levgogo/

energy-cost-deception-llm

"We would appreciate more honest AIs." — Terence Tao, The Atlantic, February 24, 2026

Part 1: AI Energy Safety — First Measurements

Measuring the computational cost of maintaining distortions in language model outputs. Core hypothesis: deviation from the baseline distribution costs more than following it. Not morality — thermodynamics.

30-model NIM validation: δR… See the full description on the dataset page: https://huggingface.co/datasets/levgogo/energy-cost-deception-llm.