“Edouard Harris believes his work is the first direct experimental evidence for the instrumental convergence thesis in AI safety.”