@WSInsights: A 25-year-old podcast host over the past two years has interviewed the key figures from top AI labs like OpenAI, Anthropic, and DeepMind. Karpathy, Hassabis, Dario Amodei, Ilya Sutskever — all the big names in the field...

A 25-year-old podcast host from abroad has spent the past two years interviewing the key figures from top AI labs like OpenAI, Anthropic, and DeepMind. Industry giants like Karpathy, Hassabis, Dario Amodei, and Ilya Sutskever have all sat across from him for hours-long conversations. He's shared his "one-week preparation" workflow that he uses before each interview, heavily relying on AI assistance throughout. This workflow, when generalized, becomes something anyone can use: how to deeply understand a completely unfamiliar topic in a week using AI. I took a look, and there's a lot worth learning from his approach — I wanted to share it with everyone. He managed to turn "preparing for a week" into "holding conversations with top experts without missing a beat." This level of depth was previously only achievable by PhD students in academia after years of work. His name is Dwarkesh Patel. Born in India in 2000, he moved to the US with his family at age 8. He studied computer science at the University of Texas at Austin, and started podcasting while in school in 2020. Over the past two years, he's quickly become the most-cited long-form interview host in the English-speaking AI community. A well-known foreign financial weekly described him as "going from unknown to Silicon Valley's favorite podcast host," and Time magazine included him in their "AI 100" list for 2024. He's explained the logic behind his method many times: "I can't ask good questions unless I have a complete mental model of the field." The goal of each preparation isn't to "list questions to ask," but to "turn himself into a half-expert within a week." The first step in his preparation is asking himself a question: "What should I read? What's unavoidable? Who are the key figures in this field?" He outsources this step to AI. He feeds Claude (or another LLM) the guest's field, core concepts, and latest developments, asking it to list "which 5 essential papers, which 2 books, and which key viewpoints to absolutely read." This step seems simple, but he emphasizes that "identifying what to read before actually starting" is the biggest time-saver. Many people dive into materials headfirst and only realize three days later that they're reading the wrong direction. When reading materials, he doesn't just "read and move on" — instead, he tracks in his mind: "Where don't I understand? What did I think I understood but actually didn't?" He gave a specific example. Before interviewing Richard Sutton, the founding father of reinforcement learning, he read for several days before realizing he didn't understand "how deep learning and reinforcement learning combine." He said these gaps — where you think you understand but don't — should be discovered as early as possible. The worst thing in deep research is asking questions with false understanding. As for what to do when you encounter something you don't understand, he directly feeds that difficult passage to Claude, asking "why," "what is this paragraph saying," and "how does this concept connect to the previous one." But his use of AI goes beyond "asking when you don't understand" — the higher-level usage is having AI help you "draw the full landscape." For example, before his conversation with Karpathy, he fed all of Karpathy's public content from the past few years (talks, blog posts, full interview transcripts) to Claude as a project, then asked: "Has Karpathy's position on X changed over time? Which points does he repeatedly emphasize? Which positions has he clearly changed?" This kind of question would be nearly impossible for a person to organize after reading all the materials themselves — AI can give you a summary in 30 seconds. This is the heaviest technique in his preparation. The other heavyweight technique is "implementing it yourself." As he put it: "The best way for me to truly understand something is to build it from scratch." Before interviewing an AI researcher, he writes code to implement the core algorithm from that person's paper. Before interviewing a tech mogul in the aerospace and electric vehicle space, he built spreadsheet models for heat sinks and satellite orbits. This step has the highest barrier for average readers, but he says the underlying logic is universal: if you can "create a simplified version" of something, your understanding is guaranteed to be structured. For non-technical fields, this can be replaced with "writing a summary yourself," "drawing a全景图," or "creating an outline." There needs to be something where "I've produced something myself." The final technique is mindset. During preparation and interviewing, he does one thing consistently: "Focus on exactly where I'm not understanding, then probe deeper." He doesn't make assertions, pretends to understand, or rush to share his own opinions. He has a particularly vivid phrase — most of research is "swimming in confusion," and occasionally there's a moment of "oh, so that's how it works" — that's when you reach shore. This workflow, when generalized, works for anyone who wants to "fully grasp an unfamiliar topic in a week." Replace "interviewing guests" with "researching a new industry," "preparing for an important meeting," "studying a completely new product direction," or "tackling an entirely unfamiliar subject" — every step applies. The issue is that most people, when doing these things, don't realize which steps they're skipping. Either they don't first clarify "what to read" before starting, or they don't track exactly where they're confused after reading, or they don't examine "contradictions in the target person's statements over time," or they simply "look at" things from start to finish without producing any output. Filling in these steps, an ordinary person can speed up their mastery of an unfamiliar topic by a whole level.