Why AI Uses So Much Electricity
Training runs are measured in months of full-power computing, and answering a question costs far more than a search. The reason is arithmetic done on a very large scale, plus the cooling to remove the heat.
Training runs are measured in months of full-power computing, and answering a question costs far more than a search. The reason is arithmetic done on a very large scale, plus the cooling to remove the heat.
A new generation of machine learning weather models produces hourly global forecasts at five kilometre resolution, and is being wired directly into consumer products.
The same month that several labs shipped their largest models yet, senior researchers published arguments for deliberately slowing development. Both things are happening at once, and that is the point.
A battery does not store electrons in a box. It stores two chemicals that want to react, and forces the electrons to take the long way round through your device.
Language models are trained to produce plausible text, not true text. Confident errors are not a glitch on top of that design; they are what the design produces when knowledge runs out.
Machine learning is not teaching a computer facts. It is adjusting millions of internal numbers until the model’s mistakes get smaller. Here is what that really means.