit immediately.
It obviously will be. An assurances are lies
🇧🇷🇧🇷🇧🇷🇧🇷👏🏻 Thank you for sharing this incredible video! As always, your work never fails to impress! No real time video yet, hope they release it soon...
And, @aiexplained-official, I have to tell you that proper physics in videos involves many factors, such as background movement, color persistence, and even the evolving appearence of plants, and othe things, among other details. A lot of those elements are already handled almost perfectly by SORA. It’s really amazing what this technology can achieve from just a prompt.
What's crazy is through an application like Krea, You get flux image generation, nearly the same amount of video generation, a real-time editor, and an image enhancer/style changer for nearly the same price as the $200 membership for open AI.
I have spoken endlessly to people about how it's horrific that Amazon offers a lower quality version of a movie for a cheaper price. Especially when it costs them pennies to stream the higher quality version one. It's like they're specifically dumbing it down to justify having a higher tier cost. And I feel like Sora is doing the same thing. They literally dumb down the quality to 720p and put a watermark on it if you aren't willing to pay $200 a month 🤦♂️ extremely far fall from what open AI seemed to promise in its inception...
I think you should definitely put that o1 "escaping" coverage on the main channel, because the media is severely misrepresenting it!
Do you think Google is going to release 2.0 sometime this month?
I think Exp might have been what they had wanted to call 2
@@aiexplained-official It'd be quite impressive if Gemini 2.0 Flash is actually the Experimental 1206 model. On twitter Logan did have the tweet referencing a "Flash" and I did DM one of the research engineers at deepmind about it (asking if it was flash or pro) but all they responded with was the emoji "🤫" lol.
But with Sora, well first, I honestly kind of think AGI will be more like a Gato model, but obviously much more scaled up and expanded in some of the more recent research directions (i.e. agents and "reasoners"). But essentially a model trained on a massive number of tasks and modalities under a generalist policy, and with Sora I think a pretty big modality may well be video generation.
kinda... melancholic. was hoping AI was going to change the world for the better. it still could i guess, but alas history seems to repeat itself over and over and over where the powerful get more powerful while the weak stay weak. promises being broken, just like always, yippee!
I get you. I feel that after covid I have just been rugpulled like 5 times now. There is never solid ground to stand on, and it seems like whatever meager contribution you can make to the world is just that, meager.
yeah, the video sadly feels like a reality check. The pessimistic prophecies of some now ring in my ears, and looks like the cool stuff will be for the rich and powerful in the end, and we could all die because of military AI. Weeeeeeee
That is a you problem. Trusting billionaires and companies with their promises. You think Altman cares about you? Musk? Gates? Dude, their priority is not to make you work less and feel better, it is to make money and maintain power. Only trust those closest to you, and have the mentality of "everything is a manipulation of truth until proven differently" for others, especially the rich folks.
Very professional and objective video, which is refreshing and hope-inducing.
If they a “letting” Microsoft have their AGI, seems like agi is still pretty far away
Sora is dissapointing, but I feel you missed an important detail about pricing and that is that you have unlimited generations in "relaxed" mode using the $200 pro plan, like you have with runway and hailuo. That being said I would wait to see how long these queues end up being.
Me too bloxy
Easily the best release so far from OpenAI this month is the folders (at least in the non-API web interface), which they call Projects. A+ idea, C- implementation.
I assume you talk to John ChatGPT. Please tell him to add:
1) subfolders
2) display of more than 5 folders at a time (I have no idea why this is a hard limit)
3) drag-and-drop reordering of folders
4) icons other than the "folder icon" to be used for folders
I don't care if it breaks the "Projects" functionality otherwise (couldn't care less about it); I just want my conversations organized.
As for how Sora could lead ro AGI: many tasks require vision. Training in simulation can be scaled much easier than training in the real world. So having a visual world model could eventually be used to faster train smarter models in simulation.
The generative models aren't good enough yet, but some time soon(ish), they might be.
I am really grateful about your great content. Especially since it is not making VPN ads 😃
Thank you for being critical and making us aware of what's going on. I love how differentiated your reports are!
I did not know how much of the original OpenAI sentiment already got watered down.
Great video. Thanks. Open source video generator will come soon. Everything else has.
Thanks, excellent content, as always. 🤩
The quality is very good. It's looking bright. 1-2 years and sora equivalent runs locally on 24GB vram ☺
There's an ex Microsoft guy called Dave (he wrote the windows task manager) with a YT channel and he borrowed a 50 thousand dollar server that had some 500gb of memory. He loaded up GPT4 level LLM and ran it on the machine. It was slower than the commercial GPT4 even though it was running locally on a machine right in front of him. That really hit home for me how much compute is actually going on server side just for the big text models. So when you factor in images, diffusion models for video. That has to be some wild compute in use. Its no wonder even the $200 tier doesn't get you much Sora use.
And some crazy people will make NF4 quantized versions so it runs on a laptop with 6 GB of VRAM. Or maybe not "runs" but at least walks with a limp.
2:26 should we remind you at each one of your video that you also sometime can hallucinate? And this one was easy to spot, so imagine all the other one we aren’t able to spot in your videos?
After the latest releases (O1, Sora), it's clear to me that proprietary AI will be confined to consumer use and evolve into exactly what we saw in old sci-fi movies. The API will likely be reserved for key stakeholders, such as governments and militaries. It's a shame, but what did we really expect? That OpenAI would democratize AGI out of the goodness of their hearts?
i hate human greed
The thing I find so annoying about these media generation models is they will consistently ignore specific details, which is where the use really comes in.
This has been the case for about 2 years now, the quality has improved a lot but the adherence to the prompt is still weak, hope this is focused on soon.
It's fascinating to see how Sora blends cutting-edge video generation with so many limitations-both technical and ethical. The physics-defying outputs are a clear reminder that generative AI, no matter how visually impressive, still struggles to ground itself in the real world. The pricing model feels steep, especially given how quickly credits can vanish, but it’s also telling about the monetization strategies tied to this technology.
The broader implications of OpenAI's evolving policies are harder to ignore. The potential pivot away from the AGI-for-humanity ethos is a major red flag, especially with the reported flexibility around their Microsoft agreement. If AGI is as transformative as they claim, consolidating control under a profit-driven entity seems at odds with their original mission-and deeply concerning for the broader impact on society.
The military collaborations are another point of contention. Defensive or not, these moves hint at a willingness to adapt to industrial demands, which could erode trust in OpenAI’s long-term vision. It’s a stark contrast to the utopian ideals that originally drove the company and raises questions about how much influence external stakeholders like governments and corporations now wield.
Overall, while Sora showcases incredible technical achievements, it’s hard to separate the tool itself from the larger narrative of OpenAI’s shifting priorities and the ethical trade-offs involved. It feels like these product launches, impressive as they are, might be serving as a smokescreen for more contentious decisions being made behind the scenes.
Just the kind of clear eyed, and quite alarming analysis. "We're all gonna die" on this timeline.
"You can't ask for a video in the style of a _living_ artist." All artists with popular visual styles now more afraid for their lives than health insurance CEOs.
How can you possibly make living artists the villain on the level of a predatory healthcare-denial-for-profit CEO?
Come up with your own style bro.
😂😂😂
I think one might note that a realistic creation of illumination effects when a scene moves could be a demonstration of an advanced understanding of physics. I wonder if Sora has not digested optical physics knowledge in its training data, which is, at least in its comprehensive visualization, difficult to model or produce-even with modern ray tracing programs.
It clearly learned physics. Whether it can follow the prompt or call on the appropriate abstractions in the network or avoid statistical contradictions / noise is another matter.
It may not have learned everything perfectly either, but again, completely different claim than saying it doesn't understand physics.
Note that this is not Sora, this is Sora Turbo. Also remember in the previous web page about Sora they explained that they got better physics and world understanding by adding compute. So I think that is what is missing, more compute. And actually video data along with transcripts and larger model size is going to be key to getting more robust cognition, common sense reasoning, etc.
This has been the enduring argument during the past couple years, and we have seen evidence that adding more compute does improve the situation, but I'm growing increasingly skeptical that compute alone would be enough.
I believe the “Sora -> AGI” pipeline will look like having a model “imagine” situations using video generation (like the classic marble/cup/microwave thing); but it seems like hallucinations in the present model mean it currently falls short.
More importantly, I think it serves as a good indicator of the model's current understanding of reality and its fundamental laws and principles. So long that it keeps doing weird stuff, it means it does not actually have understanding. Whether it ever will remains an open question.
Thank you for the video, just a heads up your chapters aren't working (must begin at zero)
Still broken
Ditching the AGI promise is incredibly damning and disappointing. That's the entire point of this technology.
I've seen so many crappy VPN ads and it would be one of the few videos, where it would have made actual sense. 😅
But he doesn't endorse the idea of getting a VPN to use Sora because it's not worth it, so I'm glad he didn't. And if he DID say it was worth it and did a VPN ad, we'd probably be skeptical. The only winning move was to make the joke.
@@mal2ksc yeah sure. :)
It'll work eventually, but right now the error rate even making single images is still high enough that I have to spam out large batches and cherry-pick the best ones for anything even remotely challenging. Expecting the same processes to stay on the rails for even five seconds seems pretty optimistic to me.
I wonder if a video model will at some point be kinda mashed with text to 3D and text to animated skeletons and all that, because that should (if not only relying on the neural net, but also on actual simulation within the generated video or something along those lines) significantly help to reduce most hallucinations. Basically mix classical and AI systems to get the best of both worlds... Or just make it 10000x bigger and pray it works.
Anyway, it's gonna be fun to try and jailbreak Sora xD, allthough much less tries than with initial DallE3.
Dropping the AGI clause is a good way to get a cyberpunk dystopia. Maybe Sam Altman is a big fan of gritty Sci Fi.
Military use is indeed a concern. I'm glad you mentioned it.
@ 2:43 - that is not quite true. You get unlimited generations in relaxed mode on sora in the pro mode AND 500 videos in fast mode.
And 50 in chatgpt plus. So each video even in plus (20 dollar) mode should just be 2% of the allowance.
I dont know how you got to 5% Per video. Thats not even true in Plus.
Great reporting as always.
As far as video generation's link to AGI goes...
If you consider "dreams" ... That kind of world simulation is likely quite useful.
I imagine that the dreaming system is part of the human brain's equivalent of generating more examples for neural net training.
Plus, OpenAI is also working on robots. So, they need a decent world simulation system...
Hopefully, someone has built a more interesting base model, than the one that is being used here for video generation.
imagine seeing a tricky physics problem that you have to intuitively try to solve. Sora getting new situations is kind of like this.
The only thing that could impress me with these 12 days of openAI is if they could show us a path forward concerning how to tackle hallucinations.
I honestly don't care about videos or other stuff. I'm mostly just interested in te bigger picture wand and so far I have seen very little to show that we have not hit a wall. It seems like a core issue they can't seem to get passed
The important part with $200/mo is that you also get unlimited slow generations. Kinda like Runway's $100/mo tier.
My most expected review today, thanks for always bringing the best content about AI as ever. ❤️
Thanks, I won't be able to sleep tonight. But good info to talk about
I appreciate your development from AI fan-girl to AI-critique - while still having a love for this technology
Its rather easy to see how sora is related to AGI. It will in time be able to generate test data. Test data will help train AGI. Imo such a lodel should be only to handle text as well as video and audio.
Great video ad always 👏
To be fair, we know well that denoising sucks at instruction-following due to its paradigm: It's pretty much a telephone game, so keeping consistency over the iterations is... not simple.
Props on the title ❤
our fiction about AGI is basically a self-fulfilling prophecy. we know ourselves too well.
This often gets overlooked, but companies need a revenue stream to fund and sustain capital intensive projects with long time horizons.
Musk literally views SpaceX’s business model as a money printing machine that exists to finance the colonization of Mars.
The development of AGI will require similar streams of revenue and I have absolutely no problem with Sora being one of them.
great video, wonderful day!
does it get a 1 out of 10 on simple bench?
Once agi is actually achieved I'm sure rival labs are going to be able to replicate the results, anthropic seems to be really close and sometimes even ahead of openAI
At least one person working at OpenAI has stated that o1 already qualifies as AGI. This is why they're ditching it at this time. Because some define AGI as a machine capable of adequately performing most computer tasks that humans do. We are very soon going to be getting into the weeds of what exactly the definition of AGI is. Both the anti-AI luddites and the capitalist investors of AI are going to be working together to ensure the definition is near impossible to meet.
Great video as always. Not surprised at all about big companies being scumbags, that's just a given.
Ilya may be our only chance now
AI adverts used in real time are horrendous but the next logical step.
You get the companies assets (logo/etc) and then when keywords are hit, advert is created in the style of the website/theme.
Great video as always, are you going to test out Gemini Experimental 1206?
Their path through latent space to get between prompts is insane and inhuman, but that’s moderately expectable. Still, so far they can’t go far coherently. They need to be xtrained on game engines.
Feels like basically a novelty. Not sure how film or other content makers can leverage tools like this. Seems like it has a long ways to go (if it ever gets there at all).
Jason Wei (AI Researcher at OAI) on twitter said on twitter "in a decade using a computer will just be interacting with a sora model" which is essentially what my own thoughts have been for the development of this video generation tech as well.
BS. You and Jason (and most people) have learned nothing from 2020, for that was just a teaser. 2.0 is almost here. "Futurists" are WAY off on what's coming for humanity.
@@FilmSpook I think the LLM OS (kind of like envisioned by Anrej Karpathy) is rather accelerationst, and I personally think we'll get this within a few years.
Press X to doubt on that one
More like, in a decade, computers will be using us.
Btw now o1-full takes less time thinking than o1-mini their "faster at reasoning" model, and dispite that o1's limit is still the same, it feels like theyre just trying to maximize profit now
I completely agree with your hypothesis about them trying to distract.