A lot of work has been going into making AIs more energy efficient, both in training and in inference stages. Electricity costs money, so obviously everyone’s interested in more efficient AIs. That makes them more profitable.
Funny you should mention blockchains. Ethereum, the second-largest blockchain after Bitcoin, switched from proof-of-work to a proof-of-stake validation system two and a half years ago. That cut its energy use by 99.95%. The “blockchains are inherently a huge waste of energy” narrative is just firmly lodged in the popular view of them now, though, despite it being long proven false.
It means that even if AI is having more environmental impact right now, there’s no reason to say “you can’t improve it that much.” Maybe you can improve it. As I said previously, a lot of research is being done on exactly that - methods to train and run AIs much more cheaply than it has so far. I see developments along those lines being discussed all the time in AI forums such as /r/localllama.
Much like with blockchains, though, it’s really popular to hate AI and “they waste enormous amounts of electricity” is an easy way to justify that. So news of such developments doesn’t spread easily.
A lot of work has been going into making AIs more energy efficient, both in training and in inference stages. Electricity costs money, so obviously everyone’s interested in more efficient AIs. That makes them more profitable.
Still you can’t improve it that much. It’s like blockchain. Computers always consume a lot of power, no matter how efficient they are.
Funny you should mention blockchains. Ethereum, the second-largest blockchain after Bitcoin, switched from proof-of-work to a proof-of-stake validation system two and a half years ago. That cut its energy use by 99.95%. The “blockchains are inherently a huge waste of energy” narrative is just firmly lodged in the popular view of them now, though, despite it being long proven false.
But that’s really good! And also means that cloud based AI is even worse than blockchain in terms of environmental impact.
It means that even if AI is having more environmental impact right now, there’s no reason to say “you can’t improve it that much.” Maybe you can improve it. As I said previously, a lot of research is being done on exactly that - methods to train and run AIs much more cheaply than it has so far. I see developments along those lines being discussed all the time in AI forums such as /r/localllama.
Much like with blockchains, though, it’s really popular to hate AI and “they waste enormous amounts of electricity” is an easy way to justify that. So news of such developments doesn’t spread easily.
You can improve it hugely. These things are very young.
There was a paper recently about removing the need for matrix multiplication from them which is a hugely expensive operation.
Dedicated hardware is also at a very early stage.