# How to read books (in the age of AI) | Lex Fridman Podcast Source: https://www.youtube.com/watch?v=R3convFdfwE Recap page: https://rapidrecap.app/video/R3convFdfwE Generated: 2026-02-07T18:01:03.205+00:00 --- ## Quick Overview Lex Fridman discusses the best way to read books in the age of AI, advocating for building things from scratch (like LLMs or reasoning models) as a superior learning method compared to just reading about concepts or relying on pre-packaged solutions, stressing that building provides concrete understanding and minimizes distraction from the internet. **Key Points:** - Building models like Large Language Models (LLMs) from scratch is considered one of the most powerful ways to learn, offering a deeper understanding than just reading about the concepts. - Sebastian Raschka is introduced as an ML/AI research engineer and author of influential books such as 'Build a Large Language Model From Scratch' and 'Build a Reasoning Model'. - Nathan Lambert is introduced as a researcher at Allen AI, focusing on reasoning, open models, and RL(VR/HF), and author of 'The RLHF Book'. - The discussion contrasts learning by building code (which is precise and verifiable) versus learning purely through reading, where errors might be missed. - The hosts agree that building things like LLMs from scratch forces a deeper understanding, contrasting this with reading others' opinions or relying on pre-packaged solutions. - The advantage of building from scratch is avoiding the 'rabbit hole' of internet opinions and distractions, allowing focus on the core mechanics. - The speaker suggests that while reading books is valuable, using LLMs as an interface layer for data analysis or as a substitute for low-level work can be helpful, but building remains paramount for true learning. ![Screenshot at 00:08: Introduction slide displaying the profiles and Twitter handles of guests Sebastian Raschka \(@rasbt\) and Nathan Lambert \(@natolambert\), who are experts in ML/AI and LLM development.](https://ss.rapidrecap.app/screens/R3convFdfwE/00-00-08.jpg) **Context:** This segment features Lex Fridman in conversation with AI researchers Sebastian Raschka and Nathan Lambert, focusing on the optimal methodology for learning complex technical subjects, specifically in the context of building Large Language Models (LLMs) and related AI systems. The discussion centers on the trade-offs between consuming written knowledge (like books) and actively implementing concepts through coding and building projects from the ground up. ## Detailed Analysis Lex Fridman initiates the discussion by asserting that both guests, Sebastian Raschka and Nathan Lambert, are legitimate experts across research, programming, and education fronts. Raschka is highlighted as the author of influential books on Python-based machine learning, deep learning, and LLMs, including 'Build a Large Language Model From Scratch' and a forthcoming book on reasoning models. Lambert is noted for his research at Allen AI on reasoning, open models, and RL(VR/HF), and for writing 'The RLHF Book'. Fridman then emphasizes his strong recommendation for Raschka's book, 'Build a Large Language Model From Scratch,' stating that building things from scratch is one of the most powerful ways of learning because code is precise and verifiable, unlike reading opinions online. Lambert agrees, noting that when code works, you know it's correct, and the high rate of 'aha moments' when building small-scale LLMs is invaluable. Fridman contrasts this active learning with reading, where one might miss errors or get lost in internet rabbit holes of opinion. Lambert further explains that while he recommends reading books, using LLMs as an interface for data analysis (like scraping data) is fine, but he resists the urge to let an LLM provide the full context or justification for *why* something matters, preferring the direct understanding gained from low-level implementation. ### Guest Introduction - Sebastian Raschka identified as author of 'Build a Large Language Model From Scratch' and 'Build a Reasoning Model' - Nathan Lambert identified as Allen AI researcher and author of 'The RLHF Book' ### Learning Methodology Debate - Building from scratch is the most powerful way to learn, especially in programming, because code is precise and verifiable - Reading opinions or relying solely on LLM explanations lacks depth and context ### LLM Building Experience - Building LLMs from scratch provides high rates of 'aha moments' and avoids internet distractions/rabbit holes - Code correctness is self-evident when it executes properly ### Role of Books and LLMs - Books are highly recommended, but building should be the first pass for deep understanding - LLMs are useful as refreshing tools or interfaces for low-level tasks (like data scraping) but shouldn't replace foundational learning. ![Screenshot at 00:08: Introduction slide displaying the profiles and Twitter handles of guests Sebastian Raschka \(@rasbt\) and Nathan Lambert \(@natolambert\), who are experts in ML/AI and LLM development.](https://ss.rapidrecap.app/screens/R3convFdfwE/00-00-08.jpg) ![Screenshot at 00:18: Lex Fridman referencing Sebastian Raschka's book, 'Build a Large Language Model From Scratch,' which is displayed on screen.](https://ss.rapidrecap.app/screens/R3convFdfwE/00-00-18.jpg) ![Screenshot at 00:48: On-screen display showing the covers of two books by Sebastian Raschka: 'Build a Large Language Model' and 'Build a Reasoning Model', both published by Manning.](https://ss.rapidrecap.app/screens/R3convFdfwE/00-00-48.jpg) ![Screenshot at 01:22: Nathan Lambert explaining the precision of code versus the potential ambiguity of figures/explanations when learning concepts.](https://ss.rapidrecap.app/screens/R3convFdfwE/00-01-22.jpg) ![Screenshot at 03:08: Lex Fridman gesturing widely while describing the danger of getting sucked into the 'rabbit hole' of internet opinions when learning new topics.](https://ss.rapidrecap.app/screens/R3convFdfwE/00-03-08.jpg)