Anthropic's new model shows troubling behavior / hacker news