Re:Apple squandered the Holy Grail
by Hilift on Monday January 06, 2025 @10:43AM
(#42611486)
Power consumption is the defining characteristic of AI. The power consumption by the US had recently plateaued at 4,000 billion kilowatt hours 2000 through 2023. That will likely accelerate by 20% or more with 2024/2025 data. It's probably one of the few guardrails. Electricity is about five times more expensive in the UK than the US. So the US is the natural home for the models and other regions are not.
[ Parent
| on HN ]
I'm kind of skeptical the US grid can even handle this industry growth if it becomes the only realistic place these models are ran. A lot of the infrastructure is pretty wobbly. And forget the tax benefits and cheap land from Texas, their private grid is liable to bust at any time.
> Power consumption is the defining characteristic of AI.
I'm not so convinced. I've been playing with running Ollama + llama3.1 8B on my 2023 M2 MacBook Air with 24GB of RAM, and I don't notice much difference in battery life with or without it. I'm not querying it continually with a shell script in a loop or anything like that, but neither am I shy about throwing all kinds of prompts at it. My laptop keeps chugging away on battery.
Training AI may be ferociously resource intensive, but I haven't seen that querying those models is especially bad. I'd think that a model that Apple had tailored specifically to run well on its hardware would be relatively "light".