Machine Learning As An Agent of Change
“We always overestimate the change that will occur in the next two years and underestimate the change that will occur in the next ten.” — Bill Gates, The Road Ahead
tldr: Machine learning should probably blow your mind more than it already does. What changed my mind on this was an 800 page book on the impact of the printing press—and how it changed our world even more than most of us moderns realize.
I haven’t been able to shut up about Barbara Eisenstein’s “The Printing Press As An Agent of Change” since I read it. The book persuasively argues that we underrate the printing press as the titular “agent of change” across virtually all domains of human endeavor—including science, commerce, religion, and culture. Eisenstein marshals 800 pages of examples, showing how much we take for granted is intimately tied to printing—things as small as “do schoolchildren learn alphabetical order” and “can scientists compare two books to each other”, and as big as “the Renaissance” and “is religion personal”.
I suspect machine learning is going to have a similar impact. If that’s true, we and our kids are about to live through a period which will be the most frenzied period of change humanity has seen in literal ages (even without considering climate change). In reviewing Agent of Change, I hope to show you that technology has previously deeply changed human thinking and culture—and it’s possibly going to happen again.
Important note: This is not a VC-style “therefore get on board this bandwagon or else” blog post; there’s plenty of interesting and important non-ML stuff to do in the world! I also don’t think machine super-intelligences are coming next week (or even next century). Instead, think of this post as a call to make your ML-related expectations much bigger, weirder, and yes—sometimes scarier.
No single type of change
There was no one simple “printing did X” — in different fields, it had different impacts. Eisenstein notes that “religious and scientific traditions were affected by printing in markedly different ways”, calling the shift “complex and contradictory” and impossible to “encapsulate… in any one formula.”
While we’ll go into the many examples in more detail in the rest of this essay, I want to call out from the beginning that there is no magic analytical bullet. Impacts on education, sex, religion, culture, economics—ML's impact on all will be different, complex, and often contradictory.
A recurring theme of Change is that a rapid change in quantity can quickly become a “revolutionary” change in quality. Eisenstein notes that the first attempt to catalog all printed books, by Conrad Gesner in the late sixteenth century, was “not only the first of its kind… it was also the last attempt.” There was just too much being printed! This volume was driven both by demand and by falling cost—in 1483 Italian printing was already at least 300x as cost-effective as a scribe. That change in cost and volume informs every other change that follows. We’ve seen this in tech recently, when the web lowered the cost of publishing. Yahoo! started as a hand-curated directory of the entire, much-smaller, web, but that quickly became impossible.
ML seems likely to make generation of content vastly cheaper, so orders of magnitude more content seems likely. This is already starting to happen—ML generated art is overwhelming some subreddits, video generation is coming along quickly, and we’re about to cross a line where everything is translated so that everyone on earth can read it. And it’s certainly possible that generation will become fast enough that search engines will shift from sifting knowledge to creating it on demand. What that will mean for every change below is anyone’s guess.
Eisenstein repeatedly makes the point that printed books changed how science was practiced. For example:
Like his predecessors [Tycho Brahe] had no telescope to aid him. But his observatory, unlike theirs, included a library well stocked with printed materials. (emphasis mine)
In other words, Brahe was smart—but his big edge over his predecessors was that, because of printing, he was one of the first people on earth with a library of multiple astronomical texts. That library gave him opportunities to compare, synthesize, and improve that no literally one else had ever had.
Brahe also used printing to enable early crowdsourcing, encouraging correspondents in distant places to make their own observations of eclipses, comets, and the like, and return them to him for inclusion in future editions of his works. Eisenstein shows how similar early crowdsourcing transformed biology and geography: as corrections and additions flowed in, they were collated, redistributed, and compounded at a speed never before possible.
Printing also literally changed how a scientist’s time was spent. Eisenstein quotes Laplace saying that “logarithm tables doubled the life of the astronomer”, pointing out that these printed tables allowed astronomers to spend less time on basic math and more time on observation and synthesis. Similarly, when DeepMind announced their ML-driven protein-folding work, a researcher said they spent eight years working out the structure of one protein. That will now be a simple lookup. That sort of ML-aided career multiplier is coming for most fields (for those who are still employed).
Science didn’t just change in volume, it also changed in mindset, from preservation to innovation. In Eisenstein’s telling, two printing-related factors helped drive this shift.
First, precise copying of technical information (particularly tables) had been a key part of the job of an expert scientist. This could now be delegated to the printer, who could do it both more quickly and with fewer errors. In David Hume’s words, this helped shift science from a focus on fighting "inevitable" corruption from hand-written errors, to a focus on “continually improving" work shared via accurate, replicable printed books.
Second, printing changed the nature of authority. In particular, widespread copies of books allowed widespread comparison and analysis—and the reputation of the ancients suffered as a result.
As fallible individuals, prone to human error… old giants began to look more like modern dwarfs. ‘The ancients were men like ourselves’, said Bude, ‘and often wrote about things they little understood’.
What an exciting change that must have been to be a part of! It’s this kind of printing-enabled deep mindset shift that I find most fascinating in Change. We often romanticize the Renaissance, saying that those “Renaissance Men” were somehow metaphysically different—more in love with knowledge, more creative, more… whatever.
Eisenstein devastates that romanticism, arguing that the Renaissance was not about people mystically becoming more inquisitive or creative, but was about change in technology that empowered already inquisitive and creative people. ML will almost certainly create similar empowerment, though to what ends (and with what new biases) remains to be seen.
As with many things, the printing knife was two-edged. In science, it led to greater certainty and more rapid progress. In religion, it led to greater uncertainty. Printing allowed much wider access to biblical sources in Aramaic and Greek. However, unlike science, where comparing things led to increased error-checking and greater certainty, in religion the many new translations led to more uncertainty, because the nascent scientific method could not be used to test which translation was superior.
It wasn’t just the things people were arguing about, but also how they argued. Printing certainly did not invent arguments among monks, but arguments that had been confined to private letters suddenly became public and widely disseminated. In Eisenstein’s words, “traditional forms of theological disputation had been transformed by entirely new publicity techniques”.
Printing also revolutionized who could argue about the bible. Pre-printing, the Bible could only have been read by the wealthy or within the physical walls of a religious institution. Post-printing, one could take a Bible home, pore over it by candlelight—and then decide one's priest was wrong.
Printing also changed the economics of critique. Clerics were more likely to criticize the church, since selling printed books to laypeople gave them financial independence. (This led, very directly, to regimes of censorship, with varying results; needless to say we’re already seeing attempts to impose state control on ML, with unpredictable results.) Similarly, the internet has created some successful patronage models, allowing for both amazing highs and very low lows. ML will do the same. For example, we're already seeing it inserted into the adversarial race between sales algorithms and … novel writing?
As with science and the enlightenment, Eisenstein argues that the Protestant Reformation was not caused by "something in the air"—it was caused by printing. People didn’t suddenly discover a personal, unmediated relationship with God—rather, they discovered an unmediated relationship with printed Bibles. That caused all kinds of changes in their relationship with their priests and the church—and ultimately contributed to decades of bitter warfare across Europe. I hope we don’t see the same in our newest wave of disintermediation but we probably shouldn’t rule it out. Facebook, for example, expects to double the amount of algorithm-driven recommendations in the coming year, making ML ever more critical to politics and religion.
Expanding commerce and form
The shift to printing from scribal copying also made commerce a key player—with all the good and bad that entails.
For example, printing impacted the debate on whether or not the church should use Latin; vernacular languages had more readers, so vernacular translations of the Bible were more profitable—and so profit-driven printers were not neutral when asked "should the Bible be restricted to Latin?" Similarly, commercial experimentation drove innovation in the written form. Footnotes, tables of contents, better figures, and cross references—these were all previously difficult because of reproducibility or non-standard page numberings, and publishers loved to innovate in them.
For example, before typesetting, if you wanted a table of contents or glossaries, you had to make a new one for each copy of the book, because hand-written pages had the same content on different pages. Similarly, my favorite factoid from the book: “A 1604 edition of an English dictionary notes … that ‘the reader must learne the alphabet, to wit: the order of the letters as they stand.’” In other words, before printing, there was not much reason to learn alphabetical order, because there wasn’t enough information to make it worth alphabetizing! Printers seized on these technologically-enabled innovations in the form, and used them to sell more books—then enabling more innovation.
We have no idea, of course, what sort of new innovations of this sort will come out of machine learning. But we can be fairly certain that some will become so basic we’ll teach them to the youngest children—and commerce will egg it along.
Another benefit of printing was standardizing knowledge made it easier to discuss and improve knowledge. As Eisenstein puts it, “printing made it possible … to publish hundreds of copies that were alike and yet might be scattered everywhere”. This meant “scholars in different regions [could] correspond with each other about the same citation” and share errata (another new technology!) While this did not invent the “Republic of Letters”, it did increase its impact by allowing scholars to refer to specific page numbers in their correspondence.
Similar multiplication of impact happened in politics:
[F]ear of disapproval, a sense of isolation, the force of local community sanctions, the habit of respectful submission to traditional authority - all might be weakened among many obscure provincial book readers by recognition that their innermost convictions were shared by fashionable and famous men of letters.
A reader in 2022 will of course find echoes (for better and for worse) in the sharing of links and message boards amongst internet scholars, weirdos, and friends. It seems likely that machine learning will have an inverse, but similarly large impact, by leading to ultra-personalized information that complements (or attacks) share knowledge. This could privilege the aggregators of the world, or those who benefit from dissensus—or most likely both and much in-between.
Printing changed the nature of what could be “known”.
Until the end of the fifteenth century, it was not always easy to decide just ‘what a statute really was’ … In ‘Englishing and printing’ the ‘Great Boke of Statutes 1530-1533’ John Rastell … was … offering a systematic review of parliamentary history - the first many readers had ever seen.
In other words, before printing, we literally had a hard time knowing what the law was. Search has already changed how judges cite, and it seems likely that our newfound ability to process and make sense of vast data sets, like “the law”, will change that further.
Similarly, Europeans had been rediscovering Greco-Roman texts for a millenia, but
the book-hunting of quattrocento humanists was of immense significance … because – unlike texts which had been retrieved previously only to be lost again - the texts which [they] rescued were duplicated in print and permanently secured.
Similarly, machine learning will absolutely change what we know—I’ve already mentioned protein folding, and in the past year ML has made advances in applied math and other areas of science as well. Its ability to comprehend and experiment on vast pools of information will absolutely change the nature of knowledge, and how we know it.
What Renaissance are we going to see in the coming 1-2 generations? I have no idea. But after reading Eisenstein, I’m pretty sure we’re going to see something of comparable scope and impact to the human condition. How we prepare for it is up to us.