Google Is Forming a New Team To Build AI That Can Simulate the Physical World 37
Google DeepMind is forming a new team to develop AI models capable of simulating the physical world. It's being led by former OpenAI Sora co-lead Tim Brooks and builds on Google's Gemini, Veo, and Genie projects. "DeepMind has ambitious plans to make massive generative models that simulate the world," Brooks wrote on X. "I'm hiring for a new team with this mission." TechCrunch reports: According to job listings Brooks linked to in his post, the new modeling team will collaborate with and build on work from Google's Gemini, Veo, and Genie teams to tackle "critical new problems" and scale models "to the highest levels of compute." Gemini is Google's flagship series of AI models for tasks like analyzing images and generating text, while Veo is Google's own video generation model. As for Genie, it's Google's take on a world model -- AI that can simulate games and 3D environments in real time. Google's latest Genie model, previewed in December, can generate a massive variety of playable 3D worlds.
"We believe scaling [AI training] on video and multimodal data is on the critical path to artificial general intelligence," reads one of the job descriptions. Artificial general intelligence, or AGI, generally refers to AI that can accomplish any task a human can. "World models will power numerous domains, such as visual reasoning and simulation, planning for embodied agents, and real-time interactive entertainment." Per the description, Brooks' new team will look to develop "real-time interactive generation" tools on top of the models they build, and study how to integrate their models with existing multimodal models such as Gemini.
"We believe scaling [AI training] on video and multimodal data is on the critical path to artificial general intelligence," reads one of the job descriptions. Artificial general intelligence, or AGI, generally refers to AI that can accomplish any task a human can. "World models will power numerous domains, such as visual reasoning and simulation, planning for embodied agents, and real-time interactive entertainment." Per the description, Brooks' new team will look to develop "real-time interactive generation" tools on top of the models they build, and study how to integrate their models with existing multimodal models such as Gemini.
Obvious... (Score:2)
Re: (Score:1)
Re: Obvious... (Score:2)
Re: (Score:2)
Freedom of speech is not freedom to lie or deceive.
Freedom of speech is freedom to lie. Politically, yesterday's lies often turn out to be the truth.
Re: Obvious... (Score:2)
cant wait (Score:2)
I'm prepared for great solutions like walking on water and walking on the ceiling. I can't even hallucinate as well as AI.
Re: cant wait (Score:2)
Try LSD
Ready Player One (Score:2)
That's how it began.
Sounds like a lost cause (Score:2)
One thing about the physical world is pretty striking: It must be simulated with perfect accuracy in order to get good results. Even tiny deviations can mess up things completely.
This idiocy has the stink of desperation.
Re: (Score:3)
Not necessarily. We are able to simulate the physical world with less than perfect accuracy and get useful results, like throwing rocks on target. Other animals can simulate the world too, birds for instance are pretty good with aerodynamics, that's how they are able to fly.
There are arguments about how this is the key to human-like intelligence. Our big brains evolved as physics simulators, so that we can do things like throw rocks on target. Incidentally, what we have not evolve to do is deal with subatom
Re: (Score:3)
Quite agree. But an important thing to realize is that the brain is NOT doing what we understand as "simulation", which is formulating an exact model of the physical phenomena and unrolling a chain of symbols to predict how they evolve.
If these AIs use statistical learning models, we should rather call them "intuition" rather than "simulation" to better understand how they with. Intuition creates a statistical model based on the observed patterns of the physical phenomena, and protecting those same patterns
Re: (Score:2)
/"how they with"/"how they work"/
Re: (Score:2)
For pure motor functions, this may work and may actually be what the brain does. But a lot of human intuition, at least in smart humans, uses General Intelligence as well both for predictions and for plausibility checking. Statistical models are useless for that.
Re: (Score:2)
Yeah but my key point is that a simulation is a different kind of simplification than that of intuition. A simulation is defined by a fundamental model that captures exactly the behaviour of the simulated thing; while the intuitive model only captures secondary properties that are derived from the behaviour, but that are only a proxy for the real phenomenon.
Re: (Score:2)
However- I think you perhaps overstepped with "a fundamental model that captures exactly the behavior of the simulated thing".
A simulation imitates a thing. The quality of the imitation is variable, as well as the domain in which it's imitated.
Where I'd say you're spot-on, is that the concept of "simulation" is trying to imitate the fundamentals of a thing, while "intuition" is really just doing... whatever the fuck it's doing, based on external
Re: (Score:2)
Re: (Score:2)
The fascinating part is this "intuition" sometimes gives better results than the simulation. This can be seen with AI models that are better at predicting some chaotic sequences than traditional models. https://www.quantamagazine.org... [quantamagazine.org]
This is currently being exploited by Google for weather prediction. Presumably the AI is able to capture some pattern in the sequen
Re: (Score:2)
You're throwing an truly absurd amount of observable data at a model who's job is to come up with any kind of internal state machine that works to basically reproduce them.
When you're hand-modeling such an internal state machine, you are subject to all of your biases. The real answer might be staring you in the face, and you will never see it, if you're intellectually opposed to it.
Training data being pumped into NNs have no such biases, and indeed, the internal mode
Re: (Score:2)
Sure. But that requires access to the physical environment to run experiments, and you need to run _physical_ experiments for each single use-case. This is not what the story is about,
Define "tiny" (Score:2)
Even tiny deviations can mess up things completely.
Define "tiny". Current games approximate normal forces which can lead to strange behaviour where objects climb slopes etc. but I doubt anyone is going to notice the fact that a human-scale simulation does not properly correct for relativistic and quantum effects and only does newtonian mechanics.
Basically an accurate simulation of newtonian mechanics, or at least accurate enough to match most people's physical intuition, should definitely be doable. Going beyond that is where it will get questionable e.
Re: (Score:2)
I doubt that you can accurately simulate the atomic structure of matter so that you can generate static charges
Not without a computer as large as the environment you're trying to simulate.
No matter how you swing it, such a simulation needs to have some kind of lossy compression of the environmental data.
30 billion year old repeat (Score:5, Funny)
Wow! This exact thing happened about 30 billion years ago when Yark proposed a universe simulator based on Plank's Constant and e. How recursive!
Re: (Score:1)
Wow! This exact thing happened about 30 billion years ago when Yark proposed a universe simulator based on Plank's Constant and e. How recursive!
Nah, it's using a much more recent technique. It goes back to 530 million years ago, when Pnae the Planaria realized she could combine input from multiple photoreceptor cells to build a mental model of the direction a predator was approaching, and flee in the opposite direction. It has been all mind games since then.
Re: (Score:1)
also, https://en.wikipedia.org/wiki/... [wikipedia.org]
Deep Mind is doing it right (Score:2)
They are building truly useful stuff while the other companies continue churning out crap generators
maybe first become experts in fake worlds? (Score:2)
so, what is this? (Score:2)
This is not possible. (Score:3)
To properly simulate something it requires an equal or greater amount of resources used by the thing you are simulating. Anything less requires leaving out some factors. If you want to perfectly simulate the Universe you need as much energy as the Universe contains in the first place. If you want to simulate our environment on Earth you need as much or more energy as the Earth uses up. More because there will be wastage in your simulation.
So any simulation run in this way can only be massively imperfect. If they want it to be perfect then they will need to run it in space and use a huge amount of energy, more than is currently available to the whole Human Race. Using 'AI' doesn't change any of this. There was a BBC Sci-Fi series about this a few years ago. This is as doomed to failure as the attempt in that series was.
Re: (Score:2)
So any simulation run in this way can only be massively imperfect.
This is undeniably true, but let's not pretend the bar is really that high.
An even rudimentary physical simulation would be more than adequate for your senses.
Metroverse and sex bots.... (Score:2)
Simulate the physical world? Why? (Score:2)
The physical world kinda sucks. Now, if they can simulate the physical world but with laser beam equipped dragons fighting spaceships on an intergalactic scale, then we'd be getting somewhere. Fuck trying to replicate the disaster we've made of the physical world. Give me something to enjoy while we burn the world!
I've seen this one before (Score:2)
This Secondlife/Metaverse crap - same. In fact, pretty sure Zuck is still delusional about anyone wanting this.
You are the training product (Score:2)