Now I can't stop thinking about _The Experience Machine_ by Andy Clark. It theorizes that this is how humans navigate and experience the real world: Our brains generate what we think the world around is like and our senses don't so much directly process visual information but instead act like a kind of loss function for our internal simulations. Then we use that error to update our internal model of the world.
In this view, we are essentially living inside a high-fidelity generative model. Our brains are constantly 'hallucinating' a predicted reality based on past experience and current goals. The data from our senses isn't the source of the image; it's the error signal used to calibrate that internal model. Much like Genie 3 uses latent actions and frames to predict the next state of a world, our brains use 'Active Inference' to minimize the gap between what we expect and what we experience.
It suggests that our sense of 'reality' isn't a direct recording of the world, but a highly optimized, interactive simulation that is continuously 'regularized' by the photons hitting our retinas.
Everyone here seems too caught up in the idea that Genie is the product, and that its purpose is to be a video game, movie, or VR environment.
That is not the goal.
The purpose of world models like Genie is to be the "imagination" of next-generation AI and robotics systems: a way for them to simulate the outcomes of potential actions in order to inform decisions.
Agreed; everyone complained that LLMs have no world model, so here we go. Next logical step is to backfill the weights with encoded video from the real world at some reasonable frame rate to ground the imagination and then branch the inference on possible interventions (actions) in the near future of the simulation, throw the results into a goal evaluator and then send the winning action-predictions to motors. Getting timing right will probably require a bit more work than literally gluing them together, but probably not much more.
Soft disagree; if you wanted imagination you don't need to make a video model. You probably don't need to decode the latents at all. That seems pretty far from information-theoretic optimality, the kind that you want in a good+fast AI model making decisions.
The whole reason for LLMs inferencing human-processable text, and "world models" inferencing human-interactive video, is precisely so that humans can connect in and debug the thing.
I think the purpose of Genie is to be a video game, but it's a video game for AI researchers developing AIs.
I do agree that the entertainment implications are kind of the research exhaust of the end goal.
> I think the purpose of Genie is to be a video game, but it's a video game for AI researchers developing AIs.
Yeah, I think this is what the person above was saying as well. This is what people at google have said already (a few podcasts on gdm's channel, hosted by Hannah Fry). They have their "agents" play in genie-powered environments. So one system "creates" the environment for the task. Say "place the ball in the basket". Genie creates an env with a ball and a basket, and the other agent learns to wasd its way around, pick up the ball and wasd to the basket, and so on. Pretty powerful combo if you have enough compute to throw at it.
Sufficiently informative latents can be decoded into video.
When you simulate a stream of those latents, you can decode them into video.
If you were trying to make an impressive demo for the public, you probably would decode them into video, even if the real applications don't require it.
Converting the latents to pixel space also makes them compatible with existing image/video models and multimodal LLMs, which (without specialized training) can't interpret the latents directly.
Didn’t the original world models paper do some training in latent space? (Edit: yes[1])
I think robots imagining the next step (in latent space) will be useful. It’s useful for people. A great way to validate that a robot is properly imagining the future is to make that latent space renderable in pixels.
[1] “By using features extracted
from the world model as inputs to an agent, we can train a very compact and simple policy that can solve the required task. We can even train our agent entirely inside of its own hallucinated dream generated by its world model, and transfer this policy back into the actual environment.”
> you don't need to make a video model. You probably don't need to decode the latents at all.
If you don't decode, how do you judge quality in a world where generative metrics are famously very hard and imprecise?
How do you go about integrating RLHF/RLAF in your pipeline if you don't decode, which is not something you can skip anymore to get SotA?
Just look at the companies that are explicitly aiming for robotics/simulation, they *are* doing video models.
> if you wanted imagination you don't need to make a video model. You probably don't need to decode the latents at all.
Soft disagree. What is the purpose of that imagination if not to map it to actual real world outfcomes. For this to compare them to the real world and possibly backpropagate through them you'll need video frames.
What model do you need then? If you want 3D real-time understanding of how realities work? Are you focusing on "imagination" in a different abstract way?
I am not sure we are at the "efficiency" phase of this.
Even if you just wire this output (or probably multiples running different counterfactuals) into a multimodal LLM that interprets the video and uses it to make decisions, you have something new.
Yeah and the goal of Instagram was to share quirky pictures you took with your friends. Now it’s a platform for influencers and brainrot; arguably it has done more damage than drugs to younger generations.
As soon as this thing is hooked up to VR and reaches a tipping point with the general public we all know exactly what is going to happen. The creation of the most profitable, addictive and ultimately dystopian technology Big Tech has ever come up with.
This is a paper that recently got popular ish and discusses the counter to your viewpoint.
> Paradox 1: Information cannot be increased by deterministic processes. For both Shannon entropy and Kolmogorov complexity, deterministic transformations cannot meaningfully increase the information content of an object. And yet, we use pseudorandom number generators to produce randomness, synthetic data improves model capabilities, mathematicians can derive new knowledge by reasoning from axioms without external information, dynamical systems produce emergent phenomena, and self-play loops like AlphaZero learn sophisticated strategies from games
In theory yes, something like the rules of chess should be enough for these mythical perfect reasoners that show up in math riddles to deduce everything that *can* be known about the game. And similarly a math textbook is no more interesting than a book with the words true and false and a bunch of true => true statements in it.
But I don't think this is the case in practice. There is something about rolling things out and leveraging the results you see that seems to have useful information in it even if the roll out is fully characterizable.
Interesting paper, thanks! But, the authors escape the three paradoxes they present by introducing training limits (compute, factorization, distribution). Kind of a different problem here.
What I object to are the "scaling maximalists" who believe that if enough training data were available, that complicated concepts like a world model will just spontaneously emerge during training. To then pile on synthetic data from a general-purpose generative model as a solution to the lack of training data becomes even more untenable.
Given that the video is fully interactive and lets you move around (in a “world” if you will) I don’t think it’s a stretch to call it a world model. It must have at least some notion of physics, cause and effect, etc etc in order to achieve what it does.
Whoa, whoa, whoa. That's just one angle. Please don't bin that as the only use case for "world models"!
First of all, there are a variety of different types of world models. Simulation, video, static asset, etc. It's a loaded term, just as the use cases are widespread.
There are world models you can play in your browser inferred entirely by your CPU:
The entertainment industry, as big as it is, just doesn't have as much profit potential as robots and AI agents that can replace human labor. Just look at how Nvidia has pivoted from gaming and rendering to AI.
The other examples you've given are neat, but for players like Google they are mostly an afterthought.
This tech is going to revolutionize "films" and gaming. The entire entertainment industry is going to transform around it.
When people aren't buying physical things, they're distracting themselves with media. Humans spend more time and money on that than anything else. Machines or otherwise.
AI impact on manufacturing will be huge. AI impact on media and entertainment will be huge. And these world models can be developed in a way that you develop exposure and competency for both domains.
edit: You can argue that manufacturing will boom when we have robotics that generalize. But you can also argue that entertainment will boom when we have holodecks people can step into.
The current robotics industry is $88B. You have to take into account the potential future industry of general purpose robots that replace a big chunk of blue-collar work.
Robots is also just one example. A hypothetically powerful AI agent (which might also use a world model) that controls a mouse and keyboard could replace a big chunk of white-collar work too.
Those are worth 10's of trillions of dollars. You can argue about whether they are actually possible, but the people backing this tech think they are.
You already can, check out Marble/World Labs, Meshy, and others.
It's not really as much of a boon as you'd think though, since throwing together a 3D model is not the bottleneck to making a sellable video game. You've had model marketplaces for a long time now.
> It's not really as much of a boon as you'd think though
It is for filmmaking! They're perfect for constructing consistent sets and blocking out how your actors and props are positioned. You can freely position the camera, control the depth of field, and then storyboard your entire scene I2V.
This I definitely agree with, before you had to massage the I2I and now you can just drag the camera.
Marble definitely changes the game if the game is "move the camera", just most people would not consider that a game (but hey there's probably a good game idea in there!)
Like LLMs, though: Do you really think a simulation will get them to all the corner cases robots/AI needs to know about, or will it be largely the same problem -- they'll be just good enough to fool the engineers and make the business ops drool and they'll be put into production and suddenly we'll see in a year or two stories about robots crushing peoples hands, stepping in drains and falling over or falling off roofs cause of some bizarre miscommunication between training and reality.
So, like, it's very important to understand the lineage of training and not just the "this is it"
The actual breakthrough with Genie is being able to turn around and look back, and seeing the same scene that was there before. A few other labs have similar world simulators, but they all struggle badly with keeping coherence of things not in view. Hence why they always walk forwards and never look around.
>Genie 3’s consistency is an emergent capability. Other methods such as NeRFs and Gaussian Splatting also allow consistent navigable 3D environments, but depend on the provision of an explicit 3D representation. By contrast, worlds generated by Genie 3 are far more dynamic and rich because they’re created frame by frame based on the world description and actions by the user.
Dunno, I want to agree, but at the same time it's spoken like someone to whom these experiences and human relationship come easily. There are many people out there who, for some reason (anxiety, etc.), cannot easily access this part of the human condition, unfortunately.
Perhaps better to roam a virtual reality than be starved in the real world.
The American public is no different than an American corporation; trying to extract as much allegiance and loyalty as possible for as little compensation as possible.
Your neighbors in the street protesting for comprehensive single payer healthcare? Yeah they're perfectly fine leaving your existence up to "market forces".
Copy-paste office workers everywhere reciting memorized platitudes and compliance demands.
You're telling me I could interact even less with such selfish (and often useless given their limited real skillset) people? Deal.
America needs to rethink the compensation package if it wants to survive as a socio-political meme. Happy to call myself Canadian or Chinese if their offer is better. No bullets needed.
>I think we'll eventually get to the point where these are real time and have consistent representations
You have a dangerously low opinion of your fellow man, and while I sympathize with your frustration, I would humbly suggest you direct that anger at owners of companies/politicians, rather than aim it at your everyday citizen.
Your suggestion is meaningless semantic differentiation.
Those owners and politicians are the result of exposure to American communities, schools, other institutions; they do not spontaneously exist.
Americans prop up the system as such Americans will defer or their faith was misplaced to begin with. And that ain't right; they're America! So the awfulness will continue until moral improves!
Atheist semantics while living theist like devotion to civil religion memes.
It is only detaching people from reality more. The internet used to be a window into the outside world and now AI is making it counterfeit manipulated fantasy.
Ironically, this brings us one step closer to believing the simulation hypothesis might be true... In which case, maybe there is no real world anyway ;)
I think this would be good for the people living in overcrowded, polluted and dirty cities in the world (let's be honest, they actually exist regardless of how that happened or why).
Maybe they can unplug from 500+ AQI pollution and spend time with their loved ones and friends in a simulated clean world?
Imagine working for 10-12 hours a day, and you come home to a pod (and a building could house thousands of pods, paid for by the company) where you plug in and relax for a few hours. Maybe a few more decades of breakthroughs can lead to simulated sleep as well so they get a full rest.
Wake up, head to the factory to make whatever the developed world needs.
(holy fuck that is a horrible existence but you know some people would LOVE for that to be real)
Hook it up to an always-on fle*light, and I'm sure you'd have millions of paying customers.
Except you'll never have to leave your pod. Extract the $$ from their attention all day, then sell them manufactured virtual happiness all night. It's just a more streamlined version of how many people live right now.
I mean if I'm totally honest, it could be beneficial to me if something like this comes to be. Even in the developed world, we have a bunch of annoying people who complain/cry constantly about dumb things. They do that instead of doing something, whereas I can excuse the people trapped in hellholes overseas because it really isn't their fault they were oppressed and mistreated.
They'd have their own economy and "life" and leave the rest of us alone. It would be completely transactional, so I'd have zero reason to feel bad if they do it voluntarily.
If they can be happy in a simulated world, and others can be happy in the real world, then everyone wins!
I think we'll eventually get to the point where these are real time and have consistent representations. I've been excited about world models since I saw the in-the-browser Pokemon demo:
At some point, we'll have the creative Holodeck. If you've seen what single improv performers can do with AI, it's ridiculously cool. I can imagine watching entertainers in the future that summon and create entire worlds before us:
I have been confused for a long time why FB is not motivated enough to invest in world models, it IS the key to unblock their "metaverse" vision. And instead they let go Yann LeCun.
LeCun wasn't producing results. He was obstinate and insistent on his own theories and ideas which weren't, and possibly aren't, going anywhere. He refused to engage with LLMs and compete in the market that exists, and spent all his effort and energy on unproven ideas and research, which split the company's mission and competitiveness. They lost their place as one of the top 4 AI companies, and are now a full generation behind, in part due to the split efforts and lack of enthusiastic participation by all the Meta AI team. If you look at the chaos and churn at the highest levels across the industry, there's not a lot of room for mission creep by leadership, and LeCun thoroughly demonstrated he wasn't suited for the mission desired by Meta.
I think he's lucky he got out with his reputation relatively intact.
To be fair, this was his job description: Fundamental AI Research (FAIR) lab. Not AI products division. You can't expect marketable products from a fundamental AI research lab.
Attentive outsider and acquaintance of a couple people who are or were employed there. Nothing I'm saying is particularly inside baseball, though, it's pretty well covered by all the blogs and podcasts.
Machine Learning Street Talk and Dwarkesh are excellent. Various discord communities, forums, and blogs downstream of the big podcasts, and following researchers on X keeps you in the loop on a lot of these things, and then you can watch for random interviews and presentations on youtube when you know who the interesting people and subjects are.
Most serious researchers want to work on interesting problems like reinforcement learning or robotics or RNN or dozen other avant-garde subjects. None want to work on "boring" LLM technology, requiring significant engineering effort and huge dataset wrangling effort.
This is true - Ilya got an exit and is engaged in serious research, but research is by its nature unpredictable. Meta wanted a product and to compete in the AI market, and JEPA was incompatible with that. Now LeCun has a lab and resources to pursue his research, and Meta has refocused efforts on LLMs and the marketplace - it remains to be seen if they'll be able to regain their position. I hope they do - open models and relatively open research are important, and the more serious AI labs that do this, the more it incentivizes others to do the same, and keeps the ones that have committed to it honest.
In an industry of big bets, especially considering the company has poured resources and renamed itself to secure a place in the VR world... staking your reputation on everyone's LLMs having peaked and shifting focus to finding a new path to AI is a pretty interesting bet, no?
Since a hot take is as good as the next one: LLMs are by the day more and more clearly understood as a "local maximum" with flawed capabilities, limited efficiency, a $trillion + a large chunk of the USA's GDP wasted, nobody even turning a profit from that nor able to build something that can't be reproduced for free within 6 months.
When the right move (strategically, economically) is to not compete, the head of the AI division acknowledging the above and deciding to focus on the next breakthrough seems absolutely reasonable.
This sounds similar to the arc of Carpathy, who also managed to preserve his reputation despite sending Tesla down a FSD deadend and missing the initial LLM boat.
Isn't it more like this:
JEPA looks at the video, "a dog walks out of the door, the mailman comes, dog is happy" and the next frame would need to look like "mailman must move to mailbox, dog will run happily towards him", which then an image/video generator would need to render.
Genie looks at the video, "when this group of pixels looks like this and the user presses 'jump', I will render the group different in this way in the next frame."
Genie is an artist drawing a flipbook. To tell you what happens next, it must draw the page. If it doesn't draw it, the story doesn't exist.
JEPA is a novelist writing a summary. To tell you what happens next, it just writes "The car crashes." It doesn't need to describe what the twisted metal looks like to know the crash happened.
You are beyond correct. World models is what saves their Reality Labs investment. I would say if Reality Labs cannot productize World Models, then that entire project needs to be scrapped.
Reminds me of this [1] HN post from 9 months ago, where the author trained a neural network to do world emulation from video recordings of their local park — you can walk around in their interactive demo [2].
I don't have access to the DeepMind demo, but from the video it looks like it takes the idea up a notch.
(I don't know the exact lineage of these ideas, but a general observation is that it's a shame that it's the norm for blog posts / indie demos to not get cited.)
Yup, similar concepts! Just at two opposite extremes of the compute/scaling spectrum.
- That forest trail world is ~5 million parameters, trained on 15 minutes of video, scoped to run on a five-year-old iPhone through a twenty-year old API (WebGL GPGPU, i.e OpenGL fragment shaders). It's the smallest '3D' world model I'm aware of.
- Genie 3 is (most likely) ~100 billion parameters trained on millions of hours of video and running across multiple TPUs. I would be shocked if it's not the largest-scale world model available to the public.
I keep on repeating myself, but it feels like I'm living in the future.
Can't wait to hook this up to my old Oculus glasses and let Genie create a fully realistic sailing simulator for me, where I can train sailing with realistic conditions. On boats I'd love to sail.
If making games out of these simulations work, it't be the end for a lot of big studios, and might be the renaissance for small to one person game studios.
Isn't this still essentially "vibe simulation" inferred from videos? Surface-level visual realism is one thing, but expecting it to figure out the exact physical mechanics of sailing just by watching boats, and usefully abstract that into a gamified form, is another thing entirely.
Yeah I have a whole lot of trouble imagining this replacing traditional video games any time soon; we have actually very good and performant representations of how physics work, and games are tuned for the player to have an enjoyable experience.
There's obviously something insanely impressive about these google experiments, and it certainly feels like there's some kind of use case for them somewhere, but I'm not sure exactly where they fit in.
Google has made it clear that Genie doesn't maintain an explicit 3D scene representation, so I don't think hooking in "assists" like that is on the table. Even if it were, the AI layer would still have to infer things like object weight, density, friction and linkages correctly. Garbage in, garbage out.
The bottleneck for games of any size is always whether they are good. There are plenty of small indies which do not put out good games. I don't see world models improving game design or fun factors.
If I am wrong, then the huge supply of fun games will completely saturate demand and be no easier for indie game devs to stand out.
It's very impressive tech but subject to the same limitations as other generative AI: Inconsistency, inaccurate physics, limited time, lag, massively expensive computation.
You COULD create a sailing sim but after ten minutes you might be walking on water, or in the bath, and it would use more power than a small ferry.
There's no way this tech can run on a PS5 or anything close to it.
Five years is nothing to wait for tech like this. I'm sure we will see the first crop of, however small, "terminally plugged in" humans on the back of this in the relatively near future.
> If making games out of these simulations work, it't be the end for a lot of big studios, and might be the renaissance for small to one person game studios.
I mean, if making a game eventually boils down to cooking a sufficient prompt (which to be clear, I'm not talking about text, these prompts are probably going to be more like video databases) then I'm not sure if it will be a renaissance for "one person game studios" any more than AI image generation has been a renaissance for "one person artists".
I want to be optimistic but it's hard to deny the massive distribution stranglehold that media publishing landscape has, and that has nothing to do with technology.
I have no idea why Google is wasting their time with this. Trying to hallucinate an entire world is a dead-end. There will never be enough predictability in the output for it to be cohesive in any meaningful way, by design. Why are they not training models to help write games instead? You wouldn't have to worry about permanence and consistency at all, since they would be enforced by the code, like all games today.
Look at how much prompting it takes to vibe code a prototype. And they want us to think we'll be able to prompt a whole world?
This was a common argument against LLMs, that the space of possible next tokens is so vast that eventually a long enough sequence will necessarily decay into nonsense, or at least that compounding error will have the same effect.
Problem is, that's not what we've observed to happen as these models get better. In reality there is some metaphysical coarse-grained substrate of physics/semantics/whatever[1] which these models can apparently construct for themselves in pursuit of ~whatever~ goal they're after.
The initially stated position, and your position: "trying to hallucinate an entire world is a dead-end", is a sort of maximally-pessimistic 'the universe is maximally-irreducible' claim.
And going back a little further, it was thought that backpropagation would be impractical, and trying to train neural networks was a dead end. Then people tried it and it worked just fine.
> Problem is, that's not what we've observed to happen as these models get better
Eh? Context rot is extremely well known. The longer you let the context grow, the worse LLMs perform. Many coding agents will pre-emptively compact the context or force you to start a new session altogether because of this. For Genie to create a consistent world, it needs to maintain context of everything, forever. No matter how good it gets, there will always be a limit. This is not a problem if you use a game engine and code it up instead.
Imo they explain pretty well what they are trying to achieve with SIMA and Genie in the Google Deepmind Podcast[1].
They see it as the way to get to AGI by letting AI agents learn for themselves in simulated worlds. Kind of like how they let AlphaGo train for Go in an enormous amount of simulated games.
That makes even less sense, because an AI agent cannot learn effectively from a hallucinated world without internal consistency guarantees. It's an even stronger case for leveraging standard game engines instead.
If that's the goal, the technology for how these agents "learn" would be the most interesting one, even more than the demos in the link.
LLMs can barely remember the coding style I keep asking it to stick to despite numerous prompts, stuffing that guideline into my (whatever is the newest flavour of product-specific markdown file). They keep expanding the context window to work around that problem.
If they have something for long-term learning and growth that can help AI agents, they should be leveraging it for competitive advantage.
Take the positive spin. What if you could put in all the inputs and it can simulate real world scenarios you can walk through to benefit mankind e.g disaster scenarios, events, plane crashes, traffic patterns. I mean there's a lot of useful applications for it. I don't like the framing at this time, but I also get where it's going. The engineer in me is drawn to it, but the Muslim in me is very scared to hear anyone talk about creating worlds.... But again I have to separate my view from the reality that this could have very positive real world benefits when you can simulate scenarios. So I could put in a 2 pager or 10 page scenario that gets played out or simulated and allow me to walk through it. Not just predictive stuff but let's say things that have happened so I can map crime scenes or anything. In the end this performance art is because they are a product company being Benchmarked by wall street and they'll need customers for the technology but at the same time they probably already have uses for it internally.
> What if you could put in all the inputs and it can simulate real world scenarios you can walk through to benefit mankind e.g disaster scenarios, events, plane crashes, traffic patterns.
This is only a useful premise if it can do any of those things accurately, as opposed to dreaming up something kinda plausible based on an amalgamation of every vaguely related YouTube video.
> What if you could put in all the inputs and it can simulate real world scenarios you can walk through to benefit mankind e.g disaster scenarios, events, plane crashes, traffic patterns.
What's the use? Current scientific models clearly showing natural disasters and how to prevent them are being ignored. Hell, ignoring scientific consensus is a fantastic political platform.
An hybrid approach could maybe work, have a more or less standard game engine for coherence and use this kind of generative AI more or less as a short term rendering and physics sim engine.
I've thought about this same idea but it probably gets very complicated.
Let's say, you simulate a long museum hallway with some vases in it. Who holds what? The basic game engine has the geometry, but once the player pushes it and moves it, it needs to inform the engine it did, and then to draw the next frame, read from the engine first, update the position in the video feed, then again feed it back to the engine.
What happens if the state diverges. Who wins? If the AI wins then...why have the engine at all?
It is possible but then who controls physics. The engine? or the AI? The AI could have a different understanding of the details of the base. What happens if the vase has water inside? who simulates that? what happens if the AI decides to break the vase? who simulates the AI.
I don't doubt that some sort of scratchpad to keep track of stuff in game would be useful, but I suspect the researchers are expecting the AI to keep track of everything in its own "head" cause that's the most flexible solution.
Then maybe the engine should be less about really simulating the 3D world and just trying best to preserve consistency, more about providing memory and saving context for consistency than truly simulating a lot besides higher level concerns (at which point we might wonder if it couldn't be directly part of the model somehow), but writing those lines I realize there would probably still be many edge cases exactly like what you are describing...
As a kid in the early 1980s, I spent a lot of time experimenting with computers by playing basic games and drawing with crude applications. And it was fun. I would have loved to have something like Google's Genie to play with. Even if it never evolved, the product in the demos looks good enough for people to get value from.
> Why are they not training models to help write games instead?
Genie isn't about making games... Granted, they for some reason they don't put this at the top. Classic Google, not communicating well...
| It simulates physics and interactions for dynamic worlds, while its breakthrough consistency enables the simulation of any real-world scenario — from robotics and modelling animation and fiction, to exploring locations and historical settings.
The key part is simulation. That's what they are building this for. Ignore everything else.
Same with Nvidia's Earth 2 and Cosmos (and a bit like Isaac). Games or VR environments are not the primary drive, the primary drive is training robots (including non-humanoids, such as Waymo) and just getting the data. It's exactly because of this that perfect physics (or let's be honest, realistic physics[0,1]). Getting 50% of the way there in simulation really does cut down the costs of development, even if we recognize that cost steepens as we approach "there". I really wish they didn't call them "world models" or more specifically didn't shove the word "physics" in there, but hey, is it really marketing if they don't claim a golden goose can not only lay actual gold eggs but also diamonds and that its honks cure cancer?
[0] Looking right does not mean it is right. Maybe it'll match your intuition or undergrad general physics classes with calculus but talk to a real physicist if you doubt me here. Even one with just an undergrad will tell you this physics is unrealistic and any one worth their salt will tell you how unintuitive physics ends up being as you get realistic, even well before approaching quantum. Go talk to the HPC folks and ask them why they need superocmputers... Sorry, physics can't be done from observation alone.
[1] Seriously, I mean look at their demo page. It really is impressive, don't get me wrong, but I can't find a single video that doesn't have major physics problems. That "A high-altitude open world featuring deformable snow terrain." looks like it is simulating Legolas[2], not a real person. The work is impressive, but it isn't anywhere near realistic https://deepmind.google/models/genie/
But it's not simulating, is it? It's hallucinating videos with an input channel to guide what the video looks like. Why do that instead of just picking Unreal, Unity, etc and having it actually simulated for a fraction of the effort?
Why is it a dead end, you don’t meaningfully explain that. These models look like you can interact with them and they seem to replicate physics models.
They don't though, they're hallucinated videos. They're feeding models tons and tons of 2D videos and hoping they figure out physics from them, instead of just using a game engine and having the LLM write something up that works 100% of the time.
This is what we were building in 2018 with Ayvri, starting from 3d tiles with the aim of building a real-world view by using AI to essentailly re-paint and add detail to what was essentially a high-resolution and faster loading Google Earth (for outside cities, we didn't have building data).
We saw a very diverse group of users, the common uses was paragliders, gliders, and pilots who wanted to view their or other peoples flights. Ultramarathons, mountain bike and some road-races where it provided an interactive way to visualize the course from any angle and distance. Transportation infrastructure to display train routes to be built. The list goes on.
It's super cool but I see it as a much more flexible open ended take on the idea of procedurally generated worlds where hard-coded deterministic math and rendering parameters are replaced by prompt-able models.
The deadness you're talking about is there in procedural worlds too, and it stems from the fact that there's not actually much "there." Think of it as a kind of illusion or a magic trick with math. It replicates some of the macro structure of the world but the true information content is low.
Search YouTube for procedural landscape examples. Some of them are actually a lot more visually impressive than this, but without the interactivity. It's a popular topic in the demo scene too where people have made tiny demos (e.g. under 1k in size) that generate impressive scenes.
I expect to see generative AI techniques like this show up in games, though it might take a bit due to their high computational cost compared to traditional procedural generation.
Best case, Google DeepMind cracks AGI by letting agents learn for themselves inside simulated worlds. Worst case, they've invented the greatest, most expensive screensaver generator in human history.
This could be the future of film. Instead of prompting where you don't know what the model will produce, you could use fine-grained motion controls to get the shot you are looking for. If you want to adjust the shot after, you could just checkpoint the model there, by taking a screenshot, and rerun. Crazy.
Compared to DeepMind's Genie 3 demo, this appears to have more morphing issues and less user interactivity with environmental consistency. Is this a stripped down version?
I've been experimenting with that from a slightly different angle: teaching Claude how to play and referee a pencil-and-paper RPG that I developed over about 20 years starting in the mid 1970s. Claude can't quite do it yet for reasons related to permanence and learning over time, but it can do surprisingly well up until it runs into those problems, and it's possible to help it past some obstacles.
The game is called "Explorers' Guild", or "xg" for short. It's easier for Claude to act as a player than a director (xg's version of a dungeon master or game master), again mainly because of permance and learning issues, but to the extent that I can help it past those issues it's also fairly good at acting as a director. It does require some pretty specific stuff in the system prompt to, for example, avoid confabulating stuff that doesn't fit the world or the scenario.
But to really build a version of xg on Claude it needs better ways to remember and improve what it has learned about playing the game, and what it has learned about a specific group of players in a specific scenario as it develops over time.
So what is it doing in the real world, microwaving an elephant on high with 80kw every second and pouring out all the water in an sub-saharan African well every 4 minutes?
Ironically the physics are kind of my biggest criticism. They call these "world models", but I think it's more accurate to call them "video game models" because they employ "video game physics" rather than real world physics, among other things
It's getting better staggeringly fast, just a year ago I wouldn't expect it to be at even video game physics level so quickly.
If there is a possibility where it continue to improve at a similar rate with llms. A way to simulate fluid dynamics or structural dynamics with reasonable accuracy and speed can unlock much faster pace of innovation in the physical world. (And validated with rigorous scientific methods)
I am stumped. Am I misreading, or are the folks at Google deliberately confounding two interpretations of "world model"? Dont get me wrong, this is really cool, and it will undoubtedly have its use. But what I am seeing is an LLM that can generate textures to be fed into a human-coded 3d engine (the "world model" that is demonstrated), and I fail to see how that brings us closer to AGI. For AGI we need "world models" as in "belief systems". The AI model must be able to reason about (learned) dynamics, which I dont see reflected in the text or video.
>an LLM that can generate textures to be fed into a human-coded 3d engine
I'm not certain but I think the LLM is also generating the physics itself. It's generating rules based on its training data, e.g. watch a cat walk enough and you can simulate how the cat moves in the generated "world".
The goal of world models like Genie is to be a way for AI and robots to "imagine" things. Then, they could practice tasks inside of the simulated world or reason about actions by simulating their outcome.
It means you should go the other way. Open world winning against smaller, handcrafted environments and stories was generally a mistake, and so is this.
What does it mean, that open world winning was a mistake? That the market is wrong, and peoples' preferences were incorrect, and they should prefer small handcrafted environments instead of what they seem to actually buy?
I would think that building a environment which can be managed by a game engine is the first pass. In a few years when we are able to render more than 60 seconds it could very well replace the game engine entirely by just rendering everything in realtime based on user interactions. The final phase is just prompts which turn directly into interactive games, maybe even multiplayer. When I see the progress we've made on things like DOOM, where it can infer the proper rendering of actions like firing weapons and even updating scores on hits and such it doesn't feel like we're very far off, a few years at most. For a game studio that could mean cutting out almost everything between keyboard and display, but for now just replacing the asset pipeline is huge.
We seem to think that Genie is good at the creative part, but bad at the consistency and performance part. How hard would it be to take 60 seconds of Genie output and pipe it into a model that generates a consistent and performant 3D environment?
I understand the ultimate end goal to be simulation of life. A near perfect replica of the real world we can use to simulate and test medicine, economy, and social impact.
I don't know if that's the simplest explanation, considering how insanely complex the generation is (these world models might literally be the most complex things to ever be created).
I mean, yes, the probability of having that level of tech in decades is quite high.
But the technology is moving very fast right now. It sounds crazy, but I think that there is a 50% chance of having ready player one level technology before 2030.
It's absolutely possible it will take more time to become economical.
Now I can't stop thinking about _The Experience Machine_ by Andy Clark. It theorizes that this is how humans navigate and experience the real world: Our brains generate what we think the world around is like and our senses don't so much directly process visual information but instead act like a kind of loss function for our internal simulations. Then we use that error to update our internal model of the world.
In this view, we are essentially living inside a high-fidelity generative model. Our brains are constantly 'hallucinating' a predicted reality based on past experience and current goals. The data from our senses isn't the source of the image; it's the error signal used to calibrate that internal model. Much like Genie 3 uses latent actions and frames to predict the next state of a world, our brains use 'Active Inference' to minimize the gap between what we expect and what we experience.
It suggests that our sense of 'reality' isn't a direct recording of the world, but a highly optimized, interactive simulation that is continuously 'regularized' by the photons hitting our retinas.
A kurzgesagt on this: Why Your Brain Blinds You For 2 Hours Every Day https://youtu.be/wo_e0EvEZn8 and the sources for that video - https://sites.google.com/view/sources-reality-is-not-real/
Everyone here seems too caught up in the idea that Genie is the product, and that its purpose is to be a video game, movie, or VR environment.
That is not the goal.
The purpose of world models like Genie is to be the "imagination" of next-generation AI and robotics systems: a way for them to simulate the outcomes of potential actions in order to inform decisions.
Agreed; everyone complained that LLMs have no world model, so here we go. Next logical step is to backfill the weights with encoded video from the real world at some reasonable frame rate to ground the imagination and then branch the inference on possible interventions (actions) in the near future of the simulation, throw the results into a goal evaluator and then send the winning action-predictions to motors. Getting timing right will probably require a bit more work than literally gluing them together, but probably not much more.
Soft disagree; if you wanted imagination you don't need to make a video model. You probably don't need to decode the latents at all. That seems pretty far from information-theoretic optimality, the kind that you want in a good+fast AI model making decisions.
The whole reason for LLMs inferencing human-processable text, and "world models" inferencing human-interactive video, is precisely so that humans can connect in and debug the thing.
I think the purpose of Genie is to be a video game, but it's a video game for AI researchers developing AIs.
I do agree that the entertainment implications are kind of the research exhaust of the end goal.
> I think the purpose of Genie is to be a video game, but it's a video game for AI researchers developing AIs.
Yeah, I think this is what the person above was saying as well. This is what people at google have said already (a few podcasts on gdm's channel, hosted by Hannah Fry). They have their "agents" play in genie-powered environments. So one system "creates" the environment for the task. Say "place the ball in the basket". Genie creates an env with a ball and a basket, and the other agent learns to wasd its way around, pick up the ball and wasd to the basket, and so on. Pretty powerful combo if you have enough compute to throw at it.
Sufficiently informative latents can be decoded into video.
When you simulate a stream of those latents, you can decode them into video.
If you were trying to make an impressive demo for the public, you probably would decode them into video, even if the real applications don't require it.
Converting the latents to pixel space also makes them compatible with existing image/video models and multimodal LLMs, which (without specialized training) can't interpret the latents directly.
Didn’t the original world models paper do some training in latent space? (Edit: yes[1])
I think robots imagining the next step (in latent space) will be useful. It’s useful for people. A great way to validate that a robot is properly imagining the future is to make that latent space renderable in pixels.
[1] “By using features extracted from the world model as inputs to an agent, we can train a very compact and simple policy that can solve the required task. We can even train our agent entirely inside of its own hallucinated dream generated by its world model, and transfer this policy back into the actual environment.”
https://arxiv.org/abs/1803.10122
> you don't need to make a video model. You probably don't need to decode the latents at all.
If you don't decode, how do you judge quality in a world where generative metrics are famously very hard and imprecise? How do you go about integrating RLHF/RLAF in your pipeline if you don't decode, which is not something you can skip anymore to get SotA?
Just look at the companies that are explicitly aiming for robotics/simulation, they *are* doing video models.
> if you wanted imagination you don't need to make a video model. You probably don't need to decode the latents at all.
Soft disagree. What is the purpose of that imagination if not to map it to actual real world outfcomes. For this to compare them to the real world and possibly backpropagate through them you'll need video frames.
If you train a video model, you by necessity train a world model for 3D worlds. Which can then be reused in robotics, potentially.
I do wonder if I can frankenstein together a passable VLA using pretrained LTX-2 as a base.
Sure, but at some point you want humans in the loop i guess?
What model do you need then? If you want 3D real-time understanding of how realities work? Are you focusing on "imagination" in a different abstract way?
Sure, but at some point you want humans in the loop i guess?
I am not sure we are at the "efficiency" phase of this.
Even if you just wire this output (or probably multiples running different counterfactuals) into a multimodal LLM that interprets the video and uses it to make decisions, you have something new.
Yeah and the goal of Instagram was to share quirky pictures you took with your friends. Now it’s a platform for influencers and brainrot; arguably it has done more damage than drugs to younger generations.
As soon as this thing is hooked up to VR and reaches a tipping point with the general public we all know exactly what is going to happen. The creation of the most profitable, addictive and ultimately dystopian technology Big Tech has ever come up with.
The good news is we’ll finally have an answer for the Fermi Paradox.
Environment mapping to AI generated alternative outcomes is the holodeck.
I prefer real danger as living in the simulation is derivative.
I think this is the key component of developing subjective experience.
Still cool though…
this.
This is a video model, not a world model. Start learning on this, and cascading errors will inevitably creep into all downstream products.
You cannot invent data.
Related: https://arxiv.org/abs/2601.03220
This is a paper that recently got popular ish and discusses the counter to your viewpoint.
> Paradox 1: Information cannot be increased by deterministic processes. For both Shannon entropy and Kolmogorov complexity, deterministic transformations cannot meaningfully increase the information content of an object. And yet, we use pseudorandom number generators to produce randomness, synthetic data improves model capabilities, mathematicians can derive new knowledge by reasoning from axioms without external information, dynamical systems produce emergent phenomena, and self-play loops like AlphaZero learn sophisticated strategies from games
In theory yes, something like the rules of chess should be enough for these mythical perfect reasoners that show up in math riddles to deduce everything that *can* be known about the game. And similarly a math textbook is no more interesting than a book with the words true and false and a bunch of true => true statements in it.
But I don't think this is the case in practice. There is something about rolling things out and leveraging the results you see that seems to have useful information in it even if the roll out is fully characterizable.
Interesting paper, thanks! But, the authors escape the three paradoxes they present by introducing training limits (compute, factorization, distribution). Kind of a different problem here.
What I object to are the "scaling maximalists" who believe that if enough training data were available, that complicated concepts like a world model will just spontaneously emerge during training. To then pile on synthetic data from a general-purpose generative model as a solution to the lack of training data becomes even more untenable.
They have a feature where you can take a photo and create a world from that.
If instead of a photo you have a video feed, this is one step closer to implementing subjective experience.
Given that the video is fully interactive and lets you move around (in a “world” if you will) I don’t think it’s a stretch to call it a world model. It must have at least some notion of physics, cause and effect, etc etc in order to achieve what it does.
No, it actually needs none of that.
Whoa, whoa, whoa. That's just one angle. Please don't bin that as the only use case for "world models"!
First of all, there are a variety of different types of world models. Simulation, video, static asset, etc. It's a loaded term, just as the use cases are widespread.
There are world models you can play in your browser inferred entirely by your CPU:
https://madebyoll.in/posts/game_emulation_via_dnn/ (my favorite, from 2022!)
https://madebyoll.in/posts/world_emulation_via_dnn/ (updated, in 3D)
There are static asset generating world models, like WorldLabs' Marble. These are useful for video games, previz, and filmmaking.
https://marble.worldlabs.ai/
I wrote open source software to leverage marble for filmmaking (I'm a filmmaker, and this tech is extremely useful for scene consistency):
https://www.youtube.com/watch?v=wJCJYdGdpHg
https://github.com/storytold/artcraft
There are playable video-oriented models, many of which are open source and will run on your 3080 and above:
https://diamond-wm.github.io/
https://github.com/Robbyant/lingbot-world
There are things termed "world models" that really shouldn't be:
https://github.com/Tencent-Hunyuan/HunyuanWorld-1.0
There are robotics training oriented world models:
https://github.com/leggedrobotics/robotic_world_model
Genie is not strictly robotics-oriented.
The entertainment industry, as big as it is, just doesn't have as much profit potential as robots and AI agents that can replace human labor. Just look at how Nvidia has pivoted from gaming and rendering to AI.
The other examples you've given are neat, but for players like Google they are mostly an afterthought.
Robotics: $88B TAM
Gaming: $350B TAM
All media and entertainment: $3T TAM
Manufacturing: $5T TAM
Roughly the same story.
This tech is going to revolutionize "films" and gaming. The entire entertainment industry is going to transform around it.
When people aren't buying physical things, they're distracting themselves with media. Humans spend more time and money on that than anything else. Machines or otherwise.
AI impact on manufacturing will be huge. AI impact on media and entertainment will be huge. And these world models can be developed in a way that you develop exposure and competency for both domains.
edit: You can argue that manufacturing will boom when we have robotics that generalize. But you can also argue that entertainment will boom when we have holodecks people can step into.
The current robotics industry is $88B. You have to take into account the potential future industry of general purpose robots that replace a big chunk of blue-collar work.
Robots is also just one example. A hypothetically powerful AI agent (which might also use a world model) that controls a mouse and keyboard could replace a big chunk of white-collar work too.
Those are worth 10's of trillions of dollars. You can argue about whether they are actually possible, but the people backing this tech think they are.
That’s part of it but if you could actually pull out 3D models from these worlds, it would massively speed up game development.
You already can, check out Marble/World Labs, Meshy, and others.
It's not really as much of a boon as you'd think though, since throwing together a 3D model is not the bottleneck to making a sellable video game. You've had model marketplaces for a long time now.
> It's not really as much of a boon as you'd think though
It is for filmmaking! They're perfect for constructing consistent sets and blocking out how your actors and props are positioned. You can freely position the camera, control the depth of field, and then storyboard your entire scene I2V.
Example of doing this with Marble: https://www.youtube.com/watch?v=wJCJYdGdpHg
This I definitely agree with, before you had to massage the I2I and now you can just drag the camera.
Marble definitely changes the game if the game is "move the camera", just most people would not consider that a game (but hey there's probably a good game idea in there!)
Like LLMs, though: Do you really think a simulation will get them to all the corner cases robots/AI needs to know about, or will it be largely the same problem -- they'll be just good enough to fool the engineers and make the business ops drool and they'll be put into production and suddenly we'll see in a year or two stories about robots crushing peoples hands, stepping in drains and falling over or falling off roofs cause of some bizarre miscommunication between training and reality.
So, like, it's very important to understand the lineage of training and not just the "this is it"
Really great to see this released! Some interesting videos from early-access users:
- https://youtu.be/15KtGNgpVnE?si=rgQ0PSRniRGcvN31&t=197 walking through various cities
- https://x.com/fofrAI/status/2016936855607136506 helicopter / flight sim
- https://x.com/venturetwins/status/2016919922727850333 space station, https://x.com/venturetwins/status/2016920340602278368 Dunkin' Donuts
- https://youtu.be/lALGud1Ynhc?si=10ERYyMFHiwL8rQ7&t=207 simulating a laptop computer, moving the mouse
- https://x.com/emollick/status/2016919989865840906 otter airline pilot with a duck on its head walking through a Rothko inspired airport
I liked that first one and I hope someone creates one of going back to dinosaur age, i want to see that.
One step closer to the science-based dinosaur MMO we were promised.
Tim is awesome.
Ironically, he covered PixVerse's world model last week and it came close to your ask: https://youtu.be/SAjKSRRJstQ?si=dqybCnaPvMmhpOnV&t=371
(Earlier in the video it shows him live prompting.)
World models are popping up everywhere, from almost every frontier lab.
Any thoughts about Project Genie?
The actual breakthrough with Genie is being able to turn around and look back, and seeing the same scene that was there before. A few other labs have similar world simulators, but they all struggle badly with keeping coherence of things not in view. Hence why they always walk forwards and never look around.
What about Fei Fei Li's lab? I think they are generating true 3D worlds rather than frames of a video?
Although that probably precludes her from having animations in those worlds...
Still amazed it took ML people so long to realize they needed and explicit representation to cache stuff.
Genie does not use an explicit representation:
>Genie 3’s consistency is an emergent capability. Other methods such as NeRFs and Gaussian Splatting also allow consistent navigable 3D environments, but depend on the provision of an explicit 3D representation. By contrast, worlds generated by Genie 3 are far more dynamic and rich because they’re created frame by frame based on the world description and actions by the user.
The representation is learned. Also, see Sutter's "Bitter Lesson" essay
And what if I go somewhere then go back there a week later?
Best they can do is 60 seconds, for now at least.
Makes you wonder what the TTL caching for our universe is.
Whatever the speed of light is I would imagine
The more of this I see the more I want to spend time away from screens and doing those things I love to do in the real world.
I love AI but I also hope it will paradoxically make people realize the value of real life experiences and human relationships.
Dunno, I want to agree, but at the same time it's spoken like someone to whom these experiences and human relationship come easily. There are many people out there who, for some reason (anxiety, etc.), cannot easily access this part of the human condition, unfortunately.
Perhaps better to roam a virtual reality than be starved in the real world.
Or maybe it will just make people realize the value of fake life experiences and human relationships.
The American public is no different than an American corporation; trying to extract as much allegiance and loyalty as possible for as little compensation as possible.
Your neighbors in the street protesting for comprehensive single payer healthcare? Yeah they're perfectly fine leaving your existence up to "market forces".
Copy-paste office workers everywhere reciting memorized platitudes and compliance demands.
You're telling me I could interact even less with such selfish (and often useless given their limited real skillset) people? Deal.
America needs to rethink the compensation package if it wants to survive as a socio-political meme. Happy to call myself Canadian or Chinese if their offer is better. No bullets needed.
>I think we'll eventually get to the point where these are real time and have consistent representations
You have a dangerously low opinion of your fellow man, and while I sympathize with your frustration, I would humbly suggest you direct that anger at owners of companies/politicians, rather than aim it at your everyday citizen.
Your suggestion is meaningless semantic differentiation.
Those owners and politicians are the result of exposure to American communities, schools, other institutions; they do not spontaneously exist.
Americans prop up the system as such Americans will defer or their faith was misplaced to begin with. And that ain't right; they're America! So the awfulness will continue until moral improves!
Atheist semantics while living theist like devotion to civil religion memes.
Maybe some folks (ahem) disappearing into virtual worlds is a good thing for those left behind.
It is only detaching people from reality more. The internet used to be a window into the outside world and now AI is making it counterfeit manipulated fantasy.
After a lifetime career in tech, I want to turn it all off.
Seriously though. These cute distractions have turned into world and culture eating monsters.
Ironically, this brings us one step closer to believing the simulation hypothesis might be true... In which case, maybe there is no real world anyway ;)
Huh, yeah, the sky is blue outside and the sun is shining.
Although, I am feeling a bit lazy so let me see if I can simulate a walk.
Most people don't have access to anything particularly nice in the real world.
It's reality privilege. Most of humanity will yearn for the worlds that AI will cook up for them, customized to their whims.
>Most people don't have access to anything particularly nice in the real world.
What data/metric are you drawing from to arrive at this conclusion? How could you even realistically make such a statement?
I think this would be good for the people living in overcrowded, polluted and dirty cities in the world (let's be honest, they actually exist regardless of how that happened or why).
Maybe they can unplug from 500+ AQI pollution and spend time with their loved ones and friends in a simulated clean world?
Imagine working for 10-12 hours a day, and you come home to a pod (and a building could house thousands of pods, paid for by the company) where you plug in and relax for a few hours. Maybe a few more decades of breakthroughs can lead to simulated sleep as well so they get a full rest.
Wake up, head to the factory to make whatever the developed world needs.
(holy fuck that is a horrible existence but you know some people would LOVE for that to be real)
Sounds similar to the black mirror ep https://en.wikipedia.org/wiki/Fifteen_Million_Merits
Hook it up to an always-on fle*light, and I'm sure you'd have millions of paying customers.
Except you'll never have to leave your pod. Extract the $$ from their attention all day, then sell them manufactured virtual happiness all night. It's just a more streamlined version of how many people live right now.
I'll be running away from that hellscape, thanks.
There are posters in this very thread chain who want this reality to come to pass.
I mean if I'm totally honest, it could be beneficial to me if something like this comes to be. Even in the developed world, we have a bunch of annoying people who complain/cry constantly about dumb things. They do that instead of doing something, whereas I can excuse the people trapped in hellholes overseas because it really isn't their fault they were oppressed and mistreated.
They'd have their own economy and "life" and leave the rest of us alone. It would be completely transactional, so I'd have zero reason to feel bad if they do it voluntarily.
If they can be happy in a simulated world, and others can be happy in the real world, then everyone wins!
The more I see of this, the faster I want it to go!
I'm developing filmmaking tools with World Labs' Marble world model:
https://www.youtube.com/watch?v=wJCJYdGdpHg
https://github.com/storytold/artcraft
I think we'll eventually get to the point where these are real time and have consistent representations. I've been excited about world models since I saw the in-the-browser Pokemon demo:
https://madebyoll.in/posts/game_emulation_via_dnn/demo/
At some point, we'll have the creative Holodeck. If you've seen what single improv performers can do with AI, it's ridiculously cool. I can imagine watching entertainers in the future that summon and create entire worlds before us:
https://www.youtube.com/watch?v=MYH3FIFH55s
(If you haven't seen CodeMiko, she's an incredibly talented engineer and streamer. She develops mocap + AI streams.)
>I think we'll eventually get to the point where these are real time and have consistent representations
Just like how people in the 50s thought we would have flying cars and nuclear fusion by 2000.
I have been confused for a long time why FB is not motivated enough to invest in world models, it IS the key to unblock their "metaverse" vision. And instead they let go Yann LeCun.
LeCun wasn't producing results. He was obstinate and insistent on his own theories and ideas which weren't, and possibly aren't, going anywhere. He refused to engage with LLMs and compete in the market that exists, and spent all his effort and energy on unproven ideas and research, which split the company's mission and competitiveness. They lost their place as one of the top 4 AI companies, and are now a full generation behind, in part due to the split efforts and lack of enthusiastic participation by all the Meta AI team. If you look at the chaos and churn at the highest levels across the industry, there's not a lot of room for mission creep by leadership, and LeCun thoroughly demonstrated he wasn't suited for the mission desired by Meta.
I think he's lucky he got out with his reputation relatively intact.
To be fair, this was his job description: Fundamental AI Research (FAIR) lab. Not AI products division. You can't expect marketable products from a fundamental AI research lab.
It's "Facebook Artificial Intelligence Research", not fundamental. So basically involves both fundamental and applied research.
[1]: https://engineering.fb.com/category/ai-research/
Ref: Yann lecun post on linkedin, 3years ago: FAIR now stand for "Fundamental AI Research"
Were you there or just an attentive outsider?
Attentive outsider and acquaintance of a couple people who are or were employed there. Nothing I'm saying is particularly inside baseball, though, it's pretty well covered by all the blogs and podcasts.
What podcast?
Machine Learning Street Talk and Dwarkesh are excellent. Various discord communities, forums, and blogs downstream of the big podcasts, and following researchers on X keeps you in the loop on a lot of these things, and then you can watch for random interviews and presentations on youtube when you know who the interesting people and subjects are.
Most serious researchers want to work on interesting problems like reinforcement learning or robotics or RNN or dozen other avant-garde subjects. None want to work on "boring" LLM technology, requiring significant engineering effort and huge dataset wrangling effort.
This is true - Ilya got an exit and is engaged in serious research, but research is by its nature unpredictable. Meta wanted a product and to compete in the AI market, and JEPA was incompatible with that. Now LeCun has a lab and resources to pursue his research, and Meta has refocused efforts on LLMs and the marketplace - it remains to be seen if they'll be able to regain their position. I hope they do - open models and relatively open research are important, and the more serious AI labs that do this, the more it incentivizes others to do the same, and keeps the ones that have committed to it honest.
In an industry of big bets, especially considering the company has poured resources and renamed itself to secure a place in the VR world... staking your reputation on everyone's LLMs having peaked and shifting focus to finding a new path to AI is a pretty interesting bet, no?
Since a hot take is as good as the next one: LLMs are by the day more and more clearly understood as a "local maximum" with flawed capabilities, limited efficiency, a $trillion + a large chunk of the USA's GDP wasted, nobody even turning a profit from that nor able to build something that can't be reproduced for free within 6 months.
When the right move (strategically, economically) is to not compete, the head of the AI division acknowledging the above and deciding to focus on the next breakthrough seems absolutely reasonable.
This sounds similar to the arc of Carpathy, who also managed to preserve his reputation despite sending Tesla down a FSD deadend and missing the initial LLM boat.
Isn't it more like this: JEPA looks at the video, "a dog walks out of the door, the mailman comes, dog is happy" and the next frame would need to look like "mailman must move to mailbox, dog will run happily towards him", which then an image/video generator would need to render.
Genie looks at the video, "when this group of pixels looks like this and the user presses 'jump', I will render the group different in this way in the next frame."
Genie is an artist drawing a flipbook. To tell you what happens next, it must draw the page. If it doesn't draw it, the story doesn't exist.
JEPA is a novelist writing a summary. To tell you what happens next, it just writes "The car crashes." It doesn't need to describe what the twisted metal looks like to know the crash happened.
You are beyond correct. World models is what saves their Reality Labs investment. I would say if Reality Labs cannot productize World Models, then that entire project needs to be scrapped.
Failures are not publicly reported, in general. Do you we know what they have invested in?
Most people don't like putting on VR headsets, no matter what the content is. It just never broke out of the tech enthusiast niche.
Reminds me of this [1] HN post from 9 months ago, where the author trained a neural network to do world emulation from video recordings of their local park — you can walk around in their interactive demo [2].
I don't have access to the DeepMind demo, but from the video it looks like it takes the idea up a notch.
(I don't know the exact lineage of these ideas, but a general observation is that it's a shame that it's the norm for blog posts / indie demos to not get cited.)
[1] https://news.ycombinator.com/item?id=43798757
[2] https://madebyoll.in/posts/world_emulation_via_dnn/demo/
Yup, similar concepts! Just at two opposite extremes of the compute/scaling spectrum.
- That forest trail world is ~5 million parameters, trained on 15 minutes of video, scoped to run on a five-year-old iPhone through a twenty-year old API (WebGL GPGPU, i.e OpenGL fragment shaders). It's the smallest '3D' world model I'm aware of.
- Genie 3 is (most likely) ~100 billion parameters trained on millions of hours of video and running across multiple TPUs. I would be shocked if it's not the largest-scale world model available to the public.
There are lots of neat intermediate-scale world models being developed as well (e.g. LingBot-World https://github.com/robbyant/lingbot-world, Waypoint 1 https://huggingface.co/blog/waypoint-1) so I expect we'll be able to play something of Genie quality locally on gaming GPUs within a year or two.
Are world models from the perspective of an observer in the world or zoomed out?
Or in gaming terms do these models think FPS or RTS?
Text models and pixel grid vision models is easy but struggling to wrap my head around what world model "sees" so to speak.
I keep on repeating myself, but it feels like I'm living in the future. Can't wait to hook this up to my old Oculus glasses and let Genie create a fully realistic sailing simulator for me, where I can train sailing with realistic conditions. On boats I'd love to sail.
If making games out of these simulations work, it't be the end for a lot of big studios, and might be the renaissance for small to one person game studios.
Isn't this still essentially "vibe simulation" inferred from videos? Surface-level visual realism is one thing, but expecting it to figure out the exact physical mechanics of sailing just by watching boats, and usefully abstract that into a gamified form, is another thing entirely.
Yeah I have a whole lot of trouble imagining this replacing traditional video games any time soon; we have actually very good and performant representations of how physics work, and games are tuned for the player to have an enjoyable experience.
There's obviously something insanely impressive about these google experiments, and it certainly feels like there's some kind of use case for them somewhere, but I'm not sure exactly where they fit in.
Why wouldn't it just hook it into something like physx?
Google has made it clear that Genie doesn't maintain an explicit 3D scene representation, so I don't think hooking in "assists" like that is on the table. Even if it were, the AI layer would still have to infer things like object weight, density, friction and linkages correctly. Garbage in, garbage out.
The bottleneck for games of any size is always whether they are good. There are plenty of small indies which do not put out good games. I don't see world models improving game design or fun factors.
If I am wrong, then the huge supply of fun games will completely saturate demand and be no easier for indie game devs to stand out.
It's very impressive tech but subject to the same limitations as other generative AI: Inconsistency, inaccurate physics, limited time, lag, massively expensive computation.
You COULD create a sailing sim but after ten minutes you might be walking on water, or in the bath, and it would use more power than a small ferry.
There's no way this tech can run on a PS5 or anything close to it.
Five years is nothing to wait for tech like this. I'm sure we will see the first crop of, however small, "terminally plugged in" humans on the back of this in the relatively near future.
You raise good points, but I think the “it’s not good enough” stance won’t last for long.
> and might be the renaissance for small to one person game studios.
Indie games are already bigger than ever as far as I know.
> If making games out of these simulations work, it't be the end for a lot of big studios, and might be the renaissance for small to one person game studios.
I mean, if making a game eventually boils down to cooking a sufficient prompt (which to be clear, I'm not talking about text, these prompts are probably going to be more like video databases) then I'm not sure if it will be a renaissance for "one person game studios" any more than AI image generation has been a renaissance for "one person artists".
I want to be optimistic but it's hard to deny the massive distribution stranglehold that media publishing landscape has, and that has nothing to do with technology.
Honestly getting a Sunfish is probably cheaper than the a VR headset if you want to "train sailing"
...and then, the pneumatics in your living room.
I have no idea why Google is wasting their time with this. Trying to hallucinate an entire world is a dead-end. There will never be enough predictability in the output for it to be cohesive in any meaningful way, by design. Why are they not training models to help write games instead? You wouldn't have to worry about permanence and consistency at all, since they would be enforced by the code, like all games today.
Look at how much prompting it takes to vibe code a prototype. And they want us to think we'll be able to prompt a whole world?
This was a common argument against LLMs, that the space of possible next tokens is so vast that eventually a long enough sequence will necessarily decay into nonsense, or at least that compounding error will have the same effect.
Problem is, that's not what we've observed to happen as these models get better. In reality there is some metaphysical coarse-grained substrate of physics/semantics/whatever[1] which these models can apparently construct for themselves in pursuit of ~whatever~ goal they're after.
The initially stated position, and your position: "trying to hallucinate an entire world is a dead-end", is a sort of maximally-pessimistic 'the universe is maximally-irreducible' claim.
The truth is much much more complicated.
[1] https://www.arxiv.org/abs/2512.03750
And going back a little further, it was thought that backpropagation would be impractical, and trying to train neural networks was a dead end. Then people tried it and it worked just fine.
> Problem is, that's not what we've observed to happen as these models get better
Eh? Context rot is extremely well known. The longer you let the context grow, the worse LLMs perform. Many coding agents will pre-emptively compact the context or force you to start a new session altogether because of this. For Genie to create a consistent world, it needs to maintain context of everything, forever. No matter how good it gets, there will always be a limit. This is not a problem if you use a game engine and code it up instead.
Imo they explain pretty well what they are trying to achieve with SIMA and Genie in the Google Deepmind Podcast[1]. They see it as the way to get to AGI by letting AI agents learn for themselves in simulated worlds. Kind of like how they let AlphaGo train for Go in an enormous amount of simulated games.
[1] https://youtu.be/n5x6yXDj0uo
"I need to go to the kitchen, but the door is closed. Easy. I'll turn around and wait for 60 seconds." -AI agent trained in this kind of world
That makes even less sense, because an AI agent cannot learn effectively from a hallucinated world without internal consistency guarantees. It's an even stronger case for leveraging standard game engines instead.
If that's the goal, the technology for how these agents "learn" would be the most interesting one, even more than the demos in the link.
LLMs can barely remember the coding style I keep asking it to stick to despite numerous prompts, stuffing that guideline into my (whatever is the newest flavour of product-specific markdown file). They keep expanding the context window to work around that problem.
If they have something for long-term learning and growth that can help AI agents, they should be leveraging it for competitive advantage.
Take the positive spin. What if you could put in all the inputs and it can simulate real world scenarios you can walk through to benefit mankind e.g disaster scenarios, events, plane crashes, traffic patterns. I mean there's a lot of useful applications for it. I don't like the framing at this time, but I also get where it's going. The engineer in me is drawn to it, but the Muslim in me is very scared to hear anyone talk about creating worlds.... But again I have to separate my view from the reality that this could have very positive real world benefits when you can simulate scenarios. So I could put in a 2 pager or 10 page scenario that gets played out or simulated and allow me to walk through it. Not just predictive stuff but let's say things that have happened so I can map crime scenes or anything. In the end this performance art is because they are a product company being Benchmarked by wall street and they'll need customers for the technology but at the same time they probably already have uses for it internally.
> What if you could put in all the inputs and it can simulate real world scenarios you can walk through to benefit mankind e.g disaster scenarios, events, plane crashes, traffic patterns.
This is only a useful premise if it can do any of those things accurately, as opposed to dreaming up something kinda plausible based on an amalgamation of every vaguely related YouTube video.
> What if you could put in all the inputs and it can simulate real world scenarios you can walk through to benefit mankind e.g disaster scenarios, events, plane crashes, traffic patterns.
What's the use? Current scientific models clearly showing natural disasters and how to prevent them are being ignored. Hell, ignoring scientific consensus is a fantastic political platform.
An hybrid approach could maybe work, have a more or less standard game engine for coherence and use this kind of generative AI more or less as a short term rendering and physics sim engine.
I've thought about this same idea but it probably gets very complicated.
Let's say, you simulate a long museum hallway with some vases in it. Who holds what? The basic game engine has the geometry, but once the player pushes it and moves it, it needs to inform the engine it did, and then to draw the next frame, read from the engine first, update the position in the video feed, then again feed it back to the engine.
What happens if the state diverges. Who wins? If the AI wins then...why have the engine at all?
It is possible but then who controls physics. The engine? or the AI? The AI could have a different understanding of the details of the base. What happens if the vase has water inside? who simulates that? what happens if the AI decides to break the vase? who simulates the AI.
I don't doubt that some sort of scratchpad to keep track of stuff in game would be useful, but I suspect the researchers are expecting the AI to keep track of everything in its own "head" cause that's the most flexible solution.
Then maybe the engine should be less about really simulating the 3D world and just trying best to preserve consistency, more about providing memory and saving context for consistency than truly simulating a lot besides higher level concerns (at which point we might wonder if it couldn't be directly part of the model somehow), but writing those lines I realize there would probably still be many edge cases exactly like what you are describing...
As a kid in the early 1980s, I spent a lot of time experimenting with computers by playing basic games and drawing with crude applications. And it was fun. I would have loved to have something like Google's Genie to play with. Even if it never evolved, the product in the demos looks good enough for people to get value from.
It's been very profitable for drug dealers for centuries, who wouldn't want a piece of that market?
Because games already exist, and it would be easier for LLMs to write games rather than hallucinate videos.
Same with Nvidia's Earth 2 and Cosmos (and a bit like Isaac). Games or VR environments are not the primary drive, the primary drive is training robots (including non-humanoids, such as Waymo) and just getting the data. It's exactly because of this that perfect physics (or let's be honest, realistic physics[0,1]). Getting 50% of the way there in simulation really does cut down the costs of development, even if we recognize that cost steepens as we approach "there". I really wish they didn't call them "world models" or more specifically didn't shove the word "physics" in there, but hey, is it really marketing if they don't claim a golden goose can not only lay actual gold eggs but also diamonds and that its honks cure cancer?
[0] Looking right does not mean it is right. Maybe it'll match your intuition or undergrad general physics classes with calculus but talk to a real physicist if you doubt me here. Even one with just an undergrad will tell you this physics is unrealistic and any one worth their salt will tell you how unintuitive physics ends up being as you get realistic, even well before approaching quantum. Go talk to the HPC folks and ask them why they need superocmputers... Sorry, physics can't be done from observation alone.
[1] Seriously, I mean look at their demo page. It really is impressive, don't get me wrong, but I can't find a single video that doesn't have major physics problems. That "A high-altitude open world featuring deformable snow terrain." looks like it is simulating Legolas[2], not a real person. The work is impressive, but it isn't anywhere near realistic https://deepmind.google/models/genie/
[2] https://www.youtube.com/watch?v=O4ZYzbKaVyQ
But it's not simulating, is it? It's hallucinating videos with an input channel to guide what the video looks like. Why do that instead of just picking Unreal, Unity, etc and having it actually simulated for a fraction of the effort?
Why is it a dead end, you don’t meaningfully explain that. These models look like you can interact with them and they seem to replicate physics models.
They don't though, they're hallucinated videos. They're feeding models tons and tons of 2D videos and hoping they figure out physics from them, instead of just using a game engine and having the LLM write something up that works 100% of the time.
This is what we were building in 2018 with Ayvri, starting from 3d tiles with the aim of building a real-world view by using AI to essentailly re-paint and add detail to what was essentially a high-resolution and faster loading Google Earth (for outside cities, we didn't have building data).
We saw a very diverse group of users, the common uses was paragliders, gliders, and pilots who wanted to view their or other peoples flights. Ultramarathons, mountain bike and some road-races where it provided an interactive way to visualize the course from any angle and distance. Transportation infrastructure to display train routes to be built. The list goes on.
This is a fascinating project. The idea of infinite interactive worlds is a huge leap for gaming and simulation.
I don't know ... it's impressive and all but the result always looks kind of dead.
This reminds me of the comments by programmers roughly two years ago:
"Sure it can write a single function but the code is terrible when it tries to write a whole class..."
You say that as if those programmers aren't still right
It's super cool but I see it as a much more flexible open ended take on the idea of procedurally generated worlds where hard-coded deterministic math and rendering parameters are replaced by prompt-able models.
The deadness you're talking about is there in procedural worlds too, and it stems from the fact that there's not actually much "there." Think of it as a kind of illusion or a magic trick with math. It replicates some of the macro structure of the world but the true information content is low.
Search YouTube for procedural landscape examples. Some of them are actually a lot more visually impressive than this, but without the interactivity. It's a popular topic in the demo scene too where people have made tiny demos (e.g. under 1k in size) that generate impressive scenes.
I expect to see generative AI techniques like this show up in games, though it might take a bit due to their high computational cost compared to traditional procedural generation.
Best case, Google DeepMind cracks AGI by letting agents learn for themselves inside simulated worlds. Worst case, they've invented the greatest, most expensive screensaver generator in human history.
This could be the future of film. Instead of prompting where you don't know what the model will produce, you could use fine-grained motion controls to get the shot you are looking for. If you want to adjust the shot after, you could just checkpoint the model there, by taking a screenshot, and rerun. Crazy.
I feel like people are already currently doing this. Essentially storyboarding first.
This guy a month ago for example: https://youtu.be/SGJC4Hnz3m0
Compared to DeepMind's Genie 3 demo, this appears to have more morphing issues and less user interactivity with environmental consistency. Is this a stripped down version?
Every character goes forward only, permanence is still out of reach apparently.
Looks good to me: https://youtu.be/15KtGNgpVnE?t=648&si=urWJGEFWuN5veh43
I've been experimenting with that from a slightly different angle: teaching Claude how to play and referee a pencil-and-paper RPG that I developed over about 20 years starting in the mid 1970s. Claude can't quite do it yet for reasons related to permanence and learning over time, but it can do surprisingly well up until it runs into those problems, and it's possible to help it past some obstacles.
The game is called "Explorers' Guild", or "xg" for short. It's easier for Claude to act as a player than a director (xg's version of a dungeon master or game master), again mainly because of permance and learning issues, but to the extent that I can help it past those issues it's also fairly good at acting as a director. It does require some pretty specific stuff in the system prompt to, for example, avoid confabulating stuff that doesn't fit the world or the scenario.
But to really build a version of xg on Claude it needs better ways to remember and improve what it has learned about playing the game, and what it has learned about a specific group of players in a specific scenario as it develops over time.
Damn that was crazy the picture of the tabletop setup/cardboard robot and it becomes 3D interactive.
Google Deepmind Page: https://deepmind.google/models/genie/
Try it in Google Labs: https://labs.google/projectgenie
(Project Genie is available to Google AI Ultra subscribers in the US 18+.)
So what is it doing in the real world, microwaving an elephant on high with 80kw every second and pouring out all the water in an sub-saharan African well every 4 minutes?
It's ability to simulate physics intact is actually a huge breakthrough.
I can't even fathom what it would be like for the future of simulation and physical world when it gets far more accurate and realistic.
Ironically the physics are kind of my biggest criticism. They call these "world models", but I think it's more accurate to call them "video game models" because they employ "video game physics" rather than real world physics, among other things
This is most evident in the way things collide.
It's getting better staggeringly fast, just a year ago I wouldn't expect it to be at even video game physics level so quickly.
If there is a possibility where it continue to improve at a similar rate with llms. A way to simulate fluid dynamics or structural dynamics with reasonable accuracy and speed can unlock much faster pace of innovation in the physical world. (And validated with rigorous scientific methods)
I am stumped. Am I misreading, or are the folks at Google deliberately confounding two interpretations of "world model"? Dont get me wrong, this is really cool, and it will undoubtedly have its use. But what I am seeing is an LLM that can generate textures to be fed into a human-coded 3d engine (the "world model" that is demonstrated), and I fail to see how that brings us closer to AGI. For AGI we need "world models" as in "belief systems". The AI model must be able to reason about (learned) dynamics, which I dont see reflected in the text or video.
>an LLM that can generate textures to be fed into a human-coded 3d engine
I'm not certain but I think the LLM is also generating the physics itself. It's generating rules based on its training data, e.g. watch a cat walk enough and you can simulate how the cat moves in the generated "world".
let's reboot Leisure Suit Larry ;-)
This is the plot of The Peripheral, right? Love the way the second half of that book turned out. Never finished Agency..
Anyone else going to try it and just keep getting a 404 page?
Finally all my anime figurines will come to life
This would be really cool if polished and integrated with VR.
Exactly this, it would essentially be a STTNG Holodeck
What’s the endgame here? For a small gaming studio, what are the actual implications?
The endgame has nothing to do with gaming.
The goal of world models like Genie is to be a way for AI and robots to "imagine" things. Then, they could practice tasks inside of the simulated world or reason about actions by simulating their outcome.
It means you should go the other way. Open world winning against smaller, handcrafted environments and stories was generally a mistake, and so is this.
What does it mean, that open world winning was a mistake? That the market is wrong, and peoples' preferences were incorrect, and they should prefer small handcrafted environments instead of what they seem to actually buy?
It seems to be generating images in real time, not 3d scenes. It might still be useful for prototyping.
There are collisions though and physics seemingly, so it doesn't seen to be a huge stretch that this could be used for games.
I would think that building a environment which can be managed by a game engine is the first pass. In a few years when we are able to render more than 60 seconds it could very well replace the game engine entirely by just rendering everything in realtime based on user interactions. The final phase is just prompts which turn directly into interactive games, maybe even multiplayer. When I see the progress we've made on things like DOOM, where it can infer the proper rendering of actions like firing weapons and even updating scores on hits and such it doesn't feel like we're very far off, a few years at most. For a game studio that could mean cutting out almost everything between keyboard and display, but for now just replacing the asset pipeline is huge.
We seem to think that Genie is good at the creative part, but bad at the consistency and performance part. How hard would it be to take 60 seconds of Genie output and pipe it into a model that generates a consistent and performant 3D environment?
I understand the ultimate end goal to be simulation of life. A near perfect replica of the real world we can use to simulate and test medicine, economy, and social impact.
Screensavers for robots?
This is as good of a place to mark it as any.
Humanity goes into the box and it never comes back out. It's better in there than it is out there for 99% of the population.
>>How we’re building responsibly
How are you justifying the enormous energy cost this toy is using, exactly?
I don't find anything "responsible" about this. And it doesn't even seem like something that has any actual use - it's literally just a toy.
Demis stays cooking
everyone will make his own game now
If only Google had the technology for game streaming... Oh wait
RIP Stadia.
Stadia was lightyears ahead, but pro-Microsoft media assassinated it with FUD
While "journalists" were busy bootlicking a laggy 720p Android only xCloud beta, Stadia was already delivering flawless 4K@60FPS in a web browser
They killed the only platform that actually worked just to protect Microsoft
This will be a textbook case study in how a legacy monopoly kills innovation to protect its own mediocrity
Microsoft won't survive the century, they are a dinosaur on borrowed time that has already lost the war in mobile, AI, and robotics
They don't create,, they just buy marrket share to suffocate the competition and ruin every product they touch
Even their cloud dominance is about to end, as they are already losing their grip on the European market to antitrust and sovereign alternatives
If creating an infinite world is so trivially easy (relatively speaking) then occam suggests that this world is generated.
I don't know if that's the simplest explanation, considering how insanely complex the generation is (these world models might literally be the most complex things to ever be created).
But I do think it's a partial existence proof.
One big simplifier is to only render what you're looking at. I wonder how one might demonstrate that.
I now believe we live in a simulation
clearly
We will probably see Ready Player One in a few decades. Hoping to stay alive till then.
The mass-poverty and climate changed ravaged world parts, I could definitely see.
Decades?
I mean, yes, the probability of having that level of tech in decades is quite high.
But the technology is moving very fast right now. It sounds crazy, but I think that there is a 50% chance of having ready player one level technology before 2030.
It's absolutely possible it will take more time to become economical.