Efficiency is a big deal in datacenters. Aside from the AI bubble, companies actually don’t like lighting money on fire. More efficient hardware saves them lots of money; less hardware, less power, less datacenter space, less maintenance, fewer support contracts, fewer software licenses, etc etc.
While the current splashy “state of the art” models in terms of cognitive ability are American, IMO the real foundation for future AI is coming out of China these days. It’s not quite as smart but they’re focusing heavily on making AI training and inference cheaper in terms of compute (and therefore more efficient in terms of energy usage). It’s a mother-of-invention situation, sure - they’ve been cut off from the latest and greatest NVIDIA cards so they’re having to find ways to make do with less powerful hardware. But that’s going to be super important once AI is “good enough” for various real world tasks and the most powerful models aren’t needed for most activities.
That’s not what I said. We’re using those hardware efficiency gains to offset the performance losses of additional abstraction layers. If we were to make the software more efficient, the hardware efficiency gains would actually be noticeable and we wouldn’t be wasting nearly as much energy overall.
Efficiency is a big deal in datacenters. Aside from the AI bubble, companies actually don’t like lighting money on fire. More efficient hardware saves them lots of money; less hardware, less power, less datacenter space, less maintenance, fewer support contracts, fewer software licenses, etc etc.
While the current splashy “state of the art” models in terms of cognitive ability are American, IMO the real foundation for future AI is coming out of China these days. It’s not quite as smart but they’re focusing heavily on making AI training and inference cheaper in terms of compute (and therefore more efficient in terms of energy usage). It’s a mother-of-invention situation, sure - they’ve been cut off from the latest and greatest NVIDIA cards so they’re having to find ways to make do with less powerful hardware. But that’s going to be super important once AI is “good enough” for various real world tasks and the most powerful models aren’t needed for most activities.
That’s not what I said. We’re using those hardware efficiency gains to offset the performance losses of additional abstraction layers. If we were to make the software more efficient, the hardware efficiency gains would actually be noticeable and we wouldn’t be wasting nearly as much energy overall.
This is all ai compute. Yes modern software is bloated but ai inference currently kind of has to run on gpus.