Mo Reads: Issue 4
AI, business, charity, culture, literature, math, money, software, rationality
In the last few issues I’ve tried and failed to keep the commentary for each share short. Commentary length is acted upon by directionally opposite sets of forces:
Longer: I like to add context to the share, and also remind myself what exactly I found so interesting about it; this usually involves weaving it into a web of related previous readings that raised questions this share partially answered, like a contribution to an ongoing conversation exploring this topic
Shorter: I’m experimenting with lowering post publication activation energy to solve a recurring problem with my past writing projects — the (self-assessed) average quality would creep up until the bar for publishing something new became prohibitive, after which I’d redirect my info-sharing to Instastories
The goal of this newsletter is to be an infinite game: just keep “playing”, i.e. sharing and commenting, thereby building upon past reads. A side goal is to have what Kevin Simler calls “an exhaust pipe for my intellectual life” — if you don’t actualize new ideas you stop having them, and vice versa. We’ll see how it goes…
As usual, there are 10 links per issue.
(Also I dropped the numbering below the fold; it felt redundant and constraining.)
Links:
TikTok and the Sorting Hat by Eugene Wei (5,600 words, 22 mins)
Speed matters by James Somers (1,000 words, 4 mins)
Wandering the web stacks by James Somers (700 words, 3 mins)
Built-in resistance in writing tools by Ted Hughes (500 words, 2 mins)
Half a year of the liquid tensor experiment: amazing developments by Peter Scholze (2,200 words, 11 mins)
The art of translation by Vladimir Nabokov (2,500 words, 10 mins)
A codebase is an organism by Kevin Simler (2,000 words, 8 mins)
The psychology of money by Morgan Hausel (8,000 words, 32 mins)
Hits-based giving by Holden Karnofsky (5,600 words, 22 mins)
How to think real good by David Chapman (6,500 words, 26 mins)
TikTok and the Sorting Hat by Eugene Wei (5,600 words, 22 mins): new tech cos from China, Eugene reasoned, might never crack the US market (and vice versa) because “the veil of cultural ignorance was too impenetrable a barrier”. TikTok is the first and sole exception; his takeaway is now “in some categories, the veil of cultural ignorance can be penetrated by a sufficiently responsive and accurate ML algo”, i.e. culture can be abstracted. Bytedance bought Musical.ly, rebranded it as TikTok, and did 2 things to jumpstart growth: (1) spend massively on US advertising to acquire users; (2) obsessively improve the For You Page algorithm. This 1-2 combo helped TikTok escape the usual “eww it’s cringe” kiss of death among fickle US teens and break out of its early adopter group by (1) bringing in lots of new people/subcultures, (2) helping them find each other quickly and branch off into their own spaces. The FYP algo, working on short videos, acts as a rapid hyper-efficient matchmaker connecting videos to audiences without an explicit follower graph, and keeping subcultures separated by personalizing FYP feeds; this is the easy “avoid it” solution to the culture wars problem arising from throwing together subcultures into low-trust conversations. (In fact, TikTok is designed around the FYP algo.) The lack of explicit follower graph is a design solution to the following problem: all networks are either heavily social (Facebook), interest (YouTube) or utility (Venmo); building an interest-based network using a social graph (e.g. FB, Twitter) is a hack with negative network effects at scale (“context collapse”) that “amplify the reach of our performative social burden”; so just don’t have a social graph!
Speed matters by James Somers (1,000 words, 4 mins): in the context of work. Reasons: (1) the cost of doing something new feels lower, inclining you to do more; the converse is true, and in fact negative-feedback loops to a halt (2) the cost in others’ minds of interacting with you feels lower, inclining them to interact more; if ‘you’ are an app (e.g. Google search) this encourages rapid user feedback (3) this suggests a prescription for anything you want to get good at: do it faster! (See also Fast, a webpage collecting examples of people quickly accomplishing ambitious things together e.g. the Eiffel Tower and COVID-19 vaccines, maintained by Stripe cofounder Patrick Collison.)
Wandering the web stacks by James Somers (700 words, 3 mins): a sobering personal reminder: if I have a curated collection of channels (e.g. RSS feed, email newsletters, Reddit) sufficiently abundant and interesting that I can afford to stop looking, info-satisfaction reduces to just feed consumption and the wider web might as well not exist, which makes me like the library patron who (having fallen in love with the sci-fi section) forgets all about the books. James suggests some Web-equivalents for books: (1) home pages of professors in fields you’re interested in (e.g. Cosma Shalizi); (2) group blogs by grad students (especially in small departments, e.g. Go Grue); (3) look at the blogrolls of many good blogs and notice that a few really good ones show up in nearly all rolls; (4) most Wiki articles draw heavily on 1-2 more comprehensive sources in the footnotes; (5) ask voracious readers what they read; (6) classic academic papers in unfamiliar fields are ‘classic’ not just because they’re top contributions but also very readable; look at Google Scholar citation counts or intro grad school courses’ syllabi
Built-in resistance in writing tools by Ted Hughes (500 words, 2 mins): typing is much lower-resistance than handwriting, but Ted argues this encourages excess verbosity (“the writer can get down almost every thought or every extension of thought”) which worsens output; the natural resistance of pen-and-paper (claims Ted) makes things “perhaps psychologically denser”. Variants of this argument are pretty common, including by Neil Gaiman, with other purported benefits ranging from therapy to creativity to calmness to age slowdown (by “engaging more motor skills”) to retention/recall to just avoiding other app distractions. I half-buy it, but the benefits for me depend on what I’m trying to do when I write, and there are shortcomings to writing on paper (can’t save/share/link to/copy etc)
Half a year of the liquid tensor experiment: amazing developments by Peter Scholze (2,200 words, 11 mins): Peter Scholze is the most celebrated young mathematician of his generation: a world leader by his mid-twenties, creating entire new subfields and transforming others. In his most important work yet, he wants to argue that topological spaces are “the wrong definition” and replace them with condensed sets; this claim ultimately turns upon a foundational theorem. Its proof is of “arithmetic nature”, so it needs close inspection, but it’s hard enough that nobody has really examined it, and Peter admits he’s occasionally been persuasive with wrong arguments; hence he can’t even be 90% sure it’s true when he needs 100% certainty. Enter the Lean community, a collective effort to build a unified library of math formalized in the Lean theorem prover, who had expressed interest in formalizing some parts of condensed math; Peter challenged them to formalize this foundational theorem. Aside from providing certainty, it would also be “a strong signal that a computer verification of current research in very abstract mathematics has become possible”. The Lean community did it in 6 months! Astoundingly, the formalization also taught Peter why his own proof worked (although he still “has no sense of the terrain” around the proof). Another tidbit: the de Bruijn factor (computer- to human-proof length ratio) is ~20, which Peter says is “amazingly small”
The art of translation by Vladimir Nabokov (2,500 words, 10 mins): there are “three grades of evil” in translation work: (1) obvious errors of ignorance or misguided knowledge (human frailty, hence excusable); (2) omissions the translator can’t bother to understand; (3) redressing it to conform to the notions and prejudices of a given public. The rest of the essay is a beautiful elaboration. I neither can do justice to it nor want to try, so do check it out!
A codebase is an organism by Kevin Simler (2,000 words, 8 mins): what a computer science college education doesn’t prepare you for, claims Kevin, is the insight that while “the computer is a machine, a codebase is an organism” — CS is all about how to control the machine to do exactly what you want, but building “real software” entails dealing with a sprawling unruly codebase managed by other people (who’ll make bug fixes, library updates, “drive-by” refactorings you won’t know about), so just telling it what to do won’t work. (To grok this, it helps to see a codeswarm visualization — an animation of the commit history of a codebase.) Codebase-as-organism suggests decay (code rot via changes to it/dependencies) and growth (usually opportunistically via shortsighted local optimizations), which suggest lessons: (1) execution in a context where someone pays attention to results is the “lifeblood” of a piece of code; (2) be paranoid of code growth — its failure mode is the lava layer antipattern; (3) growth is still necessary, so balance nurture (“coddled code won’t learn its boundaries”) and discipline (“code needs some freedom to grow optimally”); i.e. building software is less like assembling a car, and more like raising a child, or tending a garden. This insight explains received wisdom like “test code as much as possible” and “know your code smells” and “get your APIs right” and “failing fast is good”
The psychology of money by Morgan Hausel (8,000 words, 32 mins): Investing, says Morgan, isn’t the study of finance, it’s the study of how people behave with money. Behavior is hard to teach even to very smart people, which is how sometimes “someone with no education, no relevant experience, no resources, and no connections vastly outperform someone who has all these” (unthinkable in other fields!). This report gives 20 flaws/biases/causes of bad money mgmt. behavior. Examples: (1) a tendency to underestimate the role of luck & risk, and failure to recognize they’re different sides of the same coin; (2) every money reward has hidden non-financial costs (emotional, relationships etc); (3) paradox: people want wealth to signal they should be liked/admired, but admirers bypass them because they use that wealth to benchmark their own desire to be liked/admired; (4) our personal experiences make up <<0.001% of what happened in the world but form ~80% of how we think it works — current investment beliefs depend on past experiences, explaining why people manage their money so differently; (5) studying history of money is useful for general lessons (how people react to greed, fear, stress, incentives) but less so for specifics (trends, trades, sectors) because finance, unlike e.g. geology, progresses via innovation (i.e. change); (6) attachment to social proof (e.g. “I’m value-focused, just like Buffett”) in a field that demands contrarian thinking to beat the averages
Hits-based giving by Holden Karnofsky (5,600 words, 22 mins): effective altruism is a philosophy/social movement that advocates using evidence/reason to do the most good (save lives, benefit others etc) given limited resources, on pain of the opportunity costs of not doing so. The history of philanthropy is replete with examples of enormously impactful interventions — the Green Revolution, the pap smear, the combined oral contraceptive pill etc — that account for outsized share of total impact and compensate for much larger failure counts. This suggests that doing the most good should include a venture capital-like “black swan farming” of low-odds high-upside opportunities. Bonus: it’s also comparatively advantageous for philanthropists, as they’re less constrained by ROI or needing to justify to a wide audience. Holden lists working principles for doing hits-based giving well: assess impact, neglectedness and tractability, consider best and worst plausible cases, minimize number of people setting strategy and making decisions, support strong leadership with no strings attached (“get out of their way, don’t tell them what to do”), understand other funders in the space and avoid good fits for them etc. He also lists principles for sound decision making that aren’t appropriate for hits-based giving: strong evidence base, high success odds, deferring to expert opinion, avoiding controversial positions/conflicts of interest, full legible justification
How to think real good by David Chapman (6,500 words, 26 mins): David spent a ~decade doing research at the MIT AI Lab, during which he thought a lot about thinking in a context that calls your bluff if you’re unwittingly full of it (the AI either works or doesn’t). This wonderfully meandering essay is his brain dump, from those years of research, of situationally-specific rules of thumb (David’s skeptical that a “general theory of intelligence” can be developed at all). Ten tidbits (there’s more): (1) informal reasoning is probably more important to understand than technical methods; (2) most of the work of solving a problem is finding a good formulation for it — this implies (3) before applying any technical method, you need to already have a good idea of what form the answer will take; this partly entails developing a vocabulary to describe relevant factors, because (4) a key to understanding is choosing a good vocab at the right level of description: it makes distinctions used in the solution, and it makes the problem small enough to be solvable; (5) problem formulations aren’t true/false, they’re useful/not; (6) figuring out relevant distinctions to find the “right” vocab requires working through simpler specific examples first, meaning (7) problem formulation and solution are mutually-recursive processes — a “direct approach” isn’t as good; (8) you can never know enough math, given that it’s literally the study of patterns, but it’s more important to know if a branch of math applies than to know how to use it; (9) learn to think like experts in different fields i.e. collect “thinking hats”; (10) collect a bag of tricks for solving problems, and collaborate with others with complementary tricks (e.g. Danny Hillis with percolation theory)