How Andrej Karpathy Reads Books With LLMs
A friend sent me a video that Andrej Karpathy made about how he uses large language models (LLMs). According to his website, Karpathy was “a research scientist and founder member at OpenAI” before spending several years at Tesla as Senior Director of AI.
Karpathy shares lots of good info in his video, but it was one use case that prompted my friend to send it to me—and that stood out to me. Karpathy uses LLMs to help him read and understand books and research papers. Here are my takeaways:
- Nowadays, Karpathy mostly reads books using LLMs.
- Here’s his process:
- Pull up the content (i.e., text) of the book.
- Find the chapter he’s going to read.
- The LLM likely knows what the book is about but won’t remember what specific chapters are about.
- Remind the LLM about the chapter by pasting the chapter text into the LLM context window.
- Ask the LLM for a summary of the chapter.
- Read the summary.
- Read the chapter.
- If he has questions about what he’s reading, ask the LLM those questions.
- Reading a book with an LLM dramatically improves his understanding and retention of each chapter. It enhances his ability to understand material in fields he’s unfamiliar with or books written long ago.
- No tools exist that make it easy to do this, so he uses a “clunky and back-and-forth” process.
- He mentions that having the ability to highlight text and ask questions about the highlighted parts doesn’t exist but would be useful.
- Final thought: Don’t read books alone.
I’m glad my friend shared this video with me. It was useful to hear a detailed explanation of how someone leverages LLMs to understand book content. Although he has extensive experience with LLMs, his process is still cumbersome. The fact that he’s following this cumbersome process tells me that the value he’s getting from it must be immense.
The software I’m building will greatly enhance my understanding of the books I read with LLMs. Karpathy’s video confirmed I’m going in the right direction, that other readers are leveraging LLMs, and that they’re experiencing pain when doing so.
If you want to see this section of Karpathy’s video, it’s here. If you want to watch the entire video to see how he uses LLMs in life and work, see here.