Artificial Intelligence thread

horse

horse

Brigadier
Registered Member
Kalum Pupeter said:
Thoughts about the writing of this intellectual Hanjian?
Click to expand...

I am drunk, so that Xu guy sounds like a bit of an idiot.

There was a recent article about how almost all of the top download AI models were Chinese. I may be drunk, but I still remember that.

Guess that according to that Xu guy the export controls are working, (which means he what he really wants to fricking say is the USA is ahead of China when it comes to AI), he is right! When the goalposts are moved again to the best model available that no one uses because it is too good it is dangerous just trust me bro.

Cannot beat those Americans!

:cool:
 
G

gpt

Junior Member
Registered Member
horse said:
Guess that according to that Xu guy the export controls are working, (which means he what he really wants to fricking say is the USA is ahead of China when it comes to AI), he is right! When the goalposts are moved again to the best model available that no one uses because it is too good it is dangerous just trust me bro.
Click to expand...

We've done this topic to death already in this thread, but to be clear: for most devs, researchers, real-world applications, you want to optimise for cost and distribution ie. models that people can actually run and translate into practical utility. This doesn't mean frontier models are useless. Frontier-class performance (if you take all those claims around Mythos/Glasswing at face value-autonomous zero-day hunting, running thousands of huge-model instances in parallel) is heavily gated by raw high-end compute, especially HBM.

DC blob's goalpost haven't moved at all. They're just focused on the upstream to keep the mythical 'true high end' out of reach. Whether or not you think it's wise to burn that much hardware on it, that's the moat they're defending.
 
C

CMP

Captain
Registered Member
Maikeru said:
So what does this guy know about AI chips?

Please, Log in or Register to view URLs content!
Click to expand...
I'm surprised no one tried to call the podcaster out on his claim that 3nm chip is 10x or more than compute of 7nm chip.


Everything Jensen is saying about the AI competition makes perfect sense to me. Dwarkesh sounds like he doesn't understand anything.
 
Last edited:
H

HighGround

Senior Member
Registered Member
CMP said:
I'm surprised no one tried to call the podcaster out on his claim that 3nm chip is 10x or more than compute of 7nm chip.
Click to expand...

It's not important to the point they're making.

CMP said:

Everything Jensen is saying about the AI competition makes perfect sense to me. Dwarkesh sounds like he doesn't understand anything.
Click to expand...

He's just a podcaster. He doesn't need to. The content is about a relatively normal person with very layperson understanding of AI asking the CEO questions.
 
H

HighGround

Senior Member
Registered Member
CMP said:
If someone can't make a good point without exaggerating actual metrics by 2x to 5x, the point is inherently not a good one.
Click to expand...
This is not at all true. For example,

"You'll never outrun a car, it's a million times faster than you."

Now is it important whether a car is only ten times faster to the overarching point? No, it's not important.

But anyway, there is no need to attack the video, it's biggest usefulness is in confirming Nvidia's position. They want access to the Chinese market, clearly. They're more than happy to compete and they believe it's in US interest to do so. Jensen also doesn't think China is going to run out of compute either.
 
You must log in or register to reply here.
Top