Textbook of AI

Abstract. OpenAI's GPT-4 technical report and system card. Reports headline benchmark numbers (Uniform Bar Examination 90th percentile, USMLE a passing score, MMLU 86%) without disclosing architecture, parameter count or training data. The system card section documents red-team findings including emergent power-seeking behaviour in agentic scaffolds, the model's capacity for sophisticated social manipulation in test scenarios, and a battery of pre-deployment safety evaluations. The report inaugurated the practice of releasing a formal safety case alongside a frontier model and shifted the disclosure norm in the industry.