Genie 4 (deepmind.google)

It's like its simulation model is confined to the renderable area. At the cost of increased computation, it would not be hard to generalize the technique to rendering a sphere centered around the POV to keep some of those artifacts at bay; however, the fundamental limitation of "object permanence" remains if objects can evolve chaotically as soon as they are blocked from view by another object. For that to be made stable, you really need a whole 3D model of the scene, and the cost increase there might be huge.

> While it can maintain coherent environments over short sessions, visual details and object permanence can still shift or drift over time—especially as a simulation runs longer, after multiple complex prompt edits, or when navigating throughout the scene. Complex physical interactions, real-time audio accuracy, rendering legible text, spatial accuracy, and long-term consistency are active areas of ongoing research.

It just doesn't have object permanence yet!

And these are the examples they chose. I have to wonder if they add some with mistakes on purpose.

> It's very good at maintaining consistency for things that stay in frame

I noticed several inconsistencies in things that stayed in frame. See the wall on the bridge in the Van Gogh example as they walk past it to see the other side.

I don’t understand why these world models aren’t designed to be backed by an updatable world state. It could probably just lay out the world onetime and reference it as needed vs trying to invent from wholecloth via diffusion. Object persistence!

I bet it's configurable and just a matter of how much resource you want to allocate. If it's truly a world model then it seems to me that it's precise benefit is that it can keep track and order in things and they just choose to accept some noise/chaos/randomness/? for the demos when things move out frame? Does that sound sensible? I'm just guessing but don't feel it's a hard limit.

Impressive... at populating our visual world with more slop? Sure.

Would this be impressive in terms of convincing you to buy something or teach you something or entertain you?

Genuinely asking.

I'm by no means anti-anything but I think what the comments are getting at is that these "world models" do not have a world model. It's a fundamental qualitative gap that no amount of compute has been able to close, just mask by everything else getting better.

This doesn't show up as much where the verifiable reward is "code compiles and tests pass" but you can still absolutely see it in things like movies, prose, or slide decks. Even frontier models are pretty bad at keeping track of narrative threads even as the technical quality drastically improves.

This begs the question of whether the gap can be closed just with more compute (presumably yes, or at least errors can be minimized beyond human capacity to see them) and is that going to be worth it with current tech (possibly no).

I agree. This is nothing short of magic -- or perhaps 95% of magic. Imagine yourself 5 years ago, and someone from today going back in time and telling you "hey, check this out, in just 5 years you'll be able to write this short text and a computer will generate this movie from it"...

People are emotionally driven. When emotions drive us to do something we will (unconsciously) find plausible justifications for it without realizing we're doing it.

Anti-AI crowd is anti AI. Maybe they were (mostly) happy with status quo and don't like change. Maybe scared for the future. Or any other reason. Went from "AI can't even tell what 2+2 is", to "AI can't even tell whether these simply connected smooth 4-manifolds are merely homeomorphic or actually diffeomorphic after stabilization" without batting an eye. The goalposts will keep being moved and the dismissal will continue.

It's a huge leap from 3 years ago, but not a huge leap from other current AI models. People who are deep in the AI rabbit hole have seen impressive generative video that has fewer problems (or possibly just better chosen examples).

My problem with it is the shift in focus. The original Genie was pitched as a simulator of the world that an agent can interact with to learn so generally that it could be deployed into the real world. This new version looks like a good, controllable video generator. The fact that you're comparing it to the will smith video, rather than a simulator, is indicative of the fact they've strayed from that original goal. I think that's a shame

Perfection doesnt exist - technology will never free us from the human condition & constant struggles...

I watched it on mute, and even so I figured it didn't sound great, given that the guy's fingers barely move at all.

I have no idea if that's the case, but it would be funny if it didn't support negative prompting and they got someone to write the prompt who wasn't aware of that.

Edit: Okay, I read the prompts and they use negative phrasing a lot, so I guess it does work, but boy, are these prompts laborious. Requirements are repeated ad nauseam. This model seems to be quite tone deaf.

It's probably trained in a video game-like environment.

When people talk about "photorealism" it's to describe something not real (like video game graphics and artwork) that looks almost like a real photo/video. So you don't say that a real photo is photorealistic, and if you use that word when prompting models you end up making it not realistic.

Probably too much of the training data that has Magic wands were from a video game.

I might not understand what "world model" means, but from my understanding world model is the model trained on real world data, so theoretically having all laws of physics baked in the weights. So perfect world model should not be able to make the video for given prompt, as it is impossible in real world and the only possible way it can appear is in a setting of video game or movie. So this is expected, doesn't mean that is the reason behind this particular generation.

> Sometimes I really wanna skip 10 years into the future and see where all this eventually leads.

Sounds glib, but honestly ... more Ads :/

  > ... more Ads

> movies are going to become video games and video games are going to become something unimaginably better

This is the kind of detached from reality, devoid of humanity statement only a tech bro could come up with.

Why wait? It's become very clear already: https://kotaku.com/the-proliferation-of-copycats-on-steam-ma...

Instead of better it's become unimaginably worse since games can no longer serve as a means of self expression because the second you express anything worthwhile 100 people copy it immediately ... without you. It's like Disney (whose whole business model is copying public domain stories and music and ... but taking out the source and everyone who ever had anything to do with it). Except there's no longer one highly selective company destroying the world's culture one piece at a time, but 10000000 idiots looking to make a quick buck, destroying culture and self-expression like a fire raging in the rainforest.

There will be no more "celeste" games where you join for the platforming but stay for the (politically offensive to all sides) story, because AIs won't discuss the story as it's politically incorrect. Celeste is fundamentally about depression and anxiety, due to being transgender, and is about offending both sides of the argument. First, the main character is transgender, with zero explanation or need for explanation how far that goes, which of course offends conservatives. Second the goal of the main character is to run away from any provided help and professionals (which the director did too) and even under the most extreme of circumstances only reluctantly accept help from a family member that then promptly dies making things "worse".

The Qwen and other Chinese models (and ...) will refuse to discuss being transgender at all, and any western models will shut down any discussion of running away from help hard (because otherwise they get held responsible for the consequences, as has already happened)

(oh and that running away from professional help in psychological issue is actually observed to be a positive thing in almost all cases, in studies. The main way to make psychological problems worse is to let them fester and running away from help is time and time again observed to be far more productive than going to professionals and expecting them to fix things. Because that's an extreme human bias: it can never be (allowed to be) the right reaction to NOT do something about a problem, especially when countless people are paid to "solve" the problem. In other words: if you take people who expressed intent to commit suicide and compare the groups who "get help" and the group that successfully shut down help, less people actually go through with it in the second group, and worse than that, the more intense the help, the better the odds the patient ends up killing themselves. Right to the point that doctors sometimes assist such patients because after 2 years of attempts people live in constant misery and have caused enough medical damage that they will have constant pain or worse the rest of their life. And yet people keep forcing help on patients. Usually in response to an actual suicide where no help was provided. Then more help gets built, and 5 years later we yet again observe ... more suicides. But tell someone with an actual problem to avoid help and you'll be skinned alive for that)

So AI could never make a game like that. Not because it actually can't, or isn't capable, but because there already is a strong political bias, in every model you might use to make it and neither side is going to back down on their policies. The Chinese Communist Party is not suddenly going to change it's mind about transgenderism. And the western models are not suddenly going to openly discuss running away from psychological help, not even as part of a story.

So really, we've just lost these games.

So, people will be able to completely skip living in the real world, and instead spend their entire time plugged into a simulation, like a vegetable.

Sounds wonderful. May as well skip this and just spend my life hooked up to an IV drip of drugs.

    simulation becomes real enough

Type of thing he would say.

Llms wont be able to create good games. Consider for a sec that there's like 500 "Great" games ever made, maybe more but not much, in some genres that number is less than 10. There is no training data, and its largely a matter of taste and culture that cant be represented/distilled into vector embeddings.

Video games are already movies that are video games.

10yrs isn't far enough. 2016 wasn't that different from today relative to media.

> movies are going to become video games and video games are going to become something unimaginably better

God what an annoying thing to say. Suggesting that movies will essentially become games, or that future video games will be unimaginably better (presumably because they are driven by generative world models) betrays such a surface level appreciation for what makes art and culture and all this humanity so beautiful as it is.

The entire point of movies is the deliberate choices of the filmmakers! Movies will not be improved by our ability to choose what the main character looks like or add a style filter. Super Mario 64 is great because it is an expression of the very hardware that constrains it. Inevitable future versions where anything can happen because someone hooked the ROM up to a generative world model won't be better for it.

I agree that this technology is absolutely world changing, and am cautiously eager to see what comes of it, but it just bums me out, perhaps needlessly, that the people at the helm can be so artless, unreflective. Whatever.

  "What is grief, if not love persevering?"

It will lead nowhere. Graphics intensive video games as interactive environments benefit from efficient rendering done by hardware a fraction of the cost of what a neural network requires for the same output. This is a business dead end.

It will lead to garbage.

Also, does anyone want movies to become games? I don't. I want movies to be movies. Good movies. With interesting stories, beautiful artistic cinematography, great characters, moving paintings as Lynch said it.

If I want games, I'll play games.

With VR coming and then going technology has shown itself to hold within it no more immersive capability than a good book. It is not about rendering firepower it is about hitting the right notes.

For me, this does nothing. I am attracted to film and gaming for the artistry and deliberate choices made by the authors. A synthetic approach may be stimulating but I find it ultimately void.

Like, movies sometimes have alternate endings. What if they were more choice driven, and replayable. Video games have remarkably dumb ideas that keep persisting, like you can encapsulate all that is “Health” as an integer. Instead games can be an aggregation of entity bound micro-simulation that create emergent scenarios, yet follow the laws of narrative collectively. Games that build themselves. Worlds within worlds kinda jazz.

Great, I guess I can say bye bye to movies

> Sometimes I really wanna skip 10 years into the future and see where all this eventually leads.

Either the desert of the real, or something unimaginable to Mk 1 humans.

Is it surprising? Looking at the demo videos I can absolutely see why it didn't get as much attention. Not that it's not impressive, it is, but these are in different leagues.

I was under the impression he has already fly past the catapult.

That entire clip looks like it's straight out of a complete bargain-bin videogame. The weird smoothness of the ride, the pivot in mid air, the wings apparently don't have collision with the poles they fly past.

Definitely wouldn't have been one of my picks to show off my fancy new model.

There's also a whole extra bonus castle that appears to the left when the camera looks left around 0:22. Which is amusing both because it wasn't there on the initial approach, and also because nobody puts two castles each with their own distinct curtain walls right next to each other like that.

The 80's Japanese train station clip also had the conductor speaking clear english in an accent I can't quite place. Slight British tinge to it but not really? Not sure. The ambient audio also makes no sense.

I've seen weird stuff in images that OpenAI sends me in it's promotional emails too. I would think that if you were trying to show off you'd make sure there's none of these obvious AI issues in what you choose to show off with. Maybe they don't really care? "Gemini choose 10 scenes that would be impressive for our announcement and add them to the post."

It is always incredible to me how quickly we go from "this entire technology is pure science fiction" to "nitpicking over details"

The oddest thing about this, to me at least, is that they pick that as their showcase. Did they themselves not notice? Or did they think their potential customers wouldn’t? Not sure really which of the two is worse.

In the original dragon clip, at 0:29 we can see a square tower in the right side of the screen

As he passes by the tower, it becomes round, by 0:40

Elon revealed to the world that you can just lie and people will give you money, so now they all lie.

> Cool demos, but does anybody else find it troubling that companies keep claiming this tech can train robots? It’s so obviously not representative of the real world.

A proprietary version of this technology is already training Waymo's robot cars. Waymo blogged about their DeepMind-provided world model a year ago.

> It’s so obviously not representative of the real world.

That entirely depends on the use-case and level of granularity needed, doesn't it?

This stands out to me too, it seems to be representative of physics the way a video game is representative of physics, likely less so given the non-determinism

Subreddits where SolidGoldMagikarp was counting to a million in comments were used to train a machine that then writes web pages that more machines are trained on. The sheer volume of data with sufficient verisimilitude generated by these models may well work to train robots. It might not, but I think it might well work. The idea that things have to be high-fidelity high-quality to yield good results has clearly not held up in other domains.

I guess they want military contracts or something.

You can get useful levels of robot training out of Super Mario Bros 2; see John Carmack's talk from last year: https://www.youtube.com/watch?v=4epAfU1FCuQ

When your training simulation fails to reproduce certain features of the real-world environment, your control model is likely to have trouble with them, so more realistic simulations are helpful, but even fairly unrealistic training simulations are pretty useful.

In general, training neural networks to control robots is going to require simulations, ideally differentiable simulations. Even if they aren't differentiable, though, you can SGD your simulated training runs across a data center in a way that you just can't with training runs on Roombas.

Gemini Robotics is a thing: https://deepmind.google/models/gemini-robotics/

They clearly know how to train robots. Unless Genie is a misdirection...

probably meant "audio like music, ambient noises, and interactive dialogue."

Audio like: music, ambient-noises etc

As opposed to video-like music...

That was probably meant to be an em-dash, or have spaces around the hyphen.

Have you listened to "music" created by non-specialized models that use WebAudio to create?

It seems like a good descriptor to me! ;)

I'm blessed to see that even google let their most important marketing prosa be generated by ai models without any proofreading from fleshy humans.

By Genie specifically? Or by generative AI in general?

If you mean world models like Genie,I don't know. If you mean the latter, I have definitely used software that I have enjoyed that I do not believe would have been possible without generative AI.

Fluxer is the first Discord competitor that when you show to someone they can't say "well Discord does X better", and it was largely made using AI.

Herdr is designed for AI usage, but I actually really like it even for just normal terminal multiplexing. It was surely made with a lot of AI.

I have seen people make incredible things at my work that have actually solved multi-year long pain points that we've never had the resources to address before.

I have used AI to patch and revive a bunch of small services like chat bots or web services and such that I had made and had decayed into failure over the years.

Sure, it's easy to point at people using AI as a megaphone for their lack of care. But I don't think it's impossibly hard to find people doing good things with it either.

Unironically, Tung Tung Tung Sahur and some funny music video about p(doom). But yeah that’s basically it.

Throw a gifted storyteller at any medium and they will produce something worth paying attention to. AI generated content has not had that moment yet. I'm not sure it will ever, but it's worth thinking about.

It took the medium of film ~50 years to find Hitchock and ~75 to find Spielberg.

They're not really promising to do that. It's not even clear this model will be meaningfully commercialized. The key line is "Understanding and simulating the world is a crucial aspect of building AGI". This model will be used for research and robotics, and maybe the training of further models.

There are tons of cool, creative, funny and wow stuff being done with AI. A third of my Instagram feed is AI creators and some of the stuff they do is mind blowing, but even when it's not super impressive, some of it is just so creative and well done. Saiyenfeld is amazing (DBZ Seinfeld mashup by grickusai). It's taking a medium that is familiar, putting a spin on it and doing it well. You can do things that you couldn't do otherwise. People with interesting concepts can finally get them out of their head into something really interesting. People are so doomer and depressed, but they purposely don't see the stuff that is just fucking fun. I spend all day with a 5090 and courses and I still can't make stuff as good, creative, cool or funny as these people who actually have those talents and they are just fuddling through with token services.

If you want to get inspired go to civitai and look through the top stuff and be wowed. Maybe they didn't draw it, but they figured out how to make it happen and when you look at enough generated stuff you can tell the average from the holy shit. It's easy to get inspired by what people can come up with and want to do something yourself, which is a positive impact I'd say.

I agree, I don't understand why this technology has so much energy put into it. Just because we can doesn't mean we should?

What an insane set of requirements for a frontier technology and your excitement by it.

> "which is safetymaxxed to the limit,"

I feel refusal'd just *looking* at this page.

I don't get how Google is so horrible at product. Let me play with this. Otherwise it's DOA.

I’ve decide that the best (and maybe only) way for me to interact with google AI models, is on the search results

It’s where I get the most value, and my expectations are very low, no attachment if it goes away

Of course they are mining me for data anyway, but they were doing that before already (and I do have a Gmail account, although I’ve almost completely stopped using Chrome - btw, it’s crazy that you can’t add plugins/extensions to Chrome on Guest mode)

I'm not entirely convinced the new models aren't just a rollback to a clean training data state after the current model has been released for a while.

I watched all of that video, it was really good.

That's the thing about machine learning over time, and the big hole in the whole world model idea. Time and time again we make simulations and train models and we think that loop will work. It never does.

And why not? Because the first model that can actually learn to handle the simulation not only can do the simulation internally too, but it can do it better than out simulation.

It's where (ex-)"AI luminaries" go to die. It's a way to make glacial, slow progress in the correct direction while never actually getting anywhere.

So we get to a point and people keep making complicated simulations of one kind or another ... and then someone (some model) who's never used the simulation, and doesn't need it for model training gets by far the highest score on every benchmark and everyone agrees the simulation was kind of beside the point.

They were demonstrating steering.

How often do things in the real world not behave exactly as you expected them to?

I feeling I get from this is that this could be the mental model for a robot in the real world, continuously corrected by actual events.

We all have a world model in our heads of what we think is going to happen, we adjust quickly (or not) based on what does. This is pretty incredible stuff.

It also failed to properly model the physics of knights riding wyverns into battle.

This is a reflection of a complaint I frequently make about so called "world models". It's more accurate to call them video game models. They don't simulate the world, they simulate video games. For that reason a common issue with them is that the physics are video game physics rather than real world physics.

yes. I feel Google's video physics has always been bad. slow, aritifical, unnatural.

I agree but I feel like the exaggerated physics were intentional and fitting with the video theme. If they had wanted it to topple, they wouldn't have posted the video.

People are arguing if the marble bust should've toppled over or not, completely missing the fact that marble is brittle and the bust would have surely cracked

It prompts "gigantic alien tentacles that destroy the world."

I suspect you're reading the base prompt, but the video also demonstrates adding further prompts to change what is happening as the scene progresses.

It was the complete lack of fear or urgency on the part of the characters for me.

But remember, today is the worst the models will ever be. I agree they are still safe though... they will probably use AI to make even more films faster and more creative since it enables them to do more.

Yeah 2020 was the upload point when we were all corralled during covid. And trees were the data centers of the dinosaurs. I mean, what if?

Isn't that a simpler explanation of reality than the zillions of particles all existing simultaneously in this massive universe?

Some sort of simple primal generative AI construct that just feeds a conscious experience for one. You. Nothing else is real. No one else is real.

Something like that would be an infinitely simpler, than running the physics and conscious experiences of this unimaginably large universe we think we're in.

Google is already public, has a larger more normie customer base, many products to integrate AI in with established enterprise accounts in the millions... For some reason theyre not valued at 50:1 revenue... I dont think all that many people are going to be that excited for Anthropic as a business, Its an escape hatch, using the public as exit liquidity while the founders can still make a 100bil.

That actually sounds implementable even now! Except the fully consistent personality part. And maybe that's a good thing. Falling in love with beings without physical presence sounds agonizing; probably better spent on real people.

Question is if we even want that. Video games often are escapism, we want an illusion of reality but not actual reality. I mean there definitely is a market for that, but I don't think it's for everyone.

skyrim already has mods for AI chat with all of your followers. I haven't tried it but plenty of people use it, (in VR as well, of course).

man it really shows some of you have ai psychosis, nobody in the gaming world wants that, that's some wizard level virginity low even for gamers