Remix.run Logo
▲ scott_weber an hour ago

For now. Atm all compute is like 1% of global energy usage. Until hyper scalers want to 100x their energy spend, which they'll might need to get AGI (to get a step change in capability for along scaling law lines, you need both 10x model size and 10x training data, 100x as much computation. They probably want at least 2 more step changes, so even assuming they 100x efficiency again, there's another 100x that needs raw compute hence energy). Then we're doubling global energy usage.