Okay, we're here in the studio with Ethan He, most recently of XAI.
Welcome.
Yes, thank you.
Glad being here.
We're also here with Vibhu.
You were first coming to us or joining the late in space world because you were working on Cosmos and NVIDIA and you did a great paper.
We loved it.
You presented it as well.
So thank you for doing that.
Yeah, I've actually also presented the MOEs twice at Latent Space.
Yeah, yeah, yeah.
How did you actually hear about us?
Did we reach out to you?
Is that how it worked?
No, actually the community.
Like I realized, oh there's this online community yeah, that people talk about AI and also learn learn from each other through papers every, every week, through the paper club.
It's very nice.
Yeah, I learned a lot, I think three years non-stop.
We haven't stopped, even on Christmas and New Year's.
Many weeks.
I want to stop.
I think it's good, but I think you had posted that you worked on a paper and I was like oh, very cool.
We have a paper club.
Present it.
But I might have researched you after.
Yeah, because it's an amateur club, right?
Yeah.
So it's very unusual.
But we have sometimes people, authors come by and actually explain the paper.
Today we just did the poolside interview paper, which is apparently very good, came out yesterday pretty interesting right, fully open.
They talk about everything system, so it's a good one.
We'll recommend people to read it.
Bring us up to speed on your transition to XAI, because I actually don't even know when you joined, just like tell the story about the sort of transition.
Before XAI I was working on Cosmos world model at NVIDIA.
So Cosmos is a giant video foundation model that aims to simulate the world.
And it serves as a foundation for all of the roboticists to build on top of.
There.
Once I built the Cosmos 1, I realized that this thing also has a scaling law similar to the language model.
We need to scale up the video models further.
That's why I realized I need to move to somewhere with much more compute resources.
Than NVIDIA?
Yeah.
And timeline wise, when was Cosmo?
It was pretty early, right?
It was open world model, open paper.
It was like end of 2024.
End of 2024.
Yeah, then at mid-2025, I moved to XAI.
At that time, I joined by the time when XAI was about to build video models and multi-model models.
There were no infra, no data, and no model.
And just as a few engineers, we built it in three months and released the first model Grok Imagine 09.
And since then I keep working on video models and move more from pre-training and to post-training of the video models.
For example, like a reference to videos, kind of like a cameo feature and video extensions.
And before I left, I worked on a word model, leading a small team to to focus on the real-time, long-horizon video generation.
Can you give like a rough roadmap of like, okay, you're on a brand new team.
Grok previously was only tech, so they partnered with BFL for their image and stuff.
What are the building blocks, right?
You have compute, data you can procure somewhere.
What are the sequence of things that people should think about when you're setting up a new team?
I mean, actually, even deeper, not just data you can procure.
You guys had to go through getting the data too, right?
So you shipped it pretty fast.
Yeah, three months is actually very surprisingly fast.
Yeah, one thing I say like thanks to my experience at NVIDIA, because first time when we were building Cosmos together, we built it for about a year.
So this is like the second time I do it.
Roughly have an idea like what to do.
I think the most important thing is talent.
Everyone was very strong and they were very close with each other towards a common goal.
So, That speed up things a lot.
So you reduce the communication bandwidth among people and everyone can work towards the same goal.
It's like every day there's not that much meetings on the calendar, like maybe like a sync a day.
And after that, it's just all building.
It was pretty fun at that time.
And another thing is that XAI has very strong foundations of data, data inference, model inference.
And the supporting there can help the model develop a lot.
When I look at training models, the top important thing is how many iterations can you do per day?
And the more iterations you can do, you can train the model much faster.
So if you have very strong infra and you have a lot of compute, you can train these models in very short periods of time.
That can... give you a much larger buffer for arrows.
And it also gives you the opportunity to spot more bugs.
Yeah.
What is an iteration?
Is it like a few hundred steps?
Let's say just training the model.
Like from acquire new data and maybe design new algorithms and train a new model, maybe a smaller scale.
Yeah, so cycle time for like any hyperparam that you're searching.
Yeah, cycle time and thin to like eval this model.
Is this model better than my previous iteration?
Yeah.
So it's like before you, someone had already set this up that you can iterate very quickly.
Yeah, I think the foundation there is extremely good for developing and research models.
And often I find...
It's kind of boring, but a lot of the improvements does not come from new algorithms.
It comes from finding small bugs here and there in the data pipeline, in the model training pipeline.
Those give the biggest boost to the model quality.
It's interesting, right?
So you say it's like small team less communication bandwidth, but also a lot of quality is like find little bugs.
It seems counterintuitive, right?
You have a lot of people.
You can iron out more of those, but it's interesting to see the other side, right.
Yeah.
I also wonder, do you try using LLMs to look for bugs?
I don't know.
I remember at that time it was mid-2025.
So the coding model wasn't quite there yet.
I remember like December 2025, it was extremely good.
Yeah, I've been using it at that time.
It's helpful.
Sometimes it produces codes that are kind of difficult to maintain, even though, like the first time, it built something extremely fast, but it gave the like a spaghetti code thousands of lines that I couldn't maintain and the OEM itself couldn't figure out what's wrong and how to improve on top of it.
But now I find it much, much, much better.
Yeah, I want to bring up another point.
Here is like now coding models are much more efficient and can help us implement stuff much faster.
Compute might become a bottleneck again.
Because previously, if you want to train a new model, say you want to generate new synthetic data or write a new algorithm it might take a few weeks.
And during that period of time, you might not have experiments to run.
And now you can build that thing within a few hours.
Then you can immediately train a model. now you have to have enough compute to try all of the ideas.
So compute might be the bottleneck of iterating speed again.
Yeah.
Actually honestly, I think it's kind of a stressful job because you're like well, I should be trying everything And if I'm not, then I'm not doing my job well.
I mean.
There's also the stress of you're eating thousands of GPUs per hour, which is very expensive.
And, you know, compute can go to other researchers.
Daddy Elon.
Daddy Elon.
But, you know, there's still finite amount of compute.
Like you want to use it, you want to use it well.
You want more of it.
That was quite stressful indeed.
Yeah, I think one thing is, with coding models, now a lot of these jobs can be automated, which is much better.
Second, it's a marathon, so you've got to maintain good health and fitness.
A regular schedule.
It's hard to hear that when you shift from zero to nothing in three months.
Yeah, I mean, I think obviously the culture, famously, you know, people work very hard.
One thing I did want to dive into.
You know, in the notes that you sent ahead of time, you had specific comments about the cost of video gen training.
Presumably this is on the Colossus One, right?
The 200 megawatt cluster. and whatever you want to share.
I think there's three things you're talking about, right?
So there's video gen, there's also the image gen model that you put out.
Do you want to like complete that?
Okay, so zero to one, you have a few months.
Just what are the stages of create image gen?
Oh yeah, maybe I got distracted.
Sorry.
And then, you know, from there, there's video gen, there's audio gen.
I'd love to get into those next, but what is that first few months like?
So small team, a lot of bugs, iterations, but like, you know, what does it look like?
Do we take something off the shelf?
Do we just get data compute?
What's the few months like?
How do you go to state-of-the-art image and model?
How do you just start?
Yeah.
I cannot comment specifically how I did, but it's a quite standard process.
I can draw some examples from Cosmos.
So mainly it's like building a video model, you actually need to build an image model first.
And building these two models.
The data you need is 100 synthetic pair of language and image, or language to video.
Because on the internet, actually the videos don't naturally associate with text.
So you can say oh, like on YouTube, you have the title and you have the description and the comments of the video.
But usually they're not relevant to the video itself and say maybe, like the video, is a natural scene of mountains or something.
And the title is like, I'm so happy today.
So they have no correlation at all.
So the first step is to, you have to generate a synthetic pair of language with the videos.
So you get the videos from the internet and you use a VLM to caption the videos.
So that part, here's the question, how do you get the VLM to begin with?
So if there's no... You fuse the model, right?
Yeah.
Say if there's no VLM exists, how do you generate the text to the beginning?
It's impossible.
I see.
In the beginning, it's like you ask human to describe the video as detailed as possible.
For example, you ask them to describe everything, like all objects, all characters and all interaction and dialogues in the videos.
So that's in the protocol of Cosmos Labeling.
We require the.
The objective we gave to the labelers was that you have to describe the video as detailed as possible, such that a blind person hears a blob of text can reconstruct what the video is like from their head.
Video or image?
You're talking about video or image, either either one of them.
This was pretty common when we went from like Clip and Dolly right, it's all training on really detailed captioning of images.
So same is applied to video.
But Instead of using multimodal model to pass in video images and write rich descriptions, you can also I mean, I think that's the traditional perspective of supervised or very highly human, curated thing.
I feel like there's an unlock with unsupervised.
Where you have enough to bootstrap that you can just throw common corpus on it or whatever.
Unsupervised vision and language pairing, where you just have interspersed image and text and it just learns.
To me, that is the VLM breakthrough that is different from the clip, different from the pre-LM era.
Yeah, yeah.
It's interesting to see that you kind of need both data.
Yes.
You need to bootstrap it.
Yeah, for the generative model training.
There's also really like a small percentage of unlabeled data.
So the model is instructed to generate a video without any text instruction.
That can also help the model generalize.
So after this stage of generating the synthetic pair, So one important common step is to train a compressor or a tokenizer of the image or videos.
So because if you can theoretically train image or video models on pure pixels.
But the problem is that it's a lot of tokens.
So like one image, like it's a thousand by a thousand, it's like one million tokens, one million pixels.
It's impossible to transform around that.
So you need to train a tokenizer which can go from image to latent space and latent space back to image.
That's why we named the podcast.
Exactly.
Basically, you're talking about vocabulary, science.
Yeah, it's a vocab.
A million is impossible?
In generative models, the vocab is continuous.
It's a continuous space.
We can think about, like you map an image to a vector, It's a fixed length vector.
It's like 16 or 48, something like that.
And then you map that vector back to the image space.
And the mapping has...
The mapping is patch-based.
So you say you have a 16 by 16 patch, then you match, you map that patch of pixels into this latent space.
We've covered this in the Vision Transformer.
Yeah, VAEs.
You basically compress your input, you do your generation, you're reasoning all that generation in smaller dimension and then you project back out.
Yeah.
VAEs is a form compression, but for me, the patching thing is from VIT, right?
Yeah, you can't make those.
Literally, the paper is titled like 16 by 16 is all you need, something like that.
And then I think also people make a lot of comparisons with this kind of patching with convolutions.
Yes, yes.
Which is you're kind of reconstructing the old paradigm with the new.
Yeah, actually in VAEs, there are both convolution networks and transformers.
You can actually do both.
Yeah.
After this VAE.
So what you've got is you've got latent space tokens and you've got the language tokens.
So now the training of the diffusion transformer.
Usually, generation models use diffusion transformers.
It's actually quite standard.
It's very similar to how you train language transformer models.
It's not that much difference.
It's just the tokens, the visual tokens in, visual tokens out.
The only difference is there's a denoising process.
So you train the model to... unmask some of the noise.
So you add random noise to the visual tokens and then you train the model to remove those noise, to generate the clean tokens.
And in inference the model can iteratively remove noise from 100 noise, yeah.
And then there's also, to speed things along on the tech tree of diffusion, there's CFG like, and then there's also, I guess, latent diffusion that you know.
There's someone in there, I think, somewhere along the line obviously, like stability, and all these other guys pioneered a lot of this, like.
I don't know if you want to get into that or just do the video side.
Up to you.
After you train such model, such image model.
The reason it's a foundation for video models is that Image models are cheaper to train and they have much denser connection between language and text.
So, sorry, language and images.
For example, you train a billion images and there's a mapping from... from the tax to the image.
And the cost to trend the same like a billion tax to a billion videos.
That's much more expensive because videos naturally have more tokens than images.
Because the diffusion models, their understanding of language purely come from this mapping.
So if you don't have enough mapping, so if you only train on like 10 million videos or something, you might not see enough language tokens in your training.
So your model does not understand human intention enough.
So that's why you usually you first train this image diffusion models and then you bootstrap the video model from there.
One thing I did want to ask, because actually I think you're the first video model person I've ever talked to.
I think
We've talked to Luma and all those folks.
There's all these tricks in video compression where basically, frame by frame, there's not that much difference.
So actually you don't have to regenerate or it's resaved the whole frame, right?
I think MP4 compression or something else like that.
Is it tempting to use that?
Or as far as I can tell, Everyone just treats it as, no, we will just generate every frame.
Is that roughly the state of the art?
There are a few different approaches.
Let's say first, you want to just directly use MP4 compression and use that as the tokens for the transformers to train right.
So...
People actually have tried that.
But the main challenge is the latent space for the MP4.
Tokens were not very comprehensible for the models.
It's extremely hard to train on that.
And there's a So that's why we created VAEs, which creates more continuous latent space.
So the models can understand that latent space and learn from it much easier.
Even within the VAEs, there are different difficulties of the latent space.
So you can imagine something that...
The simplest, the most naive.
VAE is like you have an image and you just shuffle all of the images into a vector.
So you don't need to train any VAEs, right?
But that little space is extremely hard for models to train on top of.
That's why there's some debate on how do you compress the tokens.
So you mentioned you can compress frame by frame.
Also, you can compress the temporal dimension.
Yes.
The difference is, if you compress temporal dimension, you get a much higher compression rate because there's temporal redundancy between frames, because this frame and the last frame likely they are mostly similar
So there's only some small difference.
For example, I think in 1.2.1 VE, they have like an 8x8x4 compression rate.
So the four temporal tokens are compressed into one token.
That can save a lot of the context length.
If you do it frame by frame, you have to do maybe like 8x8x1.
Your context lens will be four times larger.
That being said, the benefit of the per-frame compression we might come back to this later is real-timeness and interactivity.
Because if you...
If you strain the output of the model frame by frame, the model can respond to any user request immediately.
So if you have like a temporal four compression, four times compression, then it might be laggy.
Yeah, there's a lag there in nature.
So you're very pilled on this.
Let's just go ahead and bring it up because we have the visual prepared anyway.
There's some frontier applications of real-time video gen.
So Flipbook is one of the examples that went viral recently, right?
What is Flipbook?
Flipbook is kind of like a web browser.
You can see it has the web browser UI on top.
The difference is all of the UIs are generated by generative image model in real time.
And anything here are fake.
But you can explore inside this imaginary world.
Say, here we have engineering the Great Pyramid.
The model generates this for us to understand how it works.
And if we want to navigate around and understand further, we can click on some of the description here and the model will generate a new page, new sub-page, describing the details we want to know about.
So it's basically kind of, we are playing a video, But it's pausing for our next interaction.
And then it just plays the next thing based on our interaction.
Yeah.
It's kind of cool.
Yes.
You kind of decide your story.
So this was, you know, how do you make a pyramid?
Levering technique seemed interesting, right?
It shows, how do you take, okay, I want to know.
The demo tweet had more animation between frames.
I think it's just skipping.
It's just skipping a lot of frames.
They also have a video mode, but I guess a lot of people are using it.
It's not available.
There's a live video stream.
We can try.
Yeah.
So this is an example of the kind of future that you see at the extreme.
We're obviously not in it today, but in a world where inference is completely free, this is better than generating code in text.
Yeah, so this is the final state of where we'll be at for our work model, I think.
Imagine internet doesn't exist, and then you're happy in google.com.
Like, what should a model show you?
The model can imagine something, and this is what the model imagined.
And these web pages, they completely do not exist.
So I think as the inference costs come down, we are going to have generative UI for everything.
If you think about how the coding model works.
So they write code for a web page and they render.
The code might be converted into binary and the binary render the pixels on the screen.
So in machine learning, every time we have some breakthrough, obviously it's more intuitive.
So why don't we have user instruction to the pixel directly?
So the generated UI will be user intention to the pixels directly.
And say like, even if I want email, let's say everyone have the same interface, but I want it slightly different.
I want the email to show to me like a TikTok so I can swipe left and right for the emails.
And, or maybe you want something else.
We can have completely different things.
Or like I have, I'm looking at Instagram stories.
I don't like the like button.
I always make clickbait.
And China is a UI without it.
So it's going to be a revolutionary replacement of the interface.
So in the future, we might have much more powerful LMs and coding models running behind the scene.
And in the front end, the distortion model will actually be the front end to show stuff to you.
That's how I imagine it.
Yeah.
Diffusion frontend, deterministic backend.
Yeah.
Something like that.
I find that very expensive, but, you know.
I find it interesting you called LM's writing code on the backend deterministic, but okay.
Yeah, you write it once and then you execute.
If you think about the cost, say, let's say H100 costs $1 per hour.
And if you use this eight hours a day and 30 days, so every month you're paying this 240.
You're likely not and not want to pay for that.
That's even more expensive than Cloud Code Max.
But if you think about the compute costs come down like two times every year.
I think that the future will likely arrive within two years.
It's everything, right?
Compute cost comes down, compute gets faster, model gets smarter, model gets smaller.
Yeah, I don't know why you say two times because I think it's like a hundred times.
In language models it is roughly 100 to a thousand times every 12 to 18 months for the same given level of LMS.
That's a net of everything, right?
That's model performance alongside compute.
So different than just compute costs come down.
But, you know, a very interesting future.
Yeah, for the web designers, we'll have to shout out that accessibility is an issue, right?
Like how do you deal with screen readers or whatever?
But yes, this is higher bandwidth storytelling than anything you can possibly generate with code, right?
So I think that's the rough idea.
I'd like to add a little bit of that.
So humans naturally have the maximum bandwidth when we are looking at things, look at videos.
And we also have maximum output bandwidth when we are talking.
So in the future it might be something like we talk to AI models and the AI model responds back with the generative UI.
So that will be the maximum input and output bandwidth to interact with AI models before Neuralink happens.
And I mean, it's also very custom, right?
Some people are very visual.
Some people are not as visual, right?
They prefer the text.
But the best thing about generative UI, right, can also be text.
Yes.
There's another project that we wanted to highlight, which is the Neural OS.
Kind of similar idea, but here you're literally operating, simulating an operating system with a video model.
Yes.
And you can play Doom, you can do Firefox.
I find this like mildly less impressive obviously because it's an OS that I can run.
But here everything is imagined.
I was, you know, used to the command W to close the Firefox tab.
It didn't crash.
It's too immersive.
It's too immersive for me.
I wanted to close the tab.
But yes, I can play generated Diffuse.
This is shockingly fast.
Yeah.
Because I remember there was a demo about like maybe one or two years ago.
Someone tried to do the first person shooter with an image model.
There was no consistency.
It was very slow.
But here, realistically, this is Doom.
I mean, I think there's two sides to that, right?
There's like, okay, what is running a game?
The heavy part of it is actually the game engine, all the lighting, all that stuff, the graphics.
This is just kind of video, right?
We've solved consistency.
This is still, you know, it looks like a few years old image generation.
There's some temporal consistency, but it's kind of just images stitched together as frame video.
But it's a good visual representation to picture the future you want to see, right?
Like that's what I see in these more so.
This reminds me of how the video models gets better and better.
So in your OS, if you just look at it, it feels like it's just a crappy version of the Windows.
We could have right.
But the difference is, this model is overfitted on the existing operating systems.
It can generate nothing different than that. but it's actually also similar to video models.
So when we're training these video model, image model, we train them on internet.
There's no imaginary supernatural stuff on the internet, but once we train this model, you can prompt the model to generate something supernatural that have never existed in the data set.
So if you train your neural OS or neural computer on the standard screen recordings on the entire internet, the model can imagine a completely new interface to interact with the computer.
Yeah.
This is one of those things that is magical to me.
Usually generalizing auto distribution is bad, but somehow we have learned some kind of internal world model.
Yes.
That you say, you know, this plus, but it looks like rainbows and butterflies.
It'll do it and it'll kind of make sense.
Yeah.
So yeah, that's kind of cool.
Yeah, I don't know if there's any comment more on there.
I did want to touch a little bit more on the model architecture stuff which I think you were getting.
It's really fascinating.
We don't get a chance to talk about this enough.
One of the papers that we covered, we've covered every... annual segment anything release.
And I don't know if you follow... I mean, you're a computer vision guy, so you know.
So they did memory attention, which is kind of interesting.
And I always think, anything where you can across the temporal dimension keep some consistency.
I think it's very fascinating.
And I don't know if...
Basically, the CV side bleeding into video gen side, I think is underexplored, right?
We talk about it for labeling, but actually you can borrow the architecture itself.
And there's also complete different approaches, right?
Like you brought up the term world model.
So we went from video model to world model.
There is diffusion, but there's also other approaches that people are doing.
So maybe we get into those after as well.
Yeah, he has a whole definition of world models and stuff.
I feel like we threw a lot at you, whatever you want to comment on.
I think one thing that we should actually comment back on is like okay, so we were talking about the steps to train image gen to video model.
One thing we don't see as much of is like, okay, you brought up the delta in training data, right?
So You won't have as much.
A video model might not generalize, but what is the cost of training a large video model?
So we know for LLMs roughly, okay, even like the poolside thing that came out today, right?
It's a Gemma level model trained on roughly 40 trillion tokens at this many H200s over this much time right.
You can see what is the exact cost of that.
So how many GPU hours over how much H200 costs?
So how do we do the backend math of, you know, Same thing for video models, image models.
How do you kind of break that down?
I can share some back to the envelope calculation.
So surprisingly, video models is like the cost is very, is comparable to language models.
And obviously the largest scale is language model, maybe like a medium scale to language models.
I said, just storing the videos alone, it costs a lot.
You can maybe look up on AWS or something.
You really like.
Say, if you have a billion videos and let's just say like each video like five megabytes, then you will need like five petabytes to just store those videos.
And also remember we talk about you use a VAE to compress the videos and you also need to store.
Typically you need to store those continuous feature on also in your storage.
That's also comparable size with the videos themselves.
So just storing these videos and the features is tens of petabytes alone.
I just looked up the calculation.
Five petabytes on S3 standard is 100K per month.
Okay.
It's a double.
And you need, you know.
And then, like, a tensor petabyte is 200K.
And even more expensive is you have the ingress and egress.
Oh, yeah.
Through the internet.
You have to...
Just to download those videos, I believe it's more expensive on AWS than just storing those videos.
And each training runs, you probably need to pull them once.
If you try multiple times, it's even more than that.
So it's like just storing on the network those costs.
It's just.
I guess it would be a few millions per month to just storing everything, not to mention the GPU costs.
My side tangent, like, you know, compute rental, like GPU rentals, very efficient.
There's one side, okay, you can be XAI and build your data center.
Should we not just build our, like, storage compute as well?
Like, cloud costs compared to just, you know.
You save so much.
Yeah, exactly.
Especially with like egress and stuff.
So, you know.
That's a good idea.
But it also comes to, there are some of its own challenges.
Of course, of course.
Yeah, like people who build the GPU data centers, they might not expect this much storage.
And yeah, people build storage, they just build it somewhere with just CPUs.
I just looked it up.
AWS only charges for egress, not ingress.
Tier 5 for 5 petabytes is 230K.
Yeah, even more expensive than soaring.
But soaring is per month, right?
You check in and you cannot check out.
That's so cool.
It's okay.
So there's that side.
So tier there, you know, my... Data is larger than you think.
Yeah, my backhand math of GPU hours times, GPU costs is also very much.
You know I'm missing some stories.
You're basically like also more IO bound than normal training.
Yes.
Because like data loading, caching, everything becomes super important.
Yeah, so in Cosmos, we did a lot of optimizations to make it not IO-bound.
Yeah, speaking of actually training the model, the GPU cost.
If you look at the open-source model, how big these video models are?
I think, like LTX has 19B parameters.
That's a dense model.
And people are also exploring MOEs.
So it might be like a 20B active and... like hundreds B total.
So that's similar size as medium size LLM models.
And if you look at number of tokens, We disclosed that in Cosmos.
It's also like tens of trillions of tokens on the visual tokens.
So putting this together, the cost of training these video models it's actually comparable with OLMs, not to mention the infra is slightly different from OLMs, so it might be less efficient to train these models.
Do you get the benefits of traditional diffusion speedup?
So For, you know, images, there's LCM, for, you know, fine tuning.
There's a lot of stuff.
Flow matching.
Yeah, there's flow matching.
There's a lot of stuff that's been done.
There's some overlap that applies to diffusion on the inference side and stuff, or...
Yeah, so the inference side is a completely different story.
I think for the training side, it might be a little bit hard to reduce that cost.
And for the inference side, The biggest gain is from the dissolution of these models.
It's called step dissolution, slightly different from knowledge dissolution in RLMs.
So you typically for flow matching models, you need like a hundred steps or something like a distortion model.
Even need even more, like a thousand steps to to generate a good image or video.
A step desolation is try to learn to generate your step from the model itself.
It's kind of like now you use the full model to generate in 100 steps.
And then you take a model that only generates 10 steps and let that model learn from the perfect one.
Why does this work?
Strong to weak.
It's kind of like strong to weak.
From the modeling perspective, The strong model, the teacher model, is trying to model the image and videos of entire internet.
And that distribution is extremely complex.
But the step distill model is just trying to learn from the teacher.
The teacher is a model and the size is fixed.
The distribution is much simpler than the whole internet.
That's an intuition I have why step distillation can work.
So usually these models serve in productions.
They only run in a few steps.
In Cosmos, I believe we have four steps and eight steps.
If you do some simpler task like image-to-image translation, it can even run in a few steps, like one step in Cosmos transfer.
I think this is the same intuition that guides a lot of the consistency model work.
I sent you a link for SCM.
I don't know if you covered that.
To me, that was actually one of the most impressive papers I've ever seen from OpenAI.
This is the unifying grand concept of consistency models.
I don't know if you have any comments on this.
So there are a few different approaches.
Oh yeah, here it is.
Two steps versus 20 or 100 steps, whatever.
It's already done.
So there are a few different approaches, for example, consistency model.
And there are also...
Actually, we shouldn't forget GAN.
So GAN, actually, that was an OG step dissolution because it trained just one step to begin with.
So actually a lot of.
For example, there's a distribution matching dissolution which uses GAN as one of the laws for dissolution.
GAN just tells you, hey, generate an image, and then...
It has a discriminator to tell, is this image real or not?
So the model just needs to learn one of the distributions, not the full distribution.
Because in training the model is asked to reconstruct the ground truth image from the internet, which is extremely hard.
And when you're training GAN, it's a one-step process.
It's just, hey, you generate an image.
Does this image look as real as the image from the internet, which is a much simpler task.
And Combining a lot of these approaches together.
People typically do that like consistency, model and distribution matching.
And again, we can get these few-step models.
Okay, then there's one step I wanted to add, which is audio and video.
Yeah, so Garkey Imagine 09.
I believe it's the first audio-video trans model deployed at a large scale.
And that was your first model?
Yeah, that was GrokImagine's first model.
It's audio-video joint generation.
I think the hard part is the modality alignment.
Because before this trend model, we have text-to-video alignment.
We have this correspondence between text and video.
Typically most of the VLMs.
They understand images and videos very well and they don't understand audio mostly.
And if you look at the audio generation on the LM side, you can talk to them perfectly fine, but if you ask them to sing a song or something, it typically is not very good.
Also, they don't have music either.
The hard part is that Actually, audio has two components.
It has like a discrete component, a continuous component.
The discrete component is like there's a language.
So when we speak, it's just some... It's an ASR issue.
Yeah.
Yeah, it's text token with some characteristics, I would say.
But music... I think the speech guys will disagree.
It's like disfluencies and then, you know... I say largely.
But the music is completely different.
It's...
It's very continuous and you cannot model them like discrete tokens in language models.
This is like.
The hard part for models is not to mention.
We have to align text, video and audio together.
Yeah.
How?
So significant.
Some significant challenges are like.
So first, like we talk about as VLMs, they cannot understand.
Most of them cannot understand audio.
So you have to have some way to do the synthetic data generation for audio.
For audio you have to caption the model and that involves data and human data effort a lot.
And not just surprisingly.
Most of the RLMs are very bad at recognizing the beat, tone and the details of music.
They can give some general prediction of which song is this, but it's very hard to describe the details of the music.
Like we mentioned in image generation, you have to describe image as detailed as possible so that someone blind can reconstruct that.
So here it's like someone deaf can reconstruct how the music sounds like without without actually listening to it.
Maybe, like you can think of it, need to have have the, what they call the, the skittles.
Yeah, You've got to have all the details of the music and the dialogue.
So is the challenge there typically stuff like music and audio, or is it just like is there a baseline?
Okay, there's enough data where we can understand.
You know narration conversation, but there's nuances in audio.
That that's where you hit all the data issues, or
Is it just from state zero, you just do it all right?
So one important thing is like the alignment.
So the model has to know, like the video and audio, it has to have a time-based alignment, like at which time step the video and the audio token correspond to each other.
We actually don't have this kind of alignment for most of the other modalities.
If you think about text and image, text and video, they are loosely aligned.
So you can have a description of what's going on in the video, but you don't have to exactly.
Uh, you typically don't have exact description or at at a time, step one second, like what happened.
It's very course.
Yeah.
Yeah.
So what was the ideal time step?
That you have to ablate it and then it's like four seconds or something.
So that comes down to how you design the model for the model to be aware of as a time modality.
So the model is like a time-aware.
And that's something pretty unique if you think about LLMs.
So if you ask LLM to complete a task, So you ask them and they will say oh, this task will probably take 12 hours to complete.
And they come back in one hour and say I've already spent two days on this and I've exhausted everything.
Yeah, so the OLMs themselves, they don't have a sense of time there?
I actually don't think that's just them not having a sense of time.
I think it's somewhat based, right?
Like you tell someone, okay, go work on this feature, go implement this.
There's a general understanding you would have of how long that would take without LLMs working at LLM speed right.
So you think back like two years ago.
If I tell you to like, build me like a new front end for latent space, have a search bar, have all this, you'll estimate that it'll take a few days, right.
So you tell an LLM, go build this, It'll take me a few days.
But you know, I think it's somewhat grounded as opposed to them not having the best, not saying that they have a great understanding.
But I think that example is like.
You can see where it comes from.
You're trained on all over the text.
They're trying to estimate what a human would say.
Yes, because that's what the data kind of represents.
It came from the corpus on the internet.
People have an estimate.
Yeah, and not even just in direct training samples, right?
Just your world understanding of tokens, of how long stuff takes, right?
Go read a book.
It'll take you a while, right?
Even if you do nothing but read a book, it takes a few days.
So yeah, I'll let my Reddit, it took me a few hours.
It'll take me a few hours to go through this research.
But this is a tangent.
So I want to, yeah.
This is a train of thought I haven't really expressed until now, which is basically like a full world model must also be recursive, meaning that the participant in the world model must also be aware that they have a world model which is like this whole recursive thing down the line.
But yes, and that the world model can be wrong and that they need to update it and blah, blah, blah.
Yeah, we've argued this on the newsletter as well, that there needs to be sort of recursive adversarial models.
Okay.
I mean, just, you know, to ask, how do you define world model?
Oh yeah, let's go there.
Yeah.
So, just for context, you know we talked about video generation, and then there's a if you say there's a distinction between world models, what's your, what's your definition?
How do you see the two?
Yeah, so disclaimer, I'm not going to debate what is word model.
There are many definitions, so I'll just talk about my definition.
Since I came from the multi-model domain, so mainly talking from video, so word model is like real-time, interactive long-horizon videos.
So there are three parts.
So let's talk about them one by one.
So interaction...
So we just look at Facebook and the neural computer.
So the interaction part of it, so your work model can allow you to interact with them through keyboard, mouse and maybe also voice.
So all of these modalities you can interact with the model and the model should respond reasonably.
Second part is real-time.
So once you move your mouse, say the world model generates a game, how fast can that game respond?
So if you're like professional CSGO players might say oh, you have to respond in sub 10 million seconds or even less.
So that's not, I guess, most of the... No, 60 FPS.
Let's go.
300 FPS.
No, 500 FPS.
Wait, okay, yeah, I didn't do the math, but yeah, okay.
Yeah, 300 FPS, that's 3 million seconds.
So you have to respond within 3 million seconds.
Most of the video models cannot do that.
Yeah.
And.
But if you say, if you have a video model, that is say, like a digital human, the response time might be more generous.
Maybe like typically like for real time voice interaction is like 200 million seconds.
So that's, that's much more generous.
But even 200 milliseconds is pretty tricky.
Because like, remember we mentioned, you have this temporal compression coming from the VAE.
So if you don't compress the temporal dimension, your sequence length is going to explode.
So if you want to have this real-timeness in your model, you have to deal with one context problem.
And the third part is long horizon.
Because if you're not going to just play with video games just like a few seconds, most of video models only a few seconds.
We're going to play with minutes, hours.
The model have to be able to generate long form content.
So putting these three together, it's a real-time, long-horizon, interactive videos.
I think the final state will be, for example, like a video version of Playbook, where you can interact with a neural computer.
You move your mouse and you click on the generative interface and it will reply to you through pixels generated in real time.
But getting there, it's a very long way to get there.
So one of the first steps at Graphic Imagine, where I led a small world model team there was, was to build video extension.
So video extension ah, it's the first step of interactivity.
It's the first step.
Yeah so, it's the first step.
You have it here.
Video editing yeah yeah, Yeah.
So the first step is because this unlocks long-horizon videos.
Typically for most of the video generation models, you give it a prompt or an image as an initial frame.
You generate a video, that's it.
That's just one time done.
And some creators would try to use the last frame as the first frame for the second video.
Sometimes it works, but if you do it a few times, the quality will decrease.
It doesn't have that context over the full video.
So the temporal Yeah, because you only gave it the last frame, of course, right.
Yeah.
Exactly.
It's actually a pretty fun hack.
Like, if you've seen, like... Oh, no, he has something better.
Yeah, yeah, yeah.
And, for example, like, if you...
I remember Vue 3 has a one-second context of the last video.
It's slightly better than using the last frame, but it has a similar problem that it the quality would degree like if you extend a few times to like one minute.
The video quality would look much worse than the first video.
Second, another problem is that the model doesn't have long range knowledge of like, what's happening before.
So if they generate some dialogue to people speaking And their voice might change over some time, especially if the one-second conditioning does not cover the previous contacts.
So these are the core challenges.
So the Galaxy Imagine video extension, it has historical context of all of the previous generation videos.
It has a context of who is speaking and what objects have appeared and everything having that to generate the next video.
So if we naively do this, you can imagine like just put all of the previous history video tokens into the context.
The context lens will easily explode.
Especially for video models, that can be like a few million context, I would imagine, context lens.
Yes.
What's wrong with that?
Yeah, for example, like in Cosmos, I think just five seconds of video is like a 50k or 60k number of tokens.
So like if you do 50 seconds as 500k tokens, if you do longer than that, easily explode.
This long horizon problem was the first step we were trying to solve word model.
It turns out people love video extension.
A lot of the creators love using video extension to create longer form videos.
This is the part I like that you have an intermediate step towards the final goal.
Instead of just a straight shot to the final version very much.
Yeah, but I can see you have a strong vision of where we want to end up.
Yeah.
Does it seem like it's an efficiency issue?
Like...
Okay, we're at a few million tokens context, you know.
If you draw the parallel to language models, we had very short context, 2,000, 8,000.
Then, you know, you scale it up, 1 million, 10 million.
Sure, there's effective context, you know, But at the end of the day, it's just what's it worth?
Like, sure, there's a whole training data side.
In video, it might be slightly easier because we have 100 million token video.
I just take a movie with the full context there.
Like is this efficiency from an inference standpoint?
That like, it's expensive, but we know how to solve it.
Or like...
Why is this not the approach?
So, like my broader point was on your second point of world models, you say it needs to be interactive and live.
Right?
You should be able to play a game and see the interaction live.
So one thing I see with research is a lot of what you actually serve is different than what you build right.
So we talked about distillation.
You train big model, you distill it, you do quantization, speculative decoding.
We do all this stuff to serve it efficiently.
Should we not just have a solution like a world model that can interact well?
Do inference optimization serve it, distill it?
Secondary
So make it real time after you solve it.
So like another parallel is say continual learning, right?
What we need is someone to solve it and show it works inefficiently.
Give it a few years, people will make it efficient.
Same thing with regular attention, right?
It worked over a few years.
People have different forms of attention and we've scaled it to be efficient at long context, you know.
So kind of two things there, right?
One is like, It seems like it works.
You've scaled it.
Can we not just scale it a lot more efficiently over time?
Do we need a separate approach if this works?
And same thing with interaction, right?
If we can get it done like, if we can solve some way that it works.
You know, we can solve making it more efficient from an inference standpoint later,
Yeah, that's actually a very good point.
So in videos, there's actually a lot of redundancies.
So we solve a lot of the pixel redundancy from VAE.
But there's more redundancy in long range and long horizon videos.
Say, if a character appear in the first clip and then it disappeared, it only reappear like as the end of the video.
You probably don't need the context, like in the middle of the generation.
So you only need that character where you need.
So that's why I helped build another feature.
It's a reference to video.
Is it here?
Is it the same model release or different one?
It's a different one.
You probably need to search on X reference to video.
Okay.
So reference video allow you to upload up to seven images as condition and generate a video save.
It can be characters or objects or even scenes.
So I want condition on Sean's selfie and the holding.
Holding a blade or whatever.
Yeah, we have a dog.
We put the dog in the thing.
Yeah, you can put them there and the video models will generate the video and copies the context over.
So that can solve a lot of the problems there.
Like the long context problem, it doesn't need to have a very long context, but I feel like it's an intermediate solution.
It's cheating, yeah.
Yeah, the model should... be able to selectively know where should I draw references.
So say, if I want to generate a movie, I generate it autoregressive, like 10 seconds at a time or something.
And now this character appears, I can look back to where it first appeared and bring that back.
Yeah, this one.
I put the references.
Yeah, that's Optimus, Einstein, myself, Annie.
Oddly enough, I used Grok search to find it and it pulled your LinkedIn post.
But you know, we found it.
Interesting.
But okay, this is a problem.
This is not your fault but like, XAI doesn't communicate all this work that you do very well, because they just have the model release and then that's it.
But like, actually these details are very, very good.
As far as I understand, everything you just described is state-of-the-art, like no one else has done it.
Thanks,
A lot of, yeah.
And then you just put this blog post with the cookies.
I'm like, this is not enough, you know?
But obviously this is like the high level numbers that people want to know.
And I wonder you know like part of that is also like some labs don't share research, research into what happens.
But this is literally bragging about how good they are.
Right?
Why would you not say that you are capable of extending with full context?
You know, this is not a secret sauce.
This is like, we did the work.
Like, yeah, I don't know.
Yeah, I guess different labs have slightly different communication styles.
Yeah.
Anyway, if anyone from XAI is listening, we are always happy to help you tell your story.
Yeah.
Okay.
So you did references and I think I think kind of the point you're making is like it is sort of like a kludge, right.
Like this is, you can do seven, but what about 100?
Yeah.
Right.
Then you need a completely different thing.
So I think this is like a mechanism to select the context from the history, and you might not put the entire history into the context.
For example, there's a paper called Frame Pack, which have a heuristic that the latest history like the last one.
Second I put the entire history and the history before that, I would compress it and make the video smaller.
So I follow this pattern this buildable pattern that the maximum sectional salience is fixed.
So the further you are from the current frame, you have a smaller image.
So this is just a heuristic.
I think it can be more automatic.
The model is very like which history part of it can be select.
So this part of the research is actually being actively researched, uh worked on by a lot of people.
It's also quite interesting, i feel this is actually this part of long contacts is a little bit ahead of the lm part.
So, for example, again in lms, if you the contacts keep growing, let's say if you you call tool and the tool call history is extremely long that's still in context And keep growing, keep growing.
Even if you switch the topic to something else, the whole context would there.
There are some agentic harnesses that help you to say prune the tool results and prune like when you query a file, only show like the top 200 lines or something.
Those are very heuristic driven.
For listeners.
We did a write up on the cloud code where there are eight different kinds of pruning including, like you prune, the two results and all that.
So you can read up on that kind of thing.
Yeah.
I think one breakthrough in container learning might be like a way to automatically manage its own context.
These are all heuristics and they will be replaced by machine learning.
Yes.
Interestingly, the same thing is being researched in both LMs and video models.
The interesting thing is also, like in the paper you showed, it's actually happening at the model level, right.
Compared to like, like with models.
Sure, we have base attention, but you know, we'll do our own compression, we'll do our own pruning, which is separate from model error.
Eventually it all just boils in, hopefully.
I think this is a form of attention, but also reasoning attention.
I feel like that's different than normal attention.
Does that make sense?
Yeah, yeah.
It's different in the sense that attention, not to mention set sparse attention aside, like normal attention.
Like UKV.
You have to attend to all of the tokens.
Yes.
So you don't have a high-level mechanism to drop which tokens you don't want to attend to.
As humans, the potential spend is surprisingly small.
You can only remember an 11-digit phone number.
But I have feature detection, right?
I can detect, oh, that's a sequence of one, two, three, four in a phone number that is 11 digit.
Very good pattern matchers.
But humans' contacts can like attention can work because we can dynamically pull in contacts from different places.
The same mechanism, I think, is going to happen for RLMs and video models.
Yeah, RLMs is one of the recent work that's there, which is not that crazy, but it's just recursive.
I think it's somewhat inherent in models too, right?
Here's a nice example.
You pull up these, you can read it fine, but language models are also very good at slot parsing.
Yeah, I throw my typos in there, it doesn't matter.
Yeah, yeah.
You have a transcript, you have whatever, just throw it in.
And it's very good at parsing through noise.
You know, that may be a brute force.
It can look over and reason over it.
But like, you know, there's parallels to both.
I think it's just really fascinating how you relate the world models stuff to the video generation, which I don't think a lot of people hear directly from people like you.
So I think that's really helpful.
Any other work?
Do we cover like video, audio, world models, any other stuff in that Omni team, I guess?
Or any other work at XAI you want to talk about?
Seems like everything we see publicly announced, oh, cool, cookies.
And then there's so much more to it.
There's a lot of that.
Any underrated stuff, you know, just at the time there?
Yeah, I feel the culture is quite interesting and a bit underrated.
So the culture is... The culture is...
It's three sentences.
Move fast.
Build.
No goal is too ambitious.
And the first principle.
Like, usually the goal set was very ambitious.
It wasn't very... It wasn't possible to achieve when I was first thinking about it.
Like, for example, I could build... build something in three months.
And was that like, okay, we're starting team, we want image, we want video, do it by this deadline?
Or, you know, how do you work back?
Like, was it just, okay, we have a rough by, you know, this date, we want something out?
Or is this like...
That's a very good point.
So it's from first principle of thinking.
If you think about people might say like first principle of thinking applied more to the physical world than the models.
I would say, for example, if you think about some limitation, for example acquiring data, how fast can we acquire the videos?
And if you think about training the models, what's the iteration speed for training a model end-to-end?
And how... how would adding more GPUs accelerate that timeline?
And maybe if you need human data, like what's the turnaround time for human data to arrive?
If you put all of those together, that is first principle thinking where oh, you know, like what is a timeline with a minimum number of days that is possible to achieve something.
I think this is a lot of Elon's type of thinking, right?
He's like.
I think he's famous for saying that the only law you can't break is the laws of physics something like that.
Yeah.
Just broadly, you worked a lot with Elon.
Yeah, I guess one benefit is like working at XAI, you got the chance to interact more with Elon.
So I was very fortunate to get a few retweets from him.
That was quite fun.
And he also worked very closely with people.
Like people imagine online, like he's very hands-on.
There are two things.
One, so I was actually looking up Elon retweeting you.
I'll pull it up.
He talked about you tweeting that you have a really good voice mode.
I don't know.
Oh, I also did it.
I actually, so I would DM you feedback on voice mode because I was like, wow, really good.
And I'm like, oh, this sucks.
But I don't know anything you want to talk about, about your voice mode building.
It was a team you worked on as well.
But that's actually not a part of the team that I worked on.
Yeah, he probably worked on more of the video.
No, but Grog Voice actually, like, very good.
There's one of those things where like, first of all, you can speak at 2x, which is fun, which I listen to 2x, so I like to speak at 2x.
But also I think, like, the interruption was better than Gemini.
I don't know how it compares to ChatGPT real time now but, like you know, as far as like driving was concerned, like having Grok in my Tesla and like driving, I think it was like it's a really good experience.
Yeah, he likes voice mode.
But also, just the crazy reach value.
50 million views are just saying, yes, true.
Oh my God.
But, you know, it's pretty cool how fast it came out.
I guess the other thing is the safety aspect of video mode.
Anything interesting to talk about there?
Spicy.
Spicy question.
A lot of the countries where they don't allow generative data, generative AI videos without watermarks.
So in all of those countries, Croc Imagine had watermarks.
And a lot of the...
But also takedowns of the videos were also happening extremely fast.
I mean, it's part of running a social platform, but also it transfers nicely to the Gen EI side.
Do you have a perspective on SynthID versus other kinds of watermarking?
Yeah, I guess it's going to be...
Yeah, it's going to be harder and harder to detect these things.
So since ID, one thing is previously it was only Google.
And now, a lot of different labs are also adapting it.
A limitation is the technology.
The paper was out there, and people can reverse engineer how to get rid of it.
And I think even as it advances, it's still... still possible to reverse engineer it?
Yeah.
So if you are interested, you can go onto Reddit and people have taken out the exact like I don't know what do you call it?
Mask or pattern that Google applies.
And then you can apply it onto any Google generated photo and you can reverse out the synth ID.
Yeah.
And it's also harder and harder to just judge by eyes.
I remember like a couple of years ago, there were like six fingers or something.
It's very obvious.
My current is actually the audio.
I feel like the audio is really lacking.
My way to tell if something is AI-generated outside of like.
Okay, I think I've seen enough.
I have a decent eye.
The audio matchup, especially of Sora, is not great.
It's all similar style.
Those are minor imperfections.
I think the point is that like, actually my closest reference to this is also Ian Goodfellow.
Cause I think he did like the adversarial game thing where like it's like okay, here's a picture of a zebra.
Then you like change one pixel and it becomes a panda.
Right.
It's like a classic computer vision issue.
Yeah, if you think about how these models were trained, like I mentioned before, GAN was in the training process.
The objective of GAN is... the model generates an image.
And the model, there's a judge to tell if the image is real or not.
The model is trained to make the image more real.
So as the model becomes more and more advanced, it's going to be harder and harder.
For me personally, now I have to judge by if these videos have logical sense.
If these videos... Have a world model.
Yeah, yeah, yeah.
No, I also like the audio is too nice, too studio quality.
The lighting is too good.
The skin is too clear.
You know, basically the lack of imperfections.
Yeah.
Do we have a good way to do reasoning in diffusion?
Is that what separates video generators from world models?
We really know how to apply it to autoregressive language models.
Is there a parallel for diffusion, video gen, world models?
On that point, right?
He has a thing on video agents.
Yeah, that's a good question.
Yeah, actually, I have a pretty big claim.
The visual intelligence are actually mostly coming from language.
Like these video models, Especially from now, since the diffusion model technology is more mature.
Every time you see there's some improvement on these models.
I would say mostly this again comes from language model, not coming from the video model itself, like the video distribution models themselves.
In Cosmos, Typically, these models, they have two parts.
There's a prompt rewriter or the prompt app sampler part.
I think in...
In Cosmos, we use Lama or we use Mixtro.
And the Cosmos video model itself is only 7B.
And the language model is a prompt rewriter.
It's bigger than that.
So the prompt rewriter's task is to take, take user instruction and convert it to extremely detailed description of the video.
So because the video is a visual, the video distortion models I would describe.
They're kind of dumb because they take the input instruction literally.
Because in the training process, remember that we have to describe the video as as detailed as possible when we are creating the synthetic text pair.
So this model, they take those kind of instruction to generate the videos.
So when you're taking the user instructions, the user instruction is really very simple.
Just say a cat or something.
If you put a cat in In the video model, they would take that instruction literally.
They would literally show a cat in maybe a white background, because you didn't describe the background.
The cat is not moving because you didn't describe it.
It takes the instruction quite literally.
It's kind of dumb.
The prompt rewriter is actually a much bigger model, which is a language model that takes the user instruction and expands it.
So the thinking process you mentioned is from there.
So if you look at a GPT image like, you generate an image in three minutes.
Certainly it's not all like a pixel generation.
A lot of time is spending in thinking.
So prompt rewriting now have evolved to Like not only just thinking, it can also be an agentic model.
For example, say, you wanted to generate the image of today's news.
So the...
So it's likely they'll go to fetch today's news online and then process them, digest them, then organize the layout and generate it.
Another thing quite interesting is, If I'm not mistaken, it's no longer a diffusion model though, right?
It's autoregressively... Or is there still... There are different approaches.
For example, like Chaminade Omni.
Since they said it's Omni, I believe it's a single model.
Maybe it's something like, it's a language model with a diffusion head or something.
Like the language model, do the thinking, do the agentic tool calling, and then it would use the diffusion head to generate the image in the end.
There were also approaches like Cosmos, where you have a separate language model and separate diffusion models.
And there are also like a purely language model, like you discretize the images and then you generate the image as discrete tokens.
So there are different approaches.
One of the claims I've seen for why these approaches struggle is because a lot of the benefits for how we currently learn reasoning with language models is you basically iteratively generate reason.
You have your thought and then you... work on that answer, right?
So if you have like Omni model and then diffusion head, you can't feed that back in.
To continue reasoning, right?
So you can't go like text image, text image.
You can't reason on the output and then go back to diffusion.
But I guess in the new Gemini Omni, you would be able to as long as you have...
Yeah, I'm not sure if they have that process.
I guess it's definitely possible in the Omni paradigm.
So if you think about traditional multi-model language model, they would have a VIT encoder that can encode the image.
So if they have a diffusion hat, they can generate the image and then put that back into the VIT encoder, encode that and then do the iterative refinement of the result.
I think you have to jointly train the VIT and the diffusion to make that somewhat feasible, because otherwise you're kind of like mismatching or feeding in slop.
Yeah.
I think it depends on the stage of training.
You might be able to freeze it.
But anyway, also just on your earlier... I wanted to also make explicit.
We do know that nanobanana and GPT image are autoregressive language model with diffusion head.
As far as I can tell from your description of grok image, it is not.
It is end-to-end.
I cannot comment on that.
The way that you described it.
But I think there's different approaches, right?
Like you started off saying prompt rewriter is like a big part of the intelligence.
And even on that, I think everyone should try using an early diffusion model.
If you've used stable diffusion one or whatever.
If you've seen the prompts, like you know, ultra high res, 4K.
This style like, oh my God, the first time I tried one.
You don't talk to them like language models, right?
Your prompting is very, you know, comma separated.
You're literally talking in the labels that were in the data set, right?
Yeah.
But basically I'm just trying to make the point that prompt writer and then image is different from autoregressive language model with diffusion head.
They're different things.
Yes, they're different.
I just wanted to establish...
I feel like the common part is the image part.
So it's quite surprising that a lot of the improvement came from the language side, the thinking, the tool calling. so I still remember like in Cosmos I I generate a happy sheep and can if without any rewriting it's it looks so CGI and after rewriting it looks it looks so beautiful I think without any joint training Yeah, actually, without any training.
With rewriting, it's already much better.
A very interesting thing I guess will happen is video agents, mostly language models, will call these generative models.
Either it's a separate model or a diffusion head or whatever as tool.
So this model can iteratively refine the results or even like, generate longer content through a very long trend of thought.
It's actually very similar to how human create art So we don't generate the pixels directly.
We literally draw something.
And I think through this process, these models not only use diffusion as one of the tool.
It can also use traditional tool.
It can also use image editing tools from Photoshop.
It can use a video editor FFmpeg whatever, to take a combination of these and the generative AI technology as a set of tools and they can iteratively create a much better video for production-grade quality.
If you look at existing professional creators, they don't end at generating a video from these models.
They will take this video to their editor and edit here and there.
So much post-production.
And sometimes actually the reason the video is good is not really the video model, it's actually the editing.
And yes, we also are engaged in the same process as well.
I would love to use a video editing model.
Yeah, actually there's a Grok Imagine agent beta.
That was the first attempt in that direction.
Yeah.
So I think the process would be similar.
It's just agent mode.
Yeah.
Yeah, you can ask it to maybe generate a one-minute video, which is not possible if you ask the same prompt to video models.
But this model will literally call different tools to do that.
So, yeah, this is actually an interesting thing.
So when they first released a video editing model, like I see on X, some people try the video editing feature with like edit this video to be one minute because they didn't understand how video editing work.
Video editing typically is just a removal, add, replace, style transfer, this kind of thing.
But that's actually a valid request under the assumption of video agents.
So these agents should be able to understand these kind of long-horizon tasks, to be able to iteratively create a long-form video.
I think this is really fascinating because it's kind of taking the same direction.
As first you have these.
AI-assisted coding, kind of like tab completion, GitHub code palette.
And from there, you gradually evolve to Codex and Cloud Code, where you do things fully automated.
So in Gorky Imagine Asian mode, you can still go in there and do stuff by yourself.
Gradually, as the model capability increases, it will be able to do everything fully automated.
Yeah, I like that.
Okay, so it looks like it's still generating.
Also, I did notice the Croc image gen was always very, very fast.
I don't know if this is something you guys benchmark, but this is just a tangent.
Compared to when I used to use before the latest OpenAI's image gen and same with Gemini Nano Banana I would oftentimes use.
Crawl.
It's in the benchmark somewhere in the Imagine API blog post that they have all the speed things.
Mostly combination of distillation plus inference.
Yeah, there are a bunch of things.
Like we have talked about distillation.
Then we have talked about thinking.
If you don't have any thinking budget, the model can just think three minutes and then come back to you.
And also, like, inference.
The inference infra team was very talented and they were able to accelerate a hell of a lot of these models.
Yeah.
Yeah, I mean you know my comment on the video agents, things like I'm trying to figure out, like when people say video agents.
When you initially told me about your bet on video agents or your vision for video agents, I was a little bit disappointed.
I was like you mean, like models are tapped out, now we have to do agents.
But like I think you have to right the question now is how much model training is it really going to make a difference versus just building a better harness?
Like you said, the models don't have to be jointly trained.
You can just take an off-the-shelf frontier reasoning model, slap it on a harness, give it grok as a tool.
That's it.
That's your video agent.
Doesn't seem super satisfying.
Obviously, you can co-train and get some more... percentage points of performance.
But, like if your central claim that the majority of video or generative media alpha or whatever, is actually coming from language intelligence and not image diffusion or video diffusion, then that is the future.
Primarily just wait.
If you pop back at the example, you know it generated frames.
Sorry to interrupt.
You know it's been saying like okay, I'm going to start stitching these frames together.
It's using FFmpeg.
This is what GPT Image Pro as well is doing, right?
Like it's also just writing code in the background and then just doing an image pass on the final output.
It feels dissatisfying for the people who want to just train models.
It's interesting, right?
It's also somewhat exciting.
Like you brought up earlier, a lot of the gains don't come as much from the video.
I think you can see that in the language model space too, right?
Anthropic, very, very good at coding.
Their multimodal, not the best, right?
They have basic input PDF but, like you know, there's clearly a disconnect in the quality of their image.
Video processing, audio processing, right.
Yet, intelligence, very top tier.
Other labs Gemini OpenAI, XAI you can add modalities, but it's not like they're unlocking crazy capabilities, right?
So it's interesting.
Yeah, it's interesting to see that, because the video model's capability increase actually come from language model being more intelligent.
I think video agent, it can unlock more stuff than you might imagine.
So there's a few things.
So one thing is When we are prompting these models.
So most of the people were actually not very good at prompting.
Actually, language models have a better sense of how to prompt AI models.
AI models know AI models better.
So if you jointly train these models, maybe as a model have a better sense of how to prompt each model.
Like a different model might be different.
Another thing is It might not as simple as just generate a few clips and slap them together using FFmpeg.
There might be more image and video editing to appear in this process.
Say, if you want to exactly add a... add a blob of text at this time step.
The video models might not get that intention very precisely, but these are possible using these deterministic tools.
The video agents can use all sorts of tools, so you don't have to put all of the capabilities into the transition model itself.
Yeah, I think that's very true.
No, so for what it's worth, I think you're right.
I think that this will be a big category.
I think probably you are predicting like the next one year in video is going to be all this.
Do you have a time prediction for how when this stuff ramps up?
I mean, they already started.
It's not very good yet.
No, it's so good.
I think the last one is just longer.
It didn't give me a minute.
It gave me 36 seconds.
But, you know, are we feeling it now?
Is there going to be inflection?
Is there any timeline predictions you want to make?
I guess by the end of this year, this is going to be a big hit.
So the inflection point will be there.
The videos generated by video agents can get to production great qualities.
It can be presented... and it can be distributed in ads.
And once that happens, I think the enterprise will have much more budget for video models, because the agents are inherently more expensive than the other video models themselves because they do this iterative process.
They generate many, many variations.
Yeah, but once these models pass this usability threshold, I think it's going to be an exponential growth beyond that.
I would fund a company right now based on this.
So I think you're right.
One thing I'm surprising, I'm reflecting on the whole past hour or so conversation.
I think you're into world models and video generation for video generation's sake.
I think that a lot of other world models people, we've interviewed a lot of them.
Genuine Tuition and Fig Fig Lee and all those guys and Moon Dream, which I think I told you about Moon Lake.
I keep saying Moon Dream, god damn it, Moon Lake.
A lot of them actually say like robotics is the end game, like embodied robotics, like you want real time, you want interactive.
It is to interact with the physical world, you're not that concerned about it.
I think robotics will be a big part of it, for sure.
I guess the process might happen naturally.
So my prediction on robotics is that the problem of physical AI might be solved without actually needing to Be in the real world.
Need to be in the real world.
So it might get solved by a video.
LLM is virus-drawn video capability.
So remember we talked about the real-time interactive long-horizon video?
So now these models are just training on screen recordings and computer screens.
Once these models can use computers and understand the future state of computers extremely well, The robots might be one of the tools a very powerful AI can use.
So the powerful AI might just be able to control the physical embodiment naturally.
I see that for sure.
Cool.
I know we are coming up on time.
You left one more spicy topic, which is why you left XAI.
For me, there's a lot of research you want to do that you cannot do as a company.
And also the priorities and objectives of a company typically can change very fast.
It's also the same for XAI.
So now is kind of like the time to do some research.
I want to do especially more on language model side.
I cannot do at XAI.
Oh, okay, yeah.
You're basically leaving.
You had this whole transition from computer vision to world models, video generation to now you're focusing on LLMs.
But it seems like you know a lot of you saying focusing on LLMs.
You really in the past hour described how it all ties together, right?
Yeah.
But I don't know.
What do you mean by focusing on LLMs?
Is there...
I realized the fact that the video models, even like in the beginning the game, might come from improvement on diffusion technology.
But this is a point where actually most of the gain come from the language models themselves.
It's a huge black pill for anyone who has spent their career in generative media.
I mean, that's an extreme view, right?
You still definitely need a bit of both, right?
There's just, it seems like more pressing, impactful work to do now on language model side.
Do you have any similar predictions?
So you predict the video agents.
I think you'll be right.
On the language side, what are you looking for in the next one year?
I think one thing pretty interesting I think might be happening soon is the language models will be context-aware and manage its own context.
So some...
Like from the video model side, we've been suffering from this long horizon issue.
Like we want to generate video longer and longer.
And we've been trying to solve the context lens issues through various ways.
One thing is just brute forcing train longer context lengths.
Another is to manage the context better.
I think the same thing in language model is also going to be happening soon.
So for example, like the language models, They're not aware of how long their own context length is.
Once they hit like 80% or something, the automatic context compression is getting triggered.
And the model is not aware of that when it's working.
And maybe it's good for the models to know, oh, I'm approaching like 80% or something.
And something also pretty interesting for example, in OpenClaw, Every time you type in something, the current local time is automatically attached to your message.
So the model actually knows what time is it.
This is making the model timer there.
And also in tool calling, a lot of the intermediate tool call results automatically prune.
So there's context removal, context addition, and context compaction.
So all of these are from the harnesses themselves.
From our experience, the heuristic engineering also have the models get absorbed into the models themselves.
I guess that's something very interesting to explore.
So infinite context?
Maybe.
No, but it's interesting, right?
It is in the space of memory and continual learning.
I don't know.
It's also like in the space of agent harness use, right?
He's saying he doesn't want to do it in a harness, right?
No, no.
But models are also being trained on using harnesses, right?
Yeah.
Some of it is, you could say, implicitly leaking in, right?
Part of that post-training of language models is okay using it in coding harnesses, In which case you know when are sub-agents spawned.
When is convection going to happen?
It's not explicit.
Like you know, you have this much token window, which I don't know if you want it to be, as that'll change, but it's somewhat leaking in there.
I'm imagining what if the model have access to the code of the agent harness itself and be able to modify it to whatever you want.
Say if the agent harness is short enough, you can just put in the context in the system prompts.
And then the model will say when I want to spawn a future version of myself, I can modify the agent harness.
For example, the agent harness can be when I'm reading a long document, I can choose to read the whole thing in chunks and come back, smash the summary together.
Or I can just read the first 200 lines and discard the rest.
And all kinds of choices if they can be made by the models themselves.
It might be very interesting to see that the model can program itself online in test time.
Yeah, so the self-modifying harness is also part of OpenClaw and Py, but I think there's a lot more work to do there.
Very cool.
I think part of me...
It's kind of curious.
I think you are a part of a big lab, right?
And there's this career path of a researcher at a big lab, which is you train models, you get more compute, you train better models and you keep going.
And somewhat, I feel like you're opting out of that.
And if I were you, I'd be like, oh, I think this is a bit of a career risk.
You know what I mean?
I don't have any comment apart from like, you're very strongly convicted.
I think that a lot of people in your shoes would not be doing what you did.
Yeah, speaking of my career, if I look back, actually there has ever been.
There were a lot of huge transitions.
So 10 years ago, I was doing research with the ResNet authors, Xiangyu Zhang and Jian Sun.
Yeah, at that time, the research were completely different.
It was, like, mostly conservation, like, image recognition, object detection, object tracking.
I was also doing neural net compression at that time.
It was quite different from knowledge dissolution these days.
And at that time, I wanted to be a professor and I applied.
When I applied for a PhD, I already had a few first author papers at top conferences.
So I confidently applied to the top schools.
But turns out I got rejected by all of the top PhD programs.
So I had to go to the industry.
At that time, I was at Facebook hair research fair led by Ian LeCun.
I want to talk about Vijapa, but it's different.
Yeah, we can leave it for another time.
Yeah, I switched to... At that time, I switched to self-supervised learning.
It was quite different from what I was doing in contribution.
And after that, it's...
NVIDIA, Cosmos.
So I realized scaling up was extremely important.
So at NVIDIA, I was mainly focusing on scaling.
So one thing is Cosmos, scaling the video distribution models to two billion parameters.
And another thing is I was working on MOEs.
The Megatron MOEs was the first framework, open source, to be able to train these MOEs at very large scales, like 100 billion parameters to even trillion parameters, efficiently at like 40 MFU.
And going to switching to xai, was trying to work on an even larger compute scale, even further.
And yeah, looking looking at this trajectory, i actually worked, worked on a lot of different things.
So i feel actually within, Within ML, it's actually easier to switch than you think.
A lot of people might have mindsets that, oh, I work on computer vision.
I always have to work on computer vision. and I cannot switch to language.
But from my experience at least at NVIDIA I worked on both language model MOEs and also video models.
It's actually not the case of others.
But also the core principles, how to train large models, are largely the same.
And...
Yeah, for me, I feel right now, the bottleneck for video models is actually the language part, the agent, which is why I want to go to work more on OLMs.
One thing is it's a bit of a challenge.
I don't think it's a huge jump.
Yeah, I mean, kudos to you.
I think you have a lot of strong vision there.
Yeah, I think that was mostly everything that we wanted to cover.
You've been very generous with your time and it's really nice that you are able to share all these things now.
We don't have to go through XAI to clear everything.
But also, I think we didn't get you in trouble.
It's a lot of good stuff about XAI compared to what you just see in the releases, right?
You don't realize how many more levels there are to it.
XAI, please do more podcasts.
Anyway, but thank you for sharing.
It's been very kind.
And also like, I want to hear more from you.
I think you are going to embark on your next phase.
You haven't announced what you're doing next, but clearly you have you know more vision and more ambition on this path.
And I think you're basically kind of gradient dissenting to like whatever your final form is.
Thank you.
Yeah, I'll share more about my next chapter soon.
Okay.
Thank you for having me.
Thanks for coming.
