NVIDIA, like everyone else on a bleeding edge node, has hardware defects. The chance goes up massively with large chips like modern GPUs. So you try to produce B200 cores but some compute units are faulty. You fuse them off and now the chip is a GP102 gaming GPU.
The gaming market allows NVIDIA to still sell partially defective chips. There’s no reason to stop doing that. It would only reduce revenue without reducing costs.
Nvidia doesn't share dies between their high-end datacenter products like B200 and consumer products. The high-end consumer dies have many more SMs than a corresponding datacenter die. Each has functionality that the other does not within an SM/TPC, nevermind the very different fabric and memory subsystem (with much higher bandwidth/SM on the datacenter parts). They run at very different clock frequencies. It just wouldn't make sense to share the dies under these constraints, especially when GPUs already present a fairly obvious yield recovery strategy.
You can't turn a GB200 into a GB202 (which I assume is what you meant since GP102 is from 2016), they are completely different designs. That kind of salvage happens between variants of the same design, for example the RTX Pro 6000 and RTX 5090 both use GB202 in different configurations, and chips which don't make the cut for the former get used for the latter.
Well the good thing for NVIDIA AI business is that most of your chips can sit unused in warehouses and still get rich. 6 million H100s sold but infrastructure (water cooled dc) for only a third of them exists in the world.
You probably mean Schwarz Gruppe, the owner of Lidl, and their subsidiary StackIT. Yes, they are growing. Schwarz is also building 11B€ AI data center in Lubbenau, so I fully agree with you. We will be fine without American digital services.
A "functioning market" doesn't prevent oligopolies. Oligopolies are natural and optimal (desirable) in many industries, if not most. That's where regulations come in.
Yay, jackpot! We taunted the monkey in the glass box into throwing the first stone.
The EU is just itching for any opportunity to get rid of US tech firms because they’re increasingly seen as sovereignty risks. And while the GDPR fines (that this likely refers to) appear huge on absolute terms, they are still low enough that US firms voluntarily decide to violate those laws and just pay the fines.
The US sees TikTok as a risk. For the EU, it’s Microsoft Office.
I think the American government is mad at the DMA more than anything. Breaking up the monopolies that are currently firmly held by American tech giants goes directly against the interests of the White House, especially now that they're able to openly bribe the president.
> The EU is just itching for any opportunity to get rid of US tech firms because they’re increasingly seen as sovereignty risks. And while the GDPR fines (that this likely refers to) appear huge on absolute terms, they are still low enough that US firms voluntarily decide to violate those laws and just pay the fines.
That is not even remotely close to the truth. The EU is not itching to get rid of Microsoft nor Windows nor Google. If these companies left tomorrow, the EU will have enormous problems replacing them if that is even possible in the first place.
The EU countries should have had a homegrown version of each US service up and running and on par with their US counterparts a decade ago, then the EU would have had leverage but as it stands, they have none.
Unless you think that every governmental office will switch to Ubuntu tomorrow morning, in which case I have a bridge to sell you.
Not to mention that the entire EU's messaging needs are met via US companies. Let's see how long the EU can last without WhatsApp, IMessage and Facebook Messenger.
My guess is not long unless you want to use Telegram which was most likely backdoor-ed by the French government not long ago.
This is the problem with the EU as it stands, there is really no mea-culpa from the institutions for their inaction and getting caught with their pants down.
All of this was foreseeable and could have been avoided, yet here we are.
There wouldn't be as big problem replacing software. It wouldn't take much. But replacing hardware? That is the real problem where you need US (or maybe China, so pick your poison).
“it’s a form of regulatory capture hidden behind a silly moral panic” is a very nice summary of gambling operators pushing an under-16 ban to protect their gambling ads by shifting responsibility onto the kids not circumventing the silly ban.
Huang's family must have had interesting connections if you consider that the daughter of his cousin is running AMD. And both were born in Taiwan and then went to MIT, which seems unlikely to happen without family money.
If you train a system to memorize A-B pairs and then you normally use it to find B when given A, then it's not surprising that finding A when given B also works, because you trained it in an almost symmetrical fashion on A-B pairs, which are, obviously, also B-A pairs.
"This seems to be largely a copy of the work done in OxCaml by @mshinwell and @spiessimon"
"The webpage credits another author: Native binary debugging for OCaml (written by Claude!)
@joelreymont, could you please explain where you obtained the code in this PR?"
That pretty much sums up the experience of coding with LLMs. They are really damn awesome at regurgitating someone else's source code. And they have memorized all of GitHub. But just like how you can get sued for using Mickey Mouse in your advertisements (yes, even if AI drew it), you can get sued for stealing someone else's source code (yes, even if AI wrote it).
Not quite. Mickey Mouse involves trademark protection (and copyright), where unauthorized commercial use of a protected mark can lead to liability regardless of who created the derivative work. Source code copyright infringement requires the copied code to be substantially similar AND protected by copyright. Not all code is copyrightable: ideas, algorithms, and functional elements often aren't protected.
The EU is worried about Trump being unpredictable, so they are pushing hard for sovereignty. See their initiatives to leave US clouds. This decision is completely in line with that strategy and, probably, also what the US military expected to happen.
The pressure to leave US controlled cloud providers actually started way back with the US Cloud Act. I’ve been surprised that that process has been as slow as it has been, especially for the public sector and adjacent services.
> I’ve been surprised that that process has been as slow as it has been, especially for the public sector and adjacent services.
Europe can't seem to get a tech sector bootstrapped no matter what it does and European governments seem to be much more comfortable with the USA having full access to everything they do then risk running on a EU cloud.
LLMs and Diffusion solve a completely different problem than world models.
If you want to predict future text, you use an LLM. If you want to predict future frames in a video, you go with Diffusion. But what both of them lack is object permanence. If a car isn't visible in the input frame, it won't be visible in the output. But in the real world, there are A LOT of things that are invisible (image) or not mentioned but only implied (text) that still strongly affect the future. Every kid knows that when you roll a marble behind your hand, it'll come out on the other side. But LLMs and Diffusion models routinely fail to predict that, as for them the object disappears when it stops being visible.
Based on what I heard from others, world models are considered the missing ingredient for useful robots and self-driving cars. If that's halfway accurate, it would make sense to pour A LOT of money into world models, because they will unlock high-value products.
Sure, if you only consider the model they have no object permanence. However you can just put your model in a loop, and feed the previous frame into the next frame. This is what LLM agent engineers do with their context histories, and it's probably also what the diffusion engineers do with their video models.
Messing with the logic in the loop and combining models has an enormous potential, but it's more engineering than researching, and it's just not the sort of work that LeCun is interested in. I think the conflict lies there, that Facebook is an engineering company, and a possible future of AI lies in AI engineering rather than AI research.
This is something that was true last year, but hanging on by a thread this year. Genie shows this off really well, but it's also in the video models as well.[1]
I think World models is way to go for Super Intelligence. One of teh patent i saw already going in this direction for Autonomous mobility is https://patents.google.com/patent/EP4379577A1 where synthetic data generation (visualization) is missing step in terms of our human intelligence.
This is the first time I have heard of world models. Based on my brief reading it does look like this is the idea model for autonomous driving. I wonder if the self driving companies are already using this architecture or something close to it.
I thoroughly disagree, I believe world models will be critical in some aspect for text generation too. A predictive world model you can help to validate your token prediction. Take a look at the Code World Model for example.
NVIDIA, like everyone else on a bleeding edge node, has hardware defects. The chance goes up massively with large chips like modern GPUs. So you try to produce B200 cores but some compute units are faulty. You fuse them off and now the chip is a GP102 gaming GPU.
The gaming market allows NVIDIA to still sell partially defective chips. There’s no reason to stop doing that. It would only reduce revenue without reducing costs.
reply