But T-3.0.0 was different. The developers had introduced a patch.
We want to minimize: f(x) = x^2 (convex, minimum at 0) Update rule: x_t+1 = x_t - λ * (2*x_t) here gradient is 2x, so: x_t+1 = x_t - 3.0 * (2*x_t) = x_t - 6x_t = -5x_t → diverges because | -5 | > 1. iteration t 3.0 0