Seeing Like an Algorithm

Seeing Like an Algorithm

English

Seeing Like an Algorithm September 20, 2020 by Eugene Wei In my previous post on TikTok I discussed why its For You Page algorithm is the connective tissue that makes TikTok work. It is the bus on its motherboard that connects and closes all its feedback loops. But in the breathless rush to understand why companies might want to acquire TikTok, should ByteDance be forced to divest itself of the popular short video app, the hype around its algorithm has taken on a bit of exoticization that often characterizes Western analysis of the Chinese tech scene these days.I kept holding off on publishing this piece because every day seemed to bring some new development in the possible ban of TikTok in the U.S. And instead of writing any introduction that would become instantly outdated, I'll just leave this sidenote here to say that as of publishing this entry, it seems Oracle will take over the TikTok cloud computing deal while also joining Wal-Mart and some VCs in assuming some ownership stake in TikTok Global. But it won't surprise me one bit if we find out even more bizarre details over the next week. This is the type of deal that I would have thought could only happen in Succession, but even in that satire it would seem hyperbolic. The 2020 Writer's Room is undefeated. In this post, I want to discuss exactly how the design of TikTok helps its algorithm work as well as it does. Last time I discussed why the FYP algorithm is at the heart of TikToks flywheel, but if the algorithm wasnt effective then the whole feedback loop would collapse. Understanding how the algorithm achieves its accuracy matters even if youre not interested in TikTok or the short video space because more and more, companies in all industries will be running up against a competitor whose advantage centers around a machine learning algorithm. What I want to discuss is how TikToks design helps its algorithm see. Seeing Like a State by James C. Scott is one of those books that turns you into one of those Silicon Valley types that use (abuse?) the term legibility. I first heard about it after reading Venkatesh Raos summary of its main themes, and that piece remains a good tldr primer on the book if you dont plan to read the text (Scott Alexander's review of the book is also good though is long enough that it could almost justify its own tldr). However, I recommend that you do. The subtitle of Scotts book is How Certain Schemes to Improve the Human Condition Have Failed. In particular, Scott dissects a failure state that recurs across a number of domains, in which a governing body like the nation-state turns to what Scott terms high modernism in an effort to increase legibility of whatever it is they are trying to exert control over, whether for the purposes of taxation or conscription or any number of goals. In doing so, they impose a false sense of order on a reality more complex than they can imagine.It would really be fascinating to hear from Scott on the case of modern China, under CCP rule, with modern technology for surveillance, and whether he thinks they will prove or violate his thesis in the fullness of time. Its a book that raises ones awareness of all sorts of examples of unintended consequences in day-to-day life. We all could use a healthier does of humility when we are too flush with great man hubris. The world is richer and more complicated than we give it credit for. As an example, much of what Scott discusses has relevance to some of the hubris of our modern social networking giants. These dominant apps are designed to increase legibility of their user bases for, among other things, driving engagement, preventing churn, and ultimately, serving targeted advertisements. That, in turn, has led their parent companies into a thicket of problems which theyre grappling with constantly now. But that is a topic for another post, another day. Whereas Scott focuses in on how the nation-state uses simplifying abstractions to see its citizens at a synoptic level, I want to discuss how TikToks application design allows its algorithm to see all the detail it needs to perform its matchmaking job efficiently and accurately. If Seeing Like a State is about a common failure state, this post is about a new model for getting the most leverage from machine learning algorithms in the design of applications and services.Im aware of the irony that the controversy around TikTok was the potential of user data being accessed by the CCP, or being seen by that state. Or that one of the sticking points of this new Cold War is the Chinese Firewall, which selects what the citizens of China see. And which most U.S. tech companies sit outside of, looking in. In recent years, one of the realizations in machine learning, at least to an outsider to the subject like myself, is just how much progress was possible just by increasing the volume of training data by several orders of magnitude. That is, even if the algorithms themselves arent that different than they were a few years ago, just by training them on a much larger datasets, AI researchers have achieved breakthroughs like GPT-3 (which temporarily gave tech Twitter a tantric orgasm). When people say that TikToks algorithms are key to its success, many picture some magical block of code as being the secret sauce of the company. The contemporary postmodernist Russian writer Viktor Pelevin has said that the protagonist of all modern cinema is a briefcase full of money. From the briefcase of radioactive material (I think thats what it was?) in Kiss Me Deadly to the briefcase of similarly glowing who knows what (Marcellus Wallaces soul?) in Pulp Fiction, from the Genesis equation in The Formula to the secret financial process in David Mamets The Spanish Prisoner, weve long been obsessed in cinema with the magical McGuffin. In recent weeks, discussion of TikToks algorithm has elevated it into something similar, akin to one of those mystical archaeological artifacts in one of the Indiana Jones films, like the Ark of the Covenant, the Holy Grail, or the lingam Shivling. But most experts in the field doubt that TikTok has made some hitherto unknown advance in machine learning recommendations algorithms. In fact, most of them would say that TikTok is likely building off of the same standard approaches to the problem that others are. But recall that the effectiveness of a machine learning algorithm isnt a function of the algorithm alone but of the algorithm after trained on some dataset. GPT-3 may not be novel, but trained on an enormous volume of data, and with a massive number of parameters, its output is often astonishing. Likewise, the TikTok FYP algorithm, trained on its dataset, is remarkably accurate and efficient at matching videos with those who will find them entertaining (and, just as importantly, at suppressing the distribution of videos to those who wont find them entertaining). For some domains, like text, good training data is readily available in large volumes. For example, to train an AI model like GPT-3, you can turn to the vast corpus of text already available on the internet, in books, and so on. If you want to train a visual AI, you can turn to the vast supply of photos online and in various databases. The training is still expensive, but at least copious training data is readily at hand. But for TikTok (or Douyin, its Chinese clone), who needed an algorithm that would excel at recommending short videos to viewers, no such massive publicly available training dataset existed. Where could you find short videos of memes, kids dancing and lip synching, pets looking adorable, influencers pushing brands, soldiers running through obstacle courses, kids impersonating brands, and on and on? Even if you had such videos, where could you find comparable data on how the general population felt about such videos? Outside of Musical.lys dataset, which consisted mostly of teen girls in the U.S. lip synching to each other, such data didnt exist. In a unique sort of chicken and egg problem, the very types of video that TikToks algorithm needed to train on werent easy to create without the apps camera tools and filters, licensed music clips, etc. This, then, is the magic of the design of TikTok: it is a closed loop of feedback which inspires and enables the creation and viewing of videos on which its algorithm can be trained. For its algorithm to become as effective as it has, TikTok became its own source of training data. To understand how TikToks created such a potent flywheel of learning, we need to delve into its design. The dominant school of thought when it comes to UI design in tech, at least that Ive grown up with the past two decades, has centered around removing friction for users in accomplishing whatever it is theyre trying to do while delighting them in the process. The goal has been design that is elegant, in every sense of the word: intuitive, ingenious, even stylish. Perhaps no company has more embodied this school of design than Apple. At its best, Apple makes hardware and software that is pleasingly elegantit just worksbut also sexy in a way that makes its users feel tasteful. Apples infamous controlling styleno replaceable batteries for its phones and laptops, the current debate over its App Store rulesput the company squarely in the camp of what Scott in Seeing Like a State refers to as high modernism. Is there any reason to show a video of how the new MacBook Pro body is crafted from one solid block of aluminum (besides the fact that Jony Ive cooing a-loo-MIN-eee-um is ASMR to Apple fans) when unveiling it at an Apple keynote? How about because its sexy AF to see industrial lasers carving that unibody out of a solid chunk of aluminum? And later, when youre cranking out an email at a coffee shop on said laptop, some residual memory of that video in your unconscious will give you just the slightest hit of dopamine? Theres a reason this user-centric design model has been so dominant for so long, especially in consumer tech. First, it works. Apples market cap was, at last check, over 2 trillion dollars. Remember when fake Sean Parker said a billion dollars was cool? That was just a decade ago and a billion dollars is no longer S-Tier. The wealth meta moves fast. Furthermore, we live in the era of massive network effects, where tech giants who apply Ben Thompsons aggregation theory and acquire a massive base of users can exert unbelievable leverage on the markets they participate in. One of the best ways to do that is to design products and services that do what users want better than your competitors. This school of design has been so dominant for so long that Ive almost managed to forget some of the brutal software design that used to the norm in a bygone era.Not to be confused with brutalist design, which can be quite beautiful in its own respect, like its architectural cousins. But what if the key to serving your users best depends in large part upon training a machine learning algorithm? What if that ML algorithm needs a massive training dataset? In an age when machine learning is in its ascendancy, this is increasingly a critical design objective. More and more, when considering how to design an app, you have to consider how best to help an algorithm see. To serve your users best, first serve the algorithm. TikTok fascinates me because it is an example of a modern app whose design, whether by accident or, uhh, design, is optimized to feed its algorithm as much useful signal as possible. It is an exemplar of what I call algorithm-friendly design.I thought about calling it algorithm-centric design but felt it went too far. Ultimately, a design that helps an algorithm see is still doing so in service of providing the user with the best possible experience. This might still be considered just a variant of user-centric design, but for those teams working on products with a heavy machine learning algorithm component, it may be useful to acknowledge explicitly. After all, when a product manager, designer, and engineer meet to design an app, the algorithm isn't in attendance. Yet its training needs must be represented. James Scott speaks of seeing like a state, of massive shifts in fields like urban design that made quantities like plots of land and their respective owners legible to tax collectors. TikToks design makes its videos, users, and user preferences legible to its For You Page algorithm. The app design fulfills one of its primary responsibilities: seeing like an algorithm. Lets take a closer look. TikTok opens into the For You Page and goes right into a video. This is what it looks like. View fullsize This is, as of right now, the most popular TikTok ever. By the time I publish this post, its 34.1M likes will likely be outdated. You can read the story of how this TikTok even came to be and it will still feel like a cultural conundrum wrapped in a riddle stuffed in a paradox, and you love to see it. I showed this to my niece, we looped it a few dozen times, then we started chanting M to the B, M to the B and laughing our asses off and it was one of the only times in this pandemic Ive truly felt anything other than despair. The entire screen is filled with one video. Just one. It is displayed fullscreen, in vertical orientation. This is not a scrolling feed. Its paginated, effectively. The video autoplays almost immediately (and the next few videos are loaded in the background so that they, too, can play quickly when its their turn on stage). This design puts the user to an immediate question: how do you feel about this short video and this short video alone? Everything you do from the moment the video begins playing is signal as to your sentiment towards that video. Do you swipe up to the next video before it has even finished playing? An implicit (though borderline explicit) signal of disinterest. Did you watch it more than once, letting it loop a few times? Seems that something about it appealed to you. Did you share the video through the built-in share pane? Another strong indicator of positive sentiment. If you tap the bottom right spinning LP icon and watch more videos with that same soundtrack, that is additional signal as to your tastes. Often the music cue is synonymous with a meme, and now TikTok has another axis on which to recommend videos for you. Did you tap into the video creators profile page? Did you watch other videos of theirs, and did you then follow them? In addition to enjoying the video, perhaps you appreciate them in particular. But lets step back even earlier, before youre even watching the video, and understand how the TikTok algorithm sees the video itself. Before the video is even sent down to your phone by the FYP algorithm, some human on TikToks operations team has already watched the video and added lots of relevant tags or labels. Is the video about dancing? Lip synching? Video games? A kitten? A chipmunk? Is it comedic? Is the subject a male or female? What age, roughly? I

中文

翻译失败，原文保留。