What I meant to say was that these data/computing centers exist, so they will be used. If any efficiency gains are integrated into western LLMs, there will be more output from a given data center.
It’s like what we see with gains in energy efficiency; instead of energy use being reduced, the thing is typically used more.
Depends on what kind of thresholds get passed I think. Yeah, they’re not gonna immediately start dismantling datacenters if LLMs get efficient cause there’s money invested in it, but well, consider this part of the article:
The day before the Stargate announcement, Trump’s inauguration date, the Chinese company DeepSeek launched its own AI model, claiming it had used far less computing power – and therefore less water – than its western rivals.
More recently, Bloomberg has reported that Microsoft is pulling back on some of its plans for new datacentres around the world.
Are those two events linked, Deepseek’s release and Microsoft pulling back? I don’t know for sure, but it’s possible. Why invest as much into it if the projections start looking like they won’t be needed. That’s where I’m coming from on this.
So far the big thing with generative AI has been that they take ridiculous amounts of compute (GPUs) to train and do inference (generate) on a trained model, with the prevailing belief being that the primary way to keep moving the needle in model quality is to keep throwing more compute at the problem. Deepseek put that more into question, doing more with less (relative to the best out there).
What I meant to say was that these data/computing centers exist, so they will be used. If any efficiency gains are integrated into western LLMs, there will be more output from a given data center.
It’s like what we see with gains in energy efficiency; instead of energy use being reduced, the thing is typically used more.
Depends on what kind of thresholds get passed I think. Yeah, they’re not gonna immediately start dismantling datacenters if LLMs get efficient cause there’s money invested in it, but well, consider this part of the article:
Are those two events linked, Deepseek’s release and Microsoft pulling back? I don’t know for sure, but it’s possible. Why invest as much into it if the projections start looking like they won’t be needed. That’s where I’m coming from on this.
So far the big thing with generative AI has been that they take ridiculous amounts of compute (GPUs) to train and do inference (generate) on a trained model, with the prevailing belief being that the primary way to keep moving the needle in model quality is to keep throwing more compute at the problem. Deepseek put that more into question, doing more with less (relative to the best out there).