AMD Acquires Fei-Fei Li's World Labs For $8.2 Billion (cnbc.com) 15
AMD has agreed to acquire Fei-Fei Li's World Labs for about $8.2 billion in stock, making it the chipmaker's second-largest acquisition and giving it a major foothold in "world models" that simulate 3D environments for robotics and other physically grounded AI. CNBC reports: AI researchers expect world models will be able to help develop robots and other physically-grounded artificial intelligence applications. AMD plans to keep World Labs separate from its chipmaking business until the transaction closes later this year. The lab's research on next-generation models will allow AMD to plan for what its AI chips need to be capable of years in advance.
World Labs was founded by AI researcher Fei-Fei Li, who is widely seen as an AI pioneer as a Stanford professor, and who previously worked for Google leading AI research. Li will become an AMD's chief scientist and an executive vice president at the company.
In a demo presented by Li and AMD CEO Lisa Su earlier this year, the two executives showed a World Labs model called Marble creating a three-dimensional scene out of a few images. "Intelligent agents, whether it's robots or vehicles or even tools, can learn inside very-rich physics-aware digital worlds before they even need to be deployed into the real one, making them much safer," Li said at the presentation.
World Labs was founded by AI researcher Fei-Fei Li, who is widely seen as an AI pioneer as a Stanford professor, and who previously worked for Google leading AI research. Li will become an AMD's chief scientist and an executive vice president at the company.
In a demo presented by Li and AMD CEO Lisa Su earlier this year, the two executives showed a World Labs model called Marble creating a three-dimensional scene out of a few images. "Intelligent agents, whether it's robots or vehicles or even tools, can learn inside very-rich physics-aware digital worlds before they even need to be deployed into the real one, making them much safer," Li said at the presentation.
Only makes training them safer and cheaper (Score:4, Interesting)
Training the robots' models in a simulation before training them in the real world won't make them safer. In fact if anything it will make them less safe due to their earliest training being subtly different from reality. What it does is make it cheaper and less dangerous to train them, and faster because you can run the simulation faster than reality.
Re: (Score:2)
Training the robots' models in a simulation before training them in the real world won't make them safer.
How could you know without ever evaluating the results of the training in the simulated environment? I'm not saying your argument is wrong just that it could be specious since simulations make it possible to train them for much longer periods of simulated time in a wider variety of environments.
Re: Only makes training them safer and cheaper (Score:2)
Because it's not reality and because the technology is fundamentally unsuited to the job. Industrial control should be predictable.
Re: (Score:2)
Because it's not reality
How you can be sure that will impact the final outcome in any tangible way?
because the technology is fundamentally unsuited to the job. Industrial control should be predictable.
That is the goal but that goal runs counter to the real world where unpredictable situations occur and people get injured or killed despite having machinery with deterministic behaviors. This can be the result of failing machinery or more often operator error. As such, the questions are if AI will behave in a more predictable manner than humans and what is the scope of the potential consequences of unpredictable behavior.
I'm highly s
Re: (Score:2)
I'm highly skeptical of large neural networks but think optimized neural networks can provide superior results to our own programming efforts which can be equally unpredictable due to edge cases.
I think neural nets have a place in industrial control, but it's only for sub-processes which are also guarded by watchdogs which prevent really OOB behavior.
Re: (Score:2)
That is a reasonable approach and it may be exactly how they plan on using the tech. I do understand your reaction though as startups tend to be lead by non-technical people that believe that AI is flawless or that it just needs to be fixed a bit and it will be flawless. However, Li is an actual researcher and professor, which makes them far more likely to recognize the limitations of the technology.
Perfect (Score:1)
So you're saying porn-bots will need lots of dick pics to properly handle their customers? Sounds like a perfect situation.
there isn't a deep enough hole (Score:1)
Re: (Score:2)
> Amazon mechanical turk slaves
What does 'slaves' mean to you in this context?
Frauds frauding frauds (Score:2)
This is laughably unscalable and noisy with only static models and zero useful topology or anything except that it looks halfways decent when covered with motion blur. Anything in a still falls apart after a bit because they just keep ignoring physics and insisting on having it both slow and imprecise, when you would rather have a tradeoff between speed and accuracy... I guess glossing it with photorealism helped, along with probably some insider deals.
Re: Frauds frauding frauds (Score:2)
I donâ(TM)t think AMD purchased them for the current product, but rather to bootstrap R&D. Itâ(TM)s difficult to bring together experts and say, hey make something using our stuff and tell us what works/doesnâ(TM)t, vs finding an operation that is already
Doing that, has the infrastructure and operations in place to continue doing it (or start new experiments etcâ¦) to then inform product.
Creating a 3D scene out of a few images. (Score:2)
Is there a link to this demo as I don't have time to watch a two hour video to get to the best bits?
Who's in the bidding for us, Universelabslabs (Score:2)
We now make some software stuff that's vaguely related to some field that investment geniuses will have read about in the WSJ last week. AR I think. Or maybe drones.
Anyway, bidding starts at $8.3 billion ( Proinvestortip: "Universe" is bigger than "World" and two labs are better than one, see. )
It is not AI but an expert system bounded by rules (Score:2)