DeepSeek: Reverse Engineering an AI Assistant by Interviewing Itself

The author conducted an extended interview with DeepSeek, an AI assistant, asking it to explain how it works, then verified its answers against published research papers. DeepSeek carefully distinguished between what it observes about itself, what it infers, and what it guesses, while acknowledging it cannot see its own weights or attention mechanisms. The analysis reveals that DeepSeek's self-reported capabilities align partially with public documentation, but architecture specifics should be verified through arXiv papers rather than relying solely on the model's self-description.