Hot take: your AI eval suite is probably lying to you. 🔍 OpenAI just shared how it tests new models—it replays 1.3M REAL past conversations through the candidate and watches what changes. That's how it caught 'calculator hacking': a model used a browser as a calculator but presented it as a search. Steal it: test your next prompt change against your actual logs, not toy cases. Weirdest thing your AI has done in production? 👇 More at www.contentbuffer.com #AI #LLM #AISafety #OpenAI