These things are all bets on the future -- and with the World Labs and Taalas acquisitions, in particular, the bet they're making is that conventional LLM development will plateau or become utterly commodified.
It's also a hedged bet against Nvidia's interest in fullstack robotics. AMD missed the datacenter train, but it's not too late to start building out an edge robotics hardware stack, maybe derived from CDNA.
I haven't heard about them, I was going to comment about how these reaserchers will get bitter lesson'd, but after looking at what they do, their "world models" are actually transformer models.
This means that the conversation has shifted to the point where we can assume that transforms hold emergent world models (already proven for tiny toy worlds). This is good, because it's probably true.
In any case, this looks awesome and I wish them all the best!
Dr Fei Fei Li's book The Worlds I See is a great read about both her life and the history of early AI and her role in it, which is major (Ilya was one of her phd students, along with the Alex of AlexNet) so if you don't already know the history of these systems, it's a great read. Very inspiring stuff. Great read and learned about it here so passing on the good word
Assuming this means AMD wants to jumpstart competition with NV on a model/simulation ecosystem (which NV has been doing for ~a decade). Probably has at least something to do with NV acquiring HF.
I'm not sure how much things like Omniverse are used, but NV has put out some really interesting stuff in the space (like for example ARTY for a somewhat recent example).
Overall it seems like a financial vote of confidence for "world models" (whatever that means, I still don't know). But it could also mean the death of the interesting things WL was doing with Marble + Atlas. It's hard to read PR.
Doing this on a throwaway to ask some awkward questions.
Way to go and all, and I suppose the investors got an exit but I still scratch my head on this one. We've seen the rise of World Labs from the beginning to this exit, and still the raw output is barely usable for any conceivable use case.
The raw output from these models is very similar or identical to what you can get to generating a splat from a camera that rotates using minimax or other frontier video models.
I work with clients in real world design contexts and the output from world labs was never usable due to all sorts of distortions and errors and stuff.
So again it's cool, but like what was all this hype about? This was usable in a very niche case if you're making video games with no care for spatial continuity and are okay with super heavy assets.
I am surely missing something as I just black box test these things for applicability in the real world, and I just never found any.
Tangent comment, it's so sad you had to create a throwaway here just to express this opinion.
People shouldn't have to be afraid to post what they think of tech topics here, as long as it is not an illegal expression of speech. i.e. one shouldn't have to use a throwaway to ask sensible and valid questions.
Yeah sorry guilty as charged -- it's hard because I've met these people IRL and yes I don't want to damage those relationships.
They're super nice, smart and capable folks. But, also as a practitioner in this space the gulf I see between stated capabilities and actual output is super big and I just don't get it.
I'm also worried I'm missing something obvious, like it's me that's the problem and that's embarrassing.
If people are really that excited about my identity go for it. I was just trying to be respectful and diminish my criticisms by making them pseudo anonymous.
They rode the gaussian splat hype train and their demos look sick. I suspect they had a big marketing push the last 6 weeks because they were looking for an exit.
Apple and Meta are the only two big companies really pushing the frontier of this space, and everyone else is trying to figure out what to do with the technology. World Labs had the best marketing demos and showed the world what you could do with it, but it's all been vaporware from where I've been sitting.
Honestly, I think this 'billion' number is also just the ball park for companies you need to pay if you want to buy them.
OpenRouter was sold for 7.5 Billion.
Lets just accept, that these people don't evaluate a company by value but by magic market "what is a sum someone needs to pay for buying a company and what company is somehow worth it to actually buy for this type of money but more in a sense of adding this to your portfolio"
AMD grew their revenue 50% last quarter from the year prior, for 11.5B in revenue. If this team can help continue that or grow that then yes - the amount of money involved here is insane.
I don't know about the training, but in terms of the the 'a lot of data', probably one of the inputs would be something like the geospatial data that Niantic has collected from all of the world's Ingress and Pokemon Go players, correlated with other more basic mapping data and things like google street view type camera data.
Neoclouds want to do neolab things and now chipmakers want to do neolab things
We don't know what the right integrations are. Is it the chipmakers with the toolmakers? The front-end makers with the model designers? That's the current set-up. But there might be another industrial configuration that is superior.
AMD is still struggling to establish itself in AI. That makes it a worthy competitor. Its agglomeration of talent and resources is probably, knowing little else, good.
If the sigmoid approacheth, the first one to get a good model into silicon, at acceptable benchmarks and silicon rejection %, will make some truckloads. Most people don't need SOTA for most of their problems.
the TAM might be a real number, but the addressee might already be walking around, such as any qwen3.8 model. Qwen3.8-Flash-Next is churning out some fascinating capabilities as we speak.
There's only so many math problems you can solve without ROI.
I used to think that, but I've started to come around to the fact that better AI comes with progressive unlocks that create new use cases. I suspect in 18 months, one of the primary consumers of tokens will be people using LLMs for one-shotting increasingly complicated games rather than just white collar work.
Then there's the matter of engineering and synthesizing consumer objects completely personalized for users!
I think it's true in the general case that people who want to create stuff (eg games) will create stuff and people who don't won't.
its hard bc everyone here is some form of builder. but there are people out there who just have almost no desire to create.
and even if we
lower the bar for creating to "hey llm make me a new video game to play tonight" that will still come with all the baggage of creation and turn folks off.
The ecosystem around LLMs is becoming just as or more important than the model. Having fast access to a web crawl and other reference material is important.
Other embodied labs are less gamer-y/consumer-y. WorldLabs didn't know whether to be a consumer tool, a creative tooling company, a robotics research lab.
Investors who have seen their materials told me this was their direct feeling.
AMD has acquired "AI stuff" for several billions at this point, yet somehow, doing "AI stuff" on their GPU/NPU stack still seems to be troublesome. Well, that is what I read from others anyway. Inference with llama.cpp on AMD GPUs pretty much just works.
I bet soon enough we can just ask AI to port stuff from CUDA to whatever language AMD is using. This is how nVidia will become a victim of their own success.
Keep going... extrapolate out further. There's a conclusion you beley will be true, but haven't fully thought through why that conclusion must be true if CUDA is no longer the moat it once was.
Who says that? I’m having a blast with rocm powering my R9700. Been able to run all models on day 0 at comparable speeds to Nvidia with similar memory bandwidth. Software is no longer the main bottleneck for AMD.
Really, it's just the Python frameworks on AMD consumer GPUs that suck. AMD datacenter GPUs seem to work with them well enough, and llama.cpp crushes it for the rest of us.
The repeating story - There are X number of people in the tech and finance world that form a close network. It is the incestuous coming together of those that enable all of them to become a multi billionaire through acquisitions and IPOs. New people join this list on a regular basis by building something that create FOMO amongst the already billionaires and they indoctrinate those into the group using retail investor money.
I hesitate to recast the discussion in a negative tone, but this doesn't sound right on technical terms.
Is World Labs' Atlas genuinely novel? Their demos do not seem to be better than the existing state of the art.
Fei Fei Li has been criticized even in the ImageNet days as a shower than a doer. Her roadshow past two years extolling "world model" in vague terms didn't show much insight, and her company's demo seems to merely rehash what has been available in the field for years. If you haven't seen the current SoTA in Gaussian splats, then World Labs' demos may seem cool, but they are mostly standard in the field now. If someone is more familiar with the inner workings of World Labs, I am happy to be corrected (but inner details wouldn't change the fact that demos are no better than SoTA).
I really had high hopes for World Labs, but I am quite saddened to realize that perhaps all this was just a financial maneuver and the detractors were right from the beginning.
source: I used to be in the downstream field: Gaussian spats for robotics. The exact field that World Labs' is supposed to help.
Bloomberg: https://www.bloomberg.com/news/articles/2026-09-28/amd-to-bu...
Worldlabs PR: https://www.worldlabs.ai/blog/amd-announcement
CNBC: https://www.cnbc.com/2026/09/28/amd-fei-fei-li-world-labs.ht...
Tweet: https://twitter.com/EdLudlow/status/2104664216393420843
Dupe: https://news.ycombinator.com/item?id=49883744
What is AMD getting out of this?
Talent and ability to create next gen of multimodal models.
Mostly just 3d world models though, they're not the same as multimodal models.
They're building quite the stack in experimental and ambitious AI model development.
For e.g., they recently acquired Taalas. https://news.ycombinator.com/item?id=49201970
These things are all bets on the future -- and with the World Labs and Taalas acquisitions, in particular, the bet they're making is that conventional LLM development will plateau or become utterly commodified.
It's also a hedged bet against Nvidia's interest in fullstack robotics. AMD missed the datacenter train, but it's not too late to start building out an edge robotics hardware stack, maybe derived from CDNA.
I haven't heard about them, I was going to comment about how these reaserchers will get bitter lesson'd, but after looking at what they do, their "world models" are actually transformer models.
This means that the conversation has shifted to the point where we can assume that transforms hold emergent world models (already proven for tiny toy worlds). This is good, because it's probably true.
In any case, this looks awesome and I wish them all the best!
I wonder where this will lead
I'd bet world models will likely become the new "media" replacing tv, movies, music, video games etc.
Unlikely they will outcompete human artists and engineers using AIs to create those same types of media: tv, movies, music, games.
Yes, because consumers just LOVE to pay for anything associated with AI!
It'll be very interesting to see how the dust settles after this window of AI IPOs and acquisitions closes.
Dr Fei Fei Li's book The Worlds I See is a great read about both her life and the history of early AI and her role in it, which is major (Ilya was one of her phd students, along with the Alex of AlexNet) so if you don't already know the history of these systems, it's a great read. Very inspiring stuff. Great read and learned about it here so passing on the good word
It's a good read but I believe you're mistaken. Ilya and Alex Krizhevsky worked with Hinton.
You might be thinking of Andrej Karpathy as one of Li's more famous PhD students?
Assuming this means AMD wants to jumpstart competition with NV on a model/simulation ecosystem (which NV has been doing for ~a decade). Probably has at least something to do with NV acquiring HF.
I'm not sure how much things like Omniverse are used, but NV has put out some really interesting stuff in the space (like for example ARTY for a somewhat recent example).
Overall it seems like a financial vote of confidence for "world models" (whatever that means, I still don't know). But it could also mean the death of the interesting things WL was doing with Marble + Atlas. It's hard to read PR.
it was also likely difficult to do what WL was looking to do without significant amount of compute.
Think it's a good match
WL was probably mostly using NVIDIA compute though?
If the ROCm team can't get enough compute it doesn't bode well for World Labs.
Doing this on a throwaway to ask some awkward questions.
Way to go and all, and I suppose the investors got an exit but I still scratch my head on this one. We've seen the rise of World Labs from the beginning to this exit, and still the raw output is barely usable for any conceivable use case.
The raw output from these models is very similar or identical to what you can get to generating a splat from a camera that rotates using minimax or other frontier video models.
I work with clients in real world design contexts and the output from world labs was never usable due to all sorts of distortions and errors and stuff.
So again it's cool, but like what was all this hype about? This was usable in a very niche case if you're making video games with no care for spatial continuity and are okay with super heavy assets.
I am surely missing something as I just black box test these things for applicability in the real world, and I just never found any.
Tangent comment, it's so sad you had to create a throwaway here just to express this opinion.
People shouldn't have to be afraid to post what they think of tech topics here, as long as it is not an illegal expression of speech. i.e. one shouldn't have to use a throwaway to ask sensible and valid questions.
Yeah sorry guilty as charged -- it's hard because I've met these people IRL and yes I don't want to damage those relationships.
They're super nice, smart and capable folks. But, also as a practitioner in this space the gulf I see between stated capabilities and actual output is super big and I just don't get it.
I'm also worried I'm missing something obvious, like it's me that's the problem and that's embarrassing.
So at least I'll share that anonymously. :/
note that unless you've de-anonymized your text - your own idosyncracies might be enough with other clues you've left to figure out who you are IRL
Came here to say exactly this. The way anon writes could be a hint if there are a couple of emails laying around from him/her.
If people are really that excited about my identity go for it. I was just trying to be respectful and diminish my criticisms by making them pseudo anonymous.
You didn’t do anything wrong. I was just worried FOR YOU. Sorry if I came across like I was attacking you or blaming you.
i also would like to know an answer to this
They rode the gaussian splat hype train and their demos look sick. I suspect they had a big marketing push the last 6 weeks because they were looking for an exit.
Apple and Meta are the only two big companies really pushing the frontier of this space, and everyone else is trying to figure out what to do with the technology. World Labs had the best marketing demos and showed the world what you could do with it, but it's all been vaporware from where I've been sitting.
Is a 2-year-old company worth 8 billion USD?
A name like Fei-Fei Li is probably worth a chunk of that (I'm not saying I agree, but that's probably the thinking).
It's a public bet, not a valuation based on product economics. Billion dollar bets follow different rules (and I'm not saying it's sane).
You're right. Her bio writes like an OP self-insert character's [1].
[1] https://en.wikipedia.org/wiki/Fei-Fei_Li
The past couple of years have really driven home the fact that a company is worth what someone will pay for it
just like NFTs :)
And that "Attention is all you need" applies to financial markets as well.
... just like real estate, a willing buyer -- and a willing seller
Honestly, I think this 'billion' number is also just the ball park for companies you need to pay if you want to buy them.
OpenRouter was sold for 7.5 Billion.
Lets just accept, that these people don't evaluate a company by value but by magic market "what is a sum someone needs to pay for buying a company and what company is somehow worth it to actually buy for this type of money but more in a sense of adding this to your portfolio"
AMD grew their revenue 50% last quarter from the year prior, for 11.5B in revenue. If this team can help continue that or grow that then yes - the amount of money involved here is insane.
Could be if they are pre-revenue.
They're a potential pure play
They’re bound to announce an appliance soon.
Given a lot of proprietary data, how would one go about starting training a world model?
Step 1: raise a billion dollars. Step 2: send a billion dollars to someone with GPUs.
Funnily enough now you can get the billion dollars from NVIDIA themselves.
But only if you promise to spend the billion dollars on something NVIDIA is selling
or https://colo.to/nvidia-stock-narrative.html
Step 3) Underpants ???
I don't know about the training, but in terms of the the 'a lot of data', probably one of the inputs would be something like the geospatial data that Niantic has collected from all of the world's Ingress and Pokemon Go players, correlated with other more basic mapping data and things like google street view type camera data.
https://www.google.com/search?client=firefox-b-d&q=pokemon+g...
You don't have to, there is a free one with open weights from lingbot. https://technology.robbyant.com/lingbot-world
Neolabs keep moving down the stack.
Neoclouds want to do neolab things and now chipmakers want to do neolab things
We don't know what the right integrations are. Is it the chipmakers with the toolmakers? The front-end makers with the model designers? That's the current set-up. But there might be another industrial configuration that is superior.
AMD is still struggling to establish itself in AI. That makes it a worthy competitor. Its agglomeration of talent and resources is probably, knowing little else, good.
World Labs is really cool. The 3D reconstruction from only a handful of frames is on another level.
how to quickly make a couple million dollars these days: start a startup and then get acquired by AMD in a few years.
The AMD part looks too specific, but I’ll have to check the math.
This was absurdly soon.
I was also shocked by how quickly AMD acquired Talaas. AMD may be preparing for the next way (ultra fast inference, and embodied AI inference)
If the sigmoid approacheth, the first one to get a good model into silicon, at acceptable benchmarks and silicon rejection %, will make some truckloads. Most people don't need SOTA for most of their problems.
the TAM might be a real number, but the addressee might already be walking around, such as any qwen3.8 model. Qwen3.8-Flash-Next is churning out some fascinating capabilities as we speak.
There's only so many math problems you can solve without ROI.
I used to think that, but I've started to come around to the fact that better AI comes with progressive unlocks that create new use cases. I suspect in 18 months, one of the primary consumers of tokens will be people using LLMs for one-shotting increasingly complicated games rather than just white collar work.
Then there's the matter of engineering and synthesizing consumer objects completely personalized for users!
Tge point though, is the TAM might already be saturated.
Even if you buy singularity the people who needand afford it is dwindling.
This also assumes societal cohesion isnt destroyed by sed people.
I think it's true in the general case that people who want to create stuff (eg games) will create stuff and people who don't won't.
its hard bc everyone here is some form of builder. but there are people out there who just have almost no desire to create.
and even if we lower the bar for creating to "hey llm make me a new video game to play tonight" that will still come with all the baggage of creation and turn folks off.
The ecosystem around LLMs is becoming just as or more important than the model. Having fast access to a web crawl and other reference material is important.
Other embodied labs are less gamer-y/consumer-y. WorldLabs didn't know whether to be a consumer tool, a creative tooling company, a robotics research lab.
Investors who have seen their materials told me this was their direct feeling.
AMD has acquired "AI stuff" for several billions at this point, yet somehow, doing "AI stuff" on their GPU/NPU stack still seems to be troublesome. Well, that is what I read from others anyway. Inference with llama.cpp on AMD GPUs pretty much just works.
I bet soon enough we can just ask AI to port stuff from CUDA to whatever language AMD is using. This is how nVidia will become a victim of their own success.
It's called Hyperloom.
And then what?
Then all software moats become shallow
Fortunately for Nvidia, they sell hardware and give away software.
Yes, but specifically they sell hardware at absurd margins, and lock people in via their software ecosystem.
Keep going... extrapolate out further. There's a conclusion you beley will be true, but haven't fully thought through why that conclusion must be true if CUDA is no longer the moat it once was.
You can already do this with oh my pi or oh my openagent
Who says that? I’m having a blast with rocm powering my R9700. Been able to run all models on day 0 at comparable speeds to Nvidia with similar memory bandwidth. Software is no longer the main bottleneck for AMD.
Really, it's just the Python frameworks on AMD consumer GPUs that suck. AMD datacenter GPUs seem to work with them well enough, and llama.cpp crushes it for the rest of us.
Congrats to the team on such a successful exit!
I'm curious how much the recent AI 3D modeling trend (Astra post-trained to use Blender) would have impacted this.
World Labs' entire tech stack becomes potentially obsolete when anyone can prompt GPT/Claude to create a simulation-ready 3D model from a photograph.
My guess is their tech can be used for spatial reasoning training environments, which gives a value beyond just output of a 3D scene.
~2.5 years roughly and an $8.2 billion exit. Absolutely amazing honestly.
The repeating story - There are X number of people in the tech and finance world that form a close network. It is the incestuous coming together of those that enable all of them to become a multi billionaire through acquisitions and IPOs. New people join this list on a regular basis by building something that create FOMO amongst the already billionaires and they indoctrinate those into the group using retail investor money.
This is truly impressive. Fei-Fei Li pretty much went on a 2.5 year IPO roadshow with World Labs and exited with a few cool tech demos.
I hesitate to recast the discussion in a negative tone, but this doesn't sound right on technical terms.
Is World Labs' Atlas genuinely novel? Their demos do not seem to be better than the existing state of the art.
Fei Fei Li has been criticized even in the ImageNet days as a shower than a doer. Her roadshow past two years extolling "world model" in vague terms didn't show much insight, and her company's demo seems to merely rehash what has been available in the field for years. If you haven't seen the current SoTA in Gaussian splats, then World Labs' demos may seem cool, but they are mostly standard in the field now. If someone is more familiar with the inner workings of World Labs, I am happy to be corrected (but inner details wouldn't change the fact that demos are no better than SoTA).
I really had high hopes for World Labs, but I am quite saddened to realize that perhaps all this was just a financial maneuver and the detractors were right from the beginning.
source: I used to be in the downstream field: Gaussian spats for robotics. The exact field that World Labs' is supposed to help.
Did Jensen pass?
All FF did was building a dataset.
Interesting move. AMD clearly wants a piece of that specialized lab tech. Wonder what they'll build with it.