October 2024 Who, What, Where, How, and Why Do I Read – Why Reading Matters Reading means my total consumption of ideas and media, learning via seeing or listening to symbols versus pure action. Reading involves books at the core, but also journal articles, news, blogs, music, video, maps, engineering and architectural drawings, code, patents, walking in cities, conversations with people, and viewing art. If it’s compressed info encoded into my brain and world models, I count it. The line between a reading algo and a learning algo is unclear. Reading is compressed learning, but obviously there are other ways to learn (direct trial and error, apprentice mentor teaching, Socratic dialectic, etc). More subtly, reading is practice directing attention, causing your awareness to curate what comes into that awareness, for a purpose. My reading goal is to use what I read, integrate it into a multidisciplinary whole in my world models, and connect it to problems I care about and am working on. It means updating my web of knowledge. But mere usefulness is too strict a criterion. An equal goal is pleasure. I’ve rewired my brain so reading to learn is highly pleasurable. Often I learn fun and useless material that my curiosity takes me to. This material may be useful later (linear algebra, graph theory, topology, thermodynamics, contemporary classical music, minimalist art). Steve Jobs learned about calligraphy and fonts, and then later created the first great GUI and desktop publishing suite. He is an inspiration. Trivia can be world-changing. Curiosity is a blazing, angelic guide. Generally I’m trying to become a learning machine – to get the best of compressed learning from many humans so I’m not limited to my own experience (like interactive fleet learning for robots and AI, who are early on this steep learning curve). I’m guided by the wisdom of Charlie Munger: In my whole life, I have known no wise people (over a broad subject matter area) who didn’t read all the time – none, zero. . . You’d be amazed at how much Warren reads – at how much I read. My children laugh at me. They think I’m a book with a couple of legs sticking out. Poor Charlie’s Almanack (1st edition, 2005; 2nd edition, 2008). Reading is highly personal. I don’t write about my reading algorithm to recommend it to you or for didactic purposes. Rather, this is descriptive, if my goal for sharing it is to get you to think deeper about your own reading, epistemology, and weltanshauung. Core of the Algorithm I’m looking to build semantic trees of knowledge. A semantic tree is an ordering of information that starts with fundamental principles and branches out into increasingly less fundamental details. For me, it’s the most important compressed knowledge that I can act and build on: math, physics, and the core of humanities. Beyond the tree trunks and branches, I’m also building search and validation functions for knowledge lookup and evaluation. I look for links that people don’t normally make – binary and analog math with biology, classics and finance via assumptions of stationarity or disruption, constitutional/statutory law and software, fixed income macro trading and photography, physics and sculpture, etc. I also love AI because it connects to everything. I take a 20/80 Pareto approach to key ideas – you really can get the most from the semantic tree trunks and branches, if you use your judgment to identify them. This is hard as it’s not always clear ex ante what is a trunk versus a tiny stem. The simplest case of reading is a book, conversing with an author in the past. Books and conversations are elemental and possibly fungible. All books are structured, frozen conversations with people from the past (not even the present, as the moment a book is published it’s already dated). We don’t know how to create self-updatable books yet. Books are not enough. Alongside books, I have conversations, probe recommender algos, and do physical search in libraries, bibliographies, and bookstores. This quickly becomes hyper-textual. The books, conversations with smart and well-read people, and recommenders chart out a bi-directional acyclic graph for me to find more books. What do I do once I have book ideas? I vet them. I may check out reviews, skim through the books, read a chapter, drill through the bibliography. Reading a book is an investment, so I want to prioritize well between my “to read” list of books and the hundreds I have at home (along with the 30-40 I’m actively reading at any point). Beyond sourcing book ideas from conversations, my own reading of references, and bibliographies, I may do a fresh search based on an interest through Amazon or Google books, random search in a library or bookstore, or via online lookup on social and mainstream media. If I’m ever in a great research library or bookstore (e.g UChicago Co-Op, Moes in Berkeley, the UC Berkeley library, the UPenn library), I can get lost for hours. I once worked as a student in an obscure desk in Penn’s Lippincott library and spent hours in the stacks – I can’t believe they paid me to disappear in a sea of books. I look for books that smart people (eg Charlie Munger, Tyler Cowen, Jason Furman, Martha Nussbaum, Richard Posner, Hannah Arendt, John Baez, Jane Goodall, May-Britt Moser, Jim O’Donnell, Balaji Srinivasan) call influential or excellent. This is also true for dead writers I admire. I trace back the genealogy of their influences. I pore over great books lists, course syllabi, end of year best books lists, and reviews of textbooks, including the short introductions series. I try to find and read primary sources whenever I can. Monographs and popular books often lead me to journal articles and textbooks. Then I can get into a ping pong loop between these three, an unceasing cross referential game, until I break the loop. Good textbooks are under-rated. They are pillars of human civilization, with much compressed knowledge. I still buy them to read chapters of material, though I’m more critical and challenge them more too. Science and engineering books demand more effort, while math is quite flexible. These function as active workouts, painful in a good way when you do them, and I feel great after. I don’t feel I need to do math calculations or proofs by hand. Sometimes I do, but often I like to take the big ideas, visualize them, and think of ways I can apply them outside of running mathematical operations. If math is the science of patterns, I want to use the patterns in art or engineering. Classics are the toughest-the great books in the widest sense. They’ve been through many filters of time and entropy, but can be painful and challenging. One of my great joys in life is to rediscover old semantic trunks and greats, or dive into obscuria: Burton’s The Anatomy of Melancholy, Lichtenbert’s The Wastebooks; Latin technical manuals like Quintilian’s Institutio Oratoria or Vitrivius’s De Architectura; Gauss/Euler/Reimann original books and papers (eg. Introductio in analysin infinitorum); early computer architecture texts like Von Neumann’s First Draft of a Report on the EDVAC; classic CS and fractal papers (ELIZA, The Complexity of Theorem Proving Procedures, etc); Francis Bacon and his circle of English degenerates; black hole physics (Bekenstein on Black Holes and Entropy); respiration, etc. Jacques Barzun said this about the classics: But why, after all, learn to read differently by tackling the classics? The answer is simple: in order to live in a wider world. Wider than what? Wider than the one that comes through the routine of our material lives and through the paper and the factual magazines — Psychology Today, House and Garden, Sports Illustrated; wider also than friends’ and neighbors’ plans and gossip; wider especially than one’s business or profession. …The great works do not yield their cargo on demand; but if one reads them with concentration, the effort gives us possession of a vast store of vicarious (indirect) experience; we come face to face with the whole range of perception that mankind has attained and that is denied by our unavoidably artificial (manufactured) existence. …I have said that the classics cannot be read like a magazine article. It takes some form of compulsion to get started, and often the eager starter bogs down in difficulties. Barzun, essay “Of What Use are the Classics?” from Begin Here: The Forgotten Conditions of Teaching and Learning. (1987) For physical versus digital books, it’s both/and. I started accumulating more physical books than my library and shelves at home could contain. So I started getting rid of lower value ones, checked out more from libraries, and indulged any penchant for hoarding via thousands of digital books. I generally retain physical books and interact better with them using scribbles, highlights, and marginalia; digital books are easier for travel and commuting, and for scrapbooks of highlights. One key lesson is to always go back to primary sources, to the original document or artifact when knowledge was created (Newton’s Philosophiæ Naturalis Principia Mathematica, Darwin’s Origin of Species, etc). Books are accumulations of facts and reasoning, but chunks can be mis-represented or just wrong, so you need need to dig back into origins of ideas, facts, experiences and experiments. Often later summarizers or citations miss key details. Translations usually work, but not always for complex languages like Sanskrit or Ancient Greek (Latin translates better). Reading as an act of visual and cognitive tracing isn’t enough. Just because your eyes traced the characters and your mind briefly held the idea doesn’t mean you actually read something. I’ve come across three axes of how to process what you read. One axis of comprehension is: curated; read; evaluated and remembered; connected; acted upon; revisited and cogitated. There’s a second axis on levels of knowledge: information, praxis, & wisdom (tied to emotion and intuition). This also relates to how well reading is properly encoded in your blood, guts, and bones. The third axis is about active cogitation and evaluation while reading, the four levels of reading that Mortimer Adler developed in How to Read a Book [Simon and Schuster, New York, 1972]: 1. Elementary Reading (trace over the text, interpret basic ideas, and retrieve general meaning) 2. Inspectional Reading (rapid reading, getting some main ideas, get a broad overview of the book or article’s structure, themes, and arguments in a short amount of time) 3. Analytical Reading (achieve a comprehensive understanding of the author’s message and systematically analyze its parts for accuracy or flaws) 4. Syntopical Reading (reading multiple texts to build a broader understanding of a subject by synthesizing insights from various sources, creating new ideas, or gaining a multi-faceted perspective). I try to operate on all these axes when I read, but I engage in inspectional reading more to curate material, while I labor with analytical and syntopical reading for deep interaction. Analytical reading means being skeptical about evidence, arguments, and proofs. Check the claims with data and the citations, especially for high stakes ideas. If it’s primary research I’m evaluating, I ask if I believe the methodology and what biases or flaws could invalidate the result. I seek disconfirming evidence to probe and test a mental model. Even the best thinkers may only be right 70-80% of time, while the worst may be right 10-20% of the time. I don’t discount people for false claims since we all make them – truth is hard to ascertain. ἀπεκρίθη ὁ Ἰησοῦς·… ἐγὼ εἰς τοῦτο γεγέννημαι καὶ εἰς τοῦτο ἐλήλυθα εἰς τὸν κόσμον ἵνα μαρτυρήσω τῇ ἀληθείᾳ· πᾶς ὁ ὢν ἐκ τῆς ἀληθείας ἀκούει μου τῆς φωνῆς. λέγει αὐτῷ ὁ Πιλᾶτος· Τί ἐστιν ἀλήθεια. Jesus replied… “In fact, the reason I was born and came into the world is to testify to the truth. Everyone on the side of truth listens to me. “What is truth?” retorted Pilate. John 18:38, SBLGNT 2010, NIV 2011. Syntopical reading means connecting themes, updating your world models, and creating virtual people. First, I like to connect intellectual themes across many books and media: the Enlightenment principles, transcendentalist thought, Darwinism/thermodynamics/organization and capitalism in Victorian England; quantum theory and modernism, post WW2 rationalist optimism, etc. Second, I’m consistently engaged in world model search, curation, read, write and update, and deletion (it’s important to let go of your wrong or bad ideas). I am my own captain; I create the main structures of my own weltanshauung and philosophy, albeit within the seas of my planet’s physics and geography, and my species’ culture and mores. I like to create virtual people in my head of my intellectual giants and heroes, from Plato and Aristotle to the Buddha, Kant, Mill, Einstein, Curie, Keynes, Bardeen and Feynman, Milton Friedman, Nussbaum, Deutsch, etc. Maybe a few dozen of these strange loops live within me, almost like judge and teacher models that help train me as a student with my many world models. As an example, the virtual Feynman teacher and judge may insist on this learning technique: 1. Select a concept to learn 2. Teach it to a child 3. Review and refine your understanding 4. Organize your notes and revisit them regularly. The Feynman judge may say to me: “What I cannot create, I do not understand,” “What do you care what other people think?”, “The first principle is that you must not fool yourself — and you are the easiest person to fool”, or the hundreds of other things he’s said or written in his formal or informal teachings that keep me grounded but also curious. Analytical and syntopical reading are hard. They take effort to execute, and even extra effort to get better at them. It pains me just to think about how bad I am at them. For reading optimization, the best times and places to read, I find that I’m limited. I read best in mornings when my mind is fresh and when all computers and phones are in other rooms. I read well while in transit on trains and planes, but cannot read in cars due to motion sickness. I prefer reading in lounge chairs, or wooden chairs with wide wooden tables. I also have tried many types of reading posture and even tried a reading stand (it didn’t work for me, but propping books at a 30 degree angle is ideal on tables). I read weakly before bedtime. I like reading multiple books at once, maybe 30-40 books over 6-8 themes. I may skim some books, or only read some chapters, or I may read every word. The percent completion doesn’t matter as long as I’ve capture the semantic trunks and maybe some branches. I have re-read some books ten or twenty to thirty times (we call this “upsampling the datamix” when training LLMs). I highlight passages with pens and pencils, or Kindle highlights and bookmarks or page earmarks. For some areas, I keep my own notes in my multi-device app Bear, or very detailed book-length notes for some subjects in Gdocs (eg deep learning, reinforcement learning, unsupervised learning, yoga, etc). I keep books all around me at all times. I’m surrounded by hundreds at home, and have stacks in every room in the house. I always travel with 2-4 books and keep 2-3 books in each car or in backpacks. And then I have hundreds on my phone and iPad. I get nervous in spaces without bookcases, and feel at home in bookstores or libraries with hundreds or thousands of books. It’s probably a character flaw; I’m hopelessly attached to books and what they represent. In the last decade, I’ve increased my consumption of audiobooks and podcasts with different authors and thinkers. These are quite useful when I need to move (walk and run) to counteract my sedentary life, or when driving in the car and reading is perilous. A key part of reading is avoiding attention sinks: social media and streaming video. While I deeply value both of these, including the heated debates on X/Threads or the Golden Age of TV on Netflix and HBO, I’ve learned they are direct competitors for my attention and in the long run I value my reading time more than my watching time. So I limit them with app controls and try to actively engage (e.g. posting versus just reading on social, or curating movies and TV to the best and taking notes). Perhaps the best app is YouTube because you can find someone lecturing on any topic under the sun (my favorites are accessible talks on the Reimann Hypothesis or quantum gravity as I’m unquenchably curious about both as an amateur – Numberphile and Leonard Susskin are brilliant). I end this discussion with two big ideas about reading that I learned the hard way. First, it’s important to stop reading. I need quiet time to think and reflect, to accept or reject what I’ve read and to connect and integrate it. This best happens on walks and runs. My axis of comprehension can only run after I read something: curated, evaluated, and remembered; connected; acted upon; revisited and cogitated. Second, only reading and not getting on with life is selfish and perhaps onanistic. Yes, I read for pure pleasure. But I also read to use and act upon what I read. I actively look for gaps and white spaces to create, invent, write, or build new things. One of my contributions to life is to make things and write about others – to create new books. Read/write is a ying/yang. Doing things gives me many entry points and questions which leads me back to reading. I end with two quotes from Samuel Johnson that I often think about: He said that for general improvement, a man should read whatever his immediate inclination prompts him to; though, to be sure, if a man has a science to learn, he must regularly and resolutely advance. He added, “what we read with inclination makes a much stronger impression. If we read without inclination, half the mind is employed in fixing the attention; so there is but one half to be employed on what we read.” He told us, he read Fielding’s Ameliathrough without stopping. He said, ‘if a man begins to read in the middle of a book, and feels an inclination to go on, let him not quit it, to go to the beginning. He may perhaps not feel again the inclination.” [656] Dr. Johnson advised me to-day to have as many books about me as I could; that I might read upon any subject upon which I had a desire for instruction at the time. “What you read then,” (said he,) “you will remember; but if you have not a book immediately ready, and the subject moulds in your mind, it is a chance if you have again a desire to study it.” He added, “If a man never has an eager desire for instruction, he shall prescribe a task for himself. But it is better when a man reads from immediate inclination.” [766]. Boswell, James and Bergen Evans. The Life of Samuel Johnson. [1st ed.] New York, Modern Library, 1952. Incidentals – Other Details about the Rao Reading Algorithm I’m a proponent of nonlinear reading, skimming, and quitting books. I may read 7 out of 10 chapters in a book out of order, or skip between 8 out of 30 chapters in 3 books. I may start a novel at the end, jump to the beginning, and then end with the middle. I skim many low quality books, and sometimes very hard high quality ones that I don’t have compute cycles to read properly. I quit books all the time – bad books are like bad conversations and I try to leave both quickly. Curation mechanisms for reading: I don’t over think it. I go for where my curiosity takes me, whatever I have a fire for reading for that moment. I can jump between 2-3 books in an hour. AIs and reading: we are very early in what can be done. The LLMs of 2024 are still pretty dumb, at the high school level. Yet they are getting smart very quickly. Exponentials exponentialize. The main perils of AI are the hallucinations, which can occur 20-40% of the time, based on a subject, but where the hallucination rates are dropping. One key problem is we haven’t taught them pragmatic epistemology yet, how as a system to weigh citations and sources and recursively judge and correct. This is an active research area and I expect it to be solved in 3-5 years. Today is the dumbest AI will ever be, so you should continually experiment with them to learn to augment yourself, adapt, and then merge. They will make you a better reader and thinker if you use them carefully with good judgment. Music and video curation: music is emotional instruction and it also helps mathematical and social development. I’ve been known to get lost at night with 2-3 hour music discover searches on Spotify or YouTube, jumping across artists, themes, instruments, music notes, interviews, etc. Videos are too good and addictive. Because they combine many modalities they should mostly be used for education and limited for entertainment. Reading other artifacts is key: architectural plans, maps, engineering diagrams, drawings, code repos, sketches and paintings. It enlarges your abilities and soul. I’m also engaged in some small and large construction projects (an ADU and Llama 4). Print representations can be much wider than words on a 6×9″ or 5.5×8.5″ page. I mostly read a lot of machine learning architecture representations in Arxiv papers, but am generally a fan of maps and technical engineering diagrams too. I’ve been known to spend entire days in places like the Met or MOMA. My test of a great city is how much I can learn by “reading” their culture via their contemporary art museums, or reading the city by examining its multiple maps and walking the streets. Raising kids who read: this is a complex topic and needs its own essay. Basically, kids need to be surrounded by print, learn to read when young, and then have reading as their main form of media pleasure and imagination, not screens. I would keep kids off screens till they are 5, and then start them with a command line prompt after they can read and type. Weekly library and bookstore trips are essential, as is quiet reading time at home with family. Reading traces: Keeping and sharing reading notes, like Jason Furman does on Goodreads or Tyler Cowen does for thoughts, lists, and reviews on Marginal Revolution is a fantastic practice. A societal reading culture is built on book reviews and reading traces. Signs: Randomness, intuition, subconscious, panpsychism- these all play a part in how I pick what to read. I may randomly pick up a book in the library or the Internet Archive, or my intuition or subconscious may tell me to spend an hour deeply examining books in a well-researched bibliography to find more. Sometimes I may just see a book on a table and a shelf and take it as a panpsychic sign that I need to read it, or at least check it out. My algorithm is mostly rational and explainable, but I keep parts open for what some call randomness and intuition, and others call signs from the universe. Limits of energy, compute, and memory – bounded cognition and rationality: as a middle-aged human, I have limited energy, compute cycles, and memory. I can only read a small fraction of what I want and am quite aware of my bounded cognition and rationality. Reading always has tradeoffs in time and energy, though I’ve never regretted reading – it is the activity in my life with the lowest regret rate, near zero. I’ve regretted friendships, meditation, yoga practice, gourmet meals, love and lovers, unskillful action, etc, but never a book I’ve read. The infinity of knowledge: less than one percent of the great books have been written. If we’re a spacefaring species over the next ten thousand years, most of the great books will be in the future. I shudder to think I won’t be able to read most of them but am hopeful for the books of our children, the species of homo interstellaris, both the carbon and silicon ones (or future substrates we cannot imagine). Written in cafes in Rappongi, Daikanyama, and Aoyama Tokyo A nerd post-script: Borges gave this interview at the National Library in 1979