What Grok’s recent OpenAI snafu teaches us about LLM model collapse