Manan's notes

Learning in the AI Era

In a world where information is abundant but time is scarce, how do we effectively learn and retain knowledge? Dwarkesh Patel discussed his approach to reading, retaining knowledge, and doing research on a recent episode of the AI & I podcast with Dan Skipper (Highly recommended). For those unfamiliar with Dwarkesh, he hosts deep, insightful conversations with experts across various intellectual fields. His thoughtful, well-researched questions make his podcast a must-listen. If you haven't already, I encourage you to check it out.

Dwarkesh is one of my heroes. This post is my take on the workflows he discussed in the podcast, including my personal highlights, takeaways, and reflections. For a quick overview, you might want to read this summary of the host's notes on the episode.

A note on Claude 3.5 Sonnet: Most of the discussion in the podcast episode centers on using Claude as an AI tool enabling Dwarkesh's workflows. Claude stands out for two main reasons:

  1. It has a large context window of 200k tokens (approximately 200,000 words), allowing users to upload extensive texts, including entire books, references, or structural guidelines.
  2. It demonstrates a level of reasoning and inference that enables meaningful discussions without frequent hallucinations, producing high-quality, thoughtful output that feels more like a discussion partner than a generic answer generator.

As of July 2024, Claude 3.5 Sonnet is arguably the best model for knowledge management, coding, and potentially many other applications. In my subjective experience, I've found Claude to be perhaps two to three times better at reasoning based on text and providing insightful responses compared to models like GPT-4. (Note: This is my personal opinion, not a paid endorsement.)

In the episode, Dwarkesh and Dan explored three ways they've found Claude to be valuable in their learning processes: as a reading partner, a tool for knowledge retention using spaced repetition, and a tool to generate insights during research.

AI as a reading partner

Dwarkesh uses Claude in three ways when reading a new, dense nonfiction book:

  1. Asking Claude to explain the main point of a chapter. For example: "Explain Lynn White Jr.'s take in 'Medieval Technology and Social Change' on how stirrups led to feudalism, individualism, chivalry, knighted class, etc." This serves as both a pre-read and a comprehension check.
  2. Posing follow-up questions about topics not covered in the book but assumed to be common knowledge. For instance: "Why was heavy cavalry so expensive in the Middle Ages?"
  3. Challenging the author's arguments, which either clarifies the point or potentially exposes flaws in the author's reasoning.

AI as a spaced retention tool

Spaced repetition has significantly improved Dwarkesh's retention of information from his interview preparations. He finds it most effective not for memorizing past learning, but for facilitating future learning and creating connections between diverse topics.

Here's how to use AI in your spaced repetition workflow:

  1. After reading, ask Claude to extract key insights and generate flashcards based on this set of guidelines for effective spaced repetition. (I use Mochi as my spaced repetition app and also incorporate this set of rules.)
  2. Create flashcards for novel thoughts or insights not explicitly stated in the text, helping to "cache" your own realizations.

At this point, you might wonder about the usefulness of spaced repetition in a world where AI can answer most questions. Dwarkesh argues that much of learning is relevant for understanding future concepts. Laying out the intellectual territory makes future learning easier. For example, a flashcard made about a concept not fully understood at first might become clearer later as you learn more about the domain. This helps build a foundation of knowledge, enabling connections between concepts that wouldn't be possible without that baseline understanding.

AI as a research assistant

Dan Skipper shared how he uses Claude for complex research tasks, like developing arguments connecting modern AI discourse to Platonic philosophy. By inputting notes and ideas into a Claude project, he can ask it to identify common threads and patterns, leveraging its ability to hold more context in working memory than humans.

Another innovative use Dan discussed is creating a Claude project called "My psychology" for researching one's own psyche. By inputting personal journal entries, goals, and self-insights, you provide Claude with rich context about your thought processes and history. This allows Claude to help you understand aspects of yourself that might not be immediately apparent, acting as a tool for self-reflection and personal growth.

Biggest personal unlocks for me

  • Consuming the same content in various contexts is crucial for learning. LLMs excel at reframing existing content into new contexts.
  • The convenience of instantly generating spaced repetition prompts from recently read content outweighs the potential benefits of creating them manually.
  • Claude's large context window is powerful for knowledge workflows, allowing users to input scattered notes and thoughts and ask questions about connections between different ideas.

I'd love to hear about your experiences using AI in knowledge workflows and any successes you've had!