Watching an interview from LaCun at the world goverment sumit

Question

Watching an interview from LaCun at the world goverment sumit

and he says that to be at the forefront of ai a country or company needs around 16.000 gpu's to have a supercomputer to train its models (1B+ dollar). Does hyperon also need so much compute power before you can train very powerfull models? @SNET_whale (maybe a stupid question and already explained in blogs but as its quite complicated for most i ask it here)

#cryptocurrency #english #singularitynet

0

19.02.2024

1 ответов

9 просмотров

Robert · Accepted Answer

If forefront is defined as training the largest multimodal transformer on the most extensive dataset then you (currently) need those unimaginable amounts of computation - and nobody can realistically compete with the megacorporations in that narrow arena, especially if they start building their own chip supply chains. However, innovations and breakthroughs so far came both from the labs of the megacorporations and from small groups in academia. I'd say it's conceivable that throwing enough data and compute at known architectures and mild variations thereof might be enough to reach AGI, but it's also conceivable to reach AGI on alternative paths enabled by algorithmic innovations that require less amounts of data and compute. So it's a race between those that throw insane amounts of resources into what's known to work well on narrow tasks, and those who keep researching more innovative ways to achieve broad intelligence in machines. The real problem, however, is imo not achieving AGI but ensuring it's safe for humans and other life forms on earth. Also, don't take LeCun's words too seriously, he's stated quite a few predictions that turned out to be completely wrong on camera, and he's heavily influenced in what he says by Meta's interests.

209 похожих чатов

Watching an interview from LaCun at the world goverment sumit

1 ответов

Похожие вопросы