In roughly the time it takes to watch a film, OpenAI's GPT-5.6 Sol Pro constructed a mathematical counterexample that disproves a 20-year-old conjecture about the most widely used error-control method ...