Like I said before, people out here expecting something like JARVIS or KITT (eighties kids represent) and what they're getting is something like JARVIS or KITT after a lobotomy.
A lot of this AI stuff is a load of garbage and it's bad for the environment because it needs loads of power which is wasting resources, it means computers run for longer, meaning they wear out quicker meaning more waste when they're junked. I know there's been some good come of it, but this chatGPT stuff is just absolute pony and trap.
Or more like expecting KITT and getting KARR instead (which has no problem in feeding you false information when it feels like doing so).
But actually, can you speak up please, I’m having a hard time hearing you over the humming of all the processor fans in my warehouse full of Crypto Mining Servers.
Although, all these are about to be scrapped anyways because they just came out with a model that can do something in a billionth of a nanosecond faster than what these current ones can.
Like I said before, people out here expecting something like JARVIS or KITT (eighties kids represent) and what they're getting is something like JARVIS or KITT after a lobotomy.
A lot of this AI stuff is a load of garbage and it's bad for the environment because it needs loads of power which is wasting resources, it means computers run for longer, meaning they wear out quicker meaning more waste when they're junked. I know there's been some good come of it, but this chatGPT stuff is just absolute pony and trap.
I wouldn't go that far. For one thing, a lot of people are assuming that the only direction LLMs can go is bigger and more computationally intensive, but I think that's not true. I think where LLMs are now is a transition state we had to pass through before we get to what they will eventually become. We have only scratched the surface of understanding both the mathematics of the phase space dynamics of LLMs, and the ability for quantization to preserve the functionality of the system.
There's an urban legend that in the 1940s Thomas Watson (then the chairman of International Business Machines - IBM) said the world market for computers was maybe five. That is not true, but the legend persists because there was in fact a belief by many in the 40s and 50s that electronic computers were giant hulking machines that might be useful, but relatively few people would actually need. Current LLMs are the IBM mainframes of the AI world. I have a bet with a few people that something functionally identical to chatGPT is going to fit in a device you can fit on a desktop in ten years. When it comes to these kinds of technology bets, I have never lost.
Also, I don't know where you get the idea that LLMs force computers to run for "longer" meaning they "wear out quicker." Maybe 30 years ago such a statement might have been true for some data centers, but that's not true for modern infrastructure. First of all, most modern compute centers are either dedicated computational workhorses that by definition are running constantly - and such systems have always run with basically a 100% duty cycle - or they are virtualized to prevent wasted rackspace, which again means they are running basically with a 100% duty cycle. What's more, it isn't duty cycle that kills systems, it is environmental control. Stable temperature and proper cooling will keep a 100% duty cycle system running for longer than its productive lifetime. The average lifespan of our hosting cluster members far exceeds the time before we cycle them out for power/performance reasons (about 4-5 years) and turn them into test servers or boat anchors. In 23 years we've had exactly four servers completely die while still in a live production cluster.
Physical spindle hard drives are mechanical devices, and they can mechanically wear out. But integrated circuits when operated within their specifications do not wear out with use. The capacitors on the motherboards will dry out and die from old age before that happens (and that sort of degradation with time is responsible for more failures than operational stress).
Comments
But actually, can you speak up please, I’m having a hard time hearing you over the humming of all the processor fans in my warehouse full of Crypto Mining Servers.
Although, all these are about to be scrapped anyways because they just came out with a model that can do something in a billionth of a nanosecond faster than what these current ones can.
There's an urban legend that in the 1940s Thomas Watson (then the chairman of International Business Machines - IBM) said the world market for computers was maybe five. That is not true, but the legend persists because there was in fact a belief by many in the 40s and 50s that electronic computers were giant hulking machines that might be useful, but relatively few people would actually need. Current LLMs are the IBM mainframes of the AI world. I have a bet with a few people that something functionally identical to chatGPT is going to fit in a device you can fit on a desktop in ten years. When it comes to these kinds of technology bets, I have never lost.
Also, I don't know where you get the idea that LLMs force computers to run for "longer" meaning they "wear out quicker." Maybe 30 years ago such a statement might have been true for some data centers, but that's not true for modern infrastructure. First of all, most modern compute centers are either dedicated computational workhorses that by definition are running constantly - and such systems have always run with basically a 100% duty cycle - or they are virtualized to prevent wasted rackspace, which again means they are running basically with a 100% duty cycle. What's more, it isn't duty cycle that kills systems, it is environmental control. Stable temperature and proper cooling will keep a 100% duty cycle system running for longer than its productive lifetime. The average lifespan of our hosting cluster members far exceeds the time before we cycle them out for power/performance reasons (about 4-5 years) and turn them into test servers or boat anchors. In 23 years we've had exactly four servers completely die while still in a live production cluster.
Physical spindle hard drives are mechanical devices, and they can mechanically wear out. But integrated circuits when operated within their specifications do not wear out with use. The capacitors on the motherboards will dry out and die from old age before that happens (and that sort of degradation with time is responsible for more failures than operational stress).