LLM "reasoning" continues to be deeply flawed / hacker news