Trillion-dollar PDFs

Greetings from Read Max HQ! In this week's column — now arriving in inboxes on Sunday mornings — notes toward an exploration of the "trillion-dollar PDF" as a form or genre.

On this week's Hard Fork, I talked with my boon companion Erin Griffith about Anthropic's S-1, before we were joined by Mike Isaac and Jasmine Sun for a broad discussion of the role of (and effects on) tech of the coming midterms, Microsoft's layoffs, cybersecurity's Black October, the California Billionaire Tax, and the sleeper stories we're tracking this week. Listen above, or on the podcast platform of your choice!

And a reminder: Read Max is a subscription enterprise. If you'd like access to our archives — including nearly six years' worth of movie and book recommendations — all you need is $5 (monthly) or $50 (yearly) — about the price of a beer or espresso drink, depending on location. You can subscribe here.

On Thursday a new single-serving website appeared: "Commanding Heights," by the investors Will Manidis and Joe Roberts. Manidis, who might be loosely and not too dismissively described as an AI thoughtfluencer — a man-about-Twitter with an influential follower list and a long CV of panel appearances — announced in a tweet: "The benefits of abundant intelligence will accrue far outside of San Fransisco. Spent the last year writing about capital, sovereignty, and the coming reordering of economic power."

The page opens with a tastefully typeset title, some acknowledgments, and a lengthy table of contents, like so:

The tech industry loves itself a long post like this. The AI sector, in particular, sets a lot of store by long posts of various stripes ("throughout Slack, we'll have giant essay-length debates with people," one Anthropic employee has said) — a cultural phenomenon that I suspect has its roots in forum debates as much as in machine-learning research's nominal status as an academic discipline. In a world of short-form social media, depleted attention spans, and automated text production, an impressively long piece of writing can serve as a signal of seriousness, thoughtfulness, intelligence, and hard work. (Of course, it can also signal windbaggery, injudiciousness, and lack of a knowledgeable editor or prudent friend.)

Perhaps it's even more straightforward than that. (This is the industry that invented blogging, after all.) Business leaders have always loved memos; bankers have always loved decks; computer nerds have always loved science fiction: What if you had a piece of writing that could be all three?

But "Commanding Heights" is something more than an investor's speculative fiction or a professionalized LessWrong post — though what it is, precisely, besides extremely long, is hard to say. What unfolds across its 50,000 words (a set of investment recommendations wrapped in a narrative description of a future of global precarity and widening inequality, dominated not by the big "labs" but by AI-infrastructure rentiers) is part sci-fi story, part footnoted economic forecast, part investor proposal, part manifesto, written in an aphoristic, portentous prose. (One person on Twitter described the style as "chuuni," which introduced me to a useful new word.) I have to admit I haven't actually read the whole thing. But I'm much less interested in the contents of "Commanding Heights" than I am in the emerging form it mimics. It is, above all else, the clearest example I've ever seen of someone specifically trying to write a "billion-dollar PDF."

First named by the investor Jeremy Giffon, the "billion-dollar PDF" is a phrase used to describe a piece of writing that comes to define — among business and investment leaders in Silicon Valley — an entire technological framework or investment strategy, in the process becoming the kind of document eager young venture investors and ambitious Twitter posters will describe as "life-changing." Famous examples (many more collected here) include investor Marc Andreessen's 2011 Wall Street Journal op-ed "Why Software Is Eating the World," which succinctly defined his firm's investment thesis; anonymous cryptographer Satoshi Nakamoto's 2008 whitepaper proposing a "peer-to-peer electronic cash system" called bitcoin; and Ashish Vaswani et al.'s 2017 research paper "Attention Is All You Need," which laid out the architecture that would come to power large language models.

As it has become clear that a particularly rousing futurological text can quickly earn its author attention (and capital) from some very wealthy and important people, the "billion-dollar PDF" has gone from a retroactive classification to a kind of genre in which one can write. "Commanding Heights," for example, is not an op-ed or memo or whitepaper that, thanks to its vision and clarity, has taken on totemic importance to a field or industry. It was fairly straightforwardly written to be regarded as a billion-dollar PDF — hailed and revered from the very start as a document of special foresight and intelligence. (In this it was fairly successful.) Its most obvious predecessor in this respect is former OpenAI employee Leopold Aschenbrenner's widely circulated "Situational Awareness," a lengthy and chart-laden narrative economic forecast that is, literally, a billion-dollar PDF, insofar as Aschenbrenner was able to use its success as a launching pad to start a hedge fund with the same name.

We might call these latter-day, lovingly crafted, excessively long, tastefully designed essays-cum-manifestoes-cum-investor-magnets-cum-Twitter-events "trillion-dollar PDFs," one of many things Manidis and Roberts' essay was hailed as after its publication. The trillion-dollar PDF distinguishes itself from the billion-dollar PDF by its straining effort toward self-importance, and it's not confined to investment strategists seeking assets to manage. OpenAI interpretability researcher Nick Cammarata's 30,000-word essay "Neural Networks Are Already Clean," a gorgeously designed piece of writing, claims to articulate a breakthrough in machine-learning interpretability research, and was published in October to absurd levels of fanfare on Twitter. Cammarata is, more than most would-be BDPDF authors, a good writer, and his descriptions of how individual neurons work in a neural network are clear and insightful. But it's also hard not to wonder why I'm reading about a supposed breakthrough at length on an extremely tasteful single-serving website with pretty animated GIFs and not as a structured paper on arXiv, at the very least.

The answer to this question, of course, is Twitter, or "the kind of discourse community cultivated when Twitter is the main hub and entire context for your work." We are all familiar with the way Twitter has made it easier than ever to circulate a link so it can find an audience, not to mention easier than ever to pretend you've just read and been astounded by a particular paper. More importantly, Twitter also effortlessly fosters the kind of social pressure necessary to make a piece of writing seem like an unquestionable masterpiece, even as it becomes an accessory deployed to cultivate the Twitter persona of every person tweeting it. Journalists who lived through the 2010s will nod as they read this because they remember what it was like to dominate and define Twitter in this way, to make it host and captive audience to your rivalries and pet issues and insider jokes, but also to let its prerogatives redefine and rework your relationship to your job and career. Now that the AI industry is Twitter's dominant subject and cast of characters, it's undergoing the same transformation. The trillion-dollar PDF is its #longreads era.