GPT-5.2 Pro directly solved an open problem in statistical learning theory. It was not given strategies or outlines of how to do so, just some prompting/verification.