but there's simply nowhere left, i'm afraid

grim

(http://www.autoadmit.com/thread.php?thread_id=5859508&forum_id=2),#49835293)

it doesn't work like that. this is actually a Clique thing, it just doesn't map directly to the specific cliques in Clique Theory

(http://www.autoadmit.com/thread.php?thread_id=5859508&forum_id=2),#49835325)

(http://www.autoadmit.com/thread.php?thread_id=5859508&forum_id=2),#49835333)

not a real thing, not a real option

the US is unironically the best option left which really shows you how bad the situation is

(http://www.autoadmit.com/thread.php?thread_id=5859508&forum_id=2),#49835341)

what countries have you checked out? i still havent even loked at SE asia and japan wis always an option

(http://www.autoadmit.com/thread.php?thread_id=5859508&forum_id=2),#49835350)

no one in the US will ever starve or die from lack of medicine/supplies

many others will

(http://www.autoadmit.com/thread.php?thread_id=5859508&forum_id=2),#49835360)

there might be some floyd-style protests periodically but nothing more. we won't even reach 60s race riot level of unrest, much less anything resembling an actual revolt

(http://www.autoadmit.com/thread.php?thread_id=5859508&forum_id=2),#49835801)

And maybe I'm naive, but I really struggle to understand what the actual need is for so many data centers. (Outside of it helps perpetuate the bubble.) What do they actually do? What are they actually for?

(http://www.autoadmit.com/thread.php?thread_id=5859508&forum_id=2),#49835965)

theyve been using free energy systems already for everything, this is all a mass cartoon heist

(http://www.autoadmit.com/thread.php?thread_id=5859508&forum_id=2),#49836017)

https://epoch.ai/blog/can-ai-scaling-continue-through-2030

10,000x GPT-4 training runs are likely possible by 2030. we will likely see reasonably large improvements in deep learning compute performance per watt for at least another decade. meanwhile the efficiency of the training algorithms continues to improve even faster than the hardware.

the more interesting argument is that inference compute will be a limiter to widespread deployment. agentic models for SWE are token hogs, so maybe we will have powerful models but only so many people can use them. i doubt this one too because it seems like distillation and overtraining models allows model developers to make increasingly cheap but powerful small models. that might not be enough, but it won't be enough to prevent companies from building very intelligent models for their own internal uses.

there is also increasingly large economic incentives to find ways around power constraints in the medium term. just as one example, most of the time there is excess capacity on the grid. there's only a constraint at peak demand. the total excess capacity on the grid is quite large and conceivably datacenters could throttle their usage down on peak demand hours.

the people thinking this will stop or meaningfully slow down are engaging in wishful thinking that is bordering on delusion. whatever obstacles anyone thinks are there, there are piles of money to tear them down

(http://www.autoadmit.com/thread.php?thread_id=5859508&forum_id=2),#49836124)