this is the equivalent to DC power. someone is going to invent the AI version of AC power and avoid this problem entirely

this is the equivalent to DC power. someone is going to invent the AI version of AC power and avoid this problem entirely
https://x.com/tsarnick/status/1835113831435321799

/bettik founder /outercorps stalwart | ex-binance, ex-celsiusX

This analogy is kind of an ironic one.

With local power generation there is really no need for AC power. Since compute uses DC power and its consumption is local (no need to transmit over large distances) generating AC power would add about 10% inefficiency.

Answering the question "where's the beef?"

the model training is centralized (due to the massive compute cost in discussion), so in this analogy, model training is equivalent to centralized power generation.

running a model can be hardware demanding, but much less than training, and you are correct can often be run locally