How much Energy to Train AGI? ☀️🔋 - Current record: GPT‑4’s full training run ≈ 55 GWh - AGI‑class system (≈ 100 million H100‑equivalent GPUs) projected annual training power draw ≈  613,000 GWh - Four orders of magnitude difference - 2031-2035 is best estimate Energy is a core dependency to AGI Caveats 1/ AI could lead to energy breakthroughs. This is recursive self improvement in a different form since its improving the underlying training making timelines shorter. 2/ Inference will be a huge demand for GPUs. Training doesn't get all GPUs. This will make timelines longer. H/T great convos with @KSimback Link to ChatGPT Question: