Long time, no talk. I’ll spare you the usual excuses and apologies and skip right to it.
As I’m currently working with a friend on Something about data and datafication, I wanted to bang out a couple of quick thoughts here that aren’t totally germane to the main thrust of our thing. As an entry point, smooth as an ocean liner, I want to draw your attention to this amazing Financial Times Alphaville feature. This is not the “Lunch with the FT” feature where Roula Khalaf interviews the OpenAI CEO Sam Altman over pasta in his Napa home. Rather, this is the blog feature accompaniment to that piece, where someone named Bryce Elder (Bryce, if you’re reading this, I love you) reads Altman for absolute filth. To wit: Altman actually cooks with the “Drizzle” offering from the trendy Graza (gra-tha, if you’re insufferable like that, which I am) olive oil brand. You should never heat a finishing oil! The bottle Altman cooks with, Elder notes with appropriate outrage, is actually one of two open bottles in his kitchen. It’s perverse, unbelievably indulgent, like paneling the exterior of your Subaru beater in solid mahogany. Sam Altman also has an expensive espresso machine that functions poorly and is widely panned online, complete with an expensive and pointless add-on doohickey perched atop it. This doohickey is called, in all solemnity, a “puck sucker.” Altman wields a knife to prepare the lunch that vexes Elder to no end. It looks expensive, but it could be cheap; he wonders about where the center of gravity in the handle is – more towards the tip, as with a Japanese knife, or more towards the center, for the kinds of vegetable-chopping more common to European techniques? Impossible to say. The blade is curved, making it, I agree, a puzzling choice to use to chop garlic. (He made them pasta aglio e olio? Give me a fucking break!)
We care about Sam Altman because, as I said, he is the CEO of OpenAI. OpenAI is in, to put it mildly, trouble. Ed Zitron has called the company “a systemic risk to the tech industry” and it’s hard to see how he’s wrong. OpenAI is currently wrangling with Microsoft, its biggest corporate supporter, over the terms of a restructuring deal whose parameters I don’t quite understand but which I’m sure is certain to pointlessly burn even more money than the company has already wantonly torched. But I didn’t come here to eulogize the most money-losing company in the history of capitalism. I came here to talk about the basis of the AI, which is to say the tech industry, which is to say the stock market’s value, which is data.
Back when I was most recently on Bluesky again (ugh), I got into an argument with somebody about how even the newest AI models don’t work – they are, in the words of Ed Zitron again, increasingly expensive and unsustainable “lossy bullshit.” (In machine learning jargon, “loss” is another word for “error.”) To support my original claim I maintained, correctly, that these models cannot really tell you what 2+2 is. My Bluesky interlocutor countered that the newer models are actually much improved over their older relatives, such that they can give answers to queries, queries like “what is 2 + 2,” to 99.999+% accuracy. Can I have a side of aglio e olio with that? You know what can give me the answer to “what is 2+2” with 100% accuracy, every time? A handheld calculator. And it doesn’t have to burn ten acres of rainforest to do it. “What is 2+2” is a simple question with a closed-form (shout out to me for the unexpectedly witty chosen name for my newsletter) arithmetic solution. If you perform the simple arithmetic operation known to initiates into the mathematical mysteries as “addition,” and perform it correctly, you will get the right answer, guaranteed. AI models like ChatGPT don’t work this way, though. An AI model answers a question like “what is 2+2” probabilistically, which means that it uses a bunch of unstructured data – let’s say, the text of every mathematics textbook it has pirated from JSTOR or wherever – to estimate the posterior probability distribution of the possible responses present in that training data set of mathematics textbooks (and whatever else, including the AI-generated slop that is beginning to creep over the face of the internet like kudzu – but this is a separate concern for now) to the question “what is 2+2.” To do this, it requires what we euphemistically call “computing power” – an awful lot of electricity and water. A very stupid and computationally intensive answer, one that is not even guaranteed to be right, to a very simple question.
In an earlier edition of this newsletter, I crankily reviewed part of the book Revolutionary Mathematics by Justin Joque. This book attributes the rise of datafied capitalism and the current charade of AI to the “Bayesian revolution” in statistical methodology, in particular to the metaphysical resonances of Bayesian epistemology with “the logic of capitalism.” This thesis is plainly wrong, but it’s wrong in an interesting way, for me at least. There hasn’t really been any Bayesian revolution. What we now call “Bayesian” probability was once just plain “classical” probability, until it, and with it the central concern of probability mathematics until the 20th century, was rechristened as Bayesianism, after Rev. Thomas Bayes, the remote Protestant who finally cracked the core problem of so-called “inverse probability.” The person who did this renaming was Ronald Fisher, founder of the until-recently-dominant-and-still-very-much-around 20th century school of so-called “frequentist” statistical inference.
The problem with Bayesian, or classical, probability, is this. It requires a ton of data and fiendishly long, tedious complications (so fiendish they vexed even the mighty Laplace) to calculate the posterior probability distribution of an unknown parameter. The genius of Fisher’s intervention with frequentism was not so much in making progress in this problem, the problem of inverse probability (which is to say, the problem of estimating the value of an unknown parameter from observed data), but rather in devising inferential tools and comparative experimental design principles such that statistics could be used to make inferences from observed frequencies. In fact, in a 1986 conference talk-turned-paper, one-time head of the American Statistical Association Bradley Efron asked, and answered, a question that is still relevant: “Why isn’t everyone a Bayesian?” Bayesian calculus answers a more straightforwardly intuitive question about probability; frequentist statistics have to be interpreted backwards, and don’t often make a lot of sense outside of the specific experimental settings they were designed for. Efron says it right there in the abstract: while Bayesianism is philosophically superior to frequentism, Bayesianism can’t compete with the obvious practical advantages of frequentism.
Or at least, it couldn’t. Bayesian methods are enjoying something of a renaissance as part of the AI boom. Joque, in his book, locates this renewed popularity of Bayesian methods somewhere in the epistemological resonances of Bayesian “metaphysics” (as interpreted, through Joque, by a few cherry-picked statisticians of the last century, de Finetti and Savage). Joque calls these metaphysical aspects, rather annoyingly, like the infinitesimals of calculus which he labors greatly to analogize them to, the “mysteries” interior to the workings of Bayesian probability. But there’s no mystery here. If we want to ask seriously why laborious and computationally intensive methods for statistical inference are now enjoying a popularity that they never have before, we need to look out, not in. If frequentism beat out classical “Bayesian” probability for the better part of a century due to its plug-and-play nature, and the fact that you don’t need masses of data and computers to calculate posterior probability distributions, then what about the world has changed to make extensive use of Bayesian methods possible? Good old Political Economic Developments.
In a 2019 paper in Big Data and Society, Jathan Sadowski capably demonstrates how data functions more like capital akin to financial capital and less like a commodity per se. Positing data as a form of capital, and putting it in Marx’s general formula of capital (folks, we love M-C-M’ on this here newsletter), helps explain how the intrinsic motivation to accumulate more data leads to the growth of data-handling infrastructure, which is also to say of the favorable conditions for a shift to Bayesianism and computationally intensive probabilistic modeling. This thesis helps address some salient epistemological questions too – like why we format ever-more knowledge in terms of machine-readable “data,” why it seems like data-sucking mechanisms are being integrated into ever more stuff, and what exactly the Trump administration is doing in terms of “data governance.” This last point is the subject of the Something we are working on, so please stay tuned for that.