A recurring theme on Stratechery is that the only technology analogous to the Internet’s impact on humanity is the printing press: Johannes Gutenberg’s invention in 1440 drastically reduced the marginal cost of printing books, dramatically increasing the amount of information that could be disseminated.

Of course you still had to actually write the book, and set the movable type in the printing press; this, though, meant we had the first version of the classic tech business model: the cost to create a book was fixed, but the potential revenue from printing a book — and overall profitability — was a function of how many copies you could sell. Every additional copy increased the leverage on the up-front costs of producing the book in the first place, improving the overall profitability; this, by extension, meant there were strong incentives to produce popular books.

This set off a number of changes that transformed history. Before the printing press:

• The Bible was the province of the Catholic Church; it was only available in Latin and laboriously reproduced by monks. In practice this meant that the Catholic Church was the source of religious authority throughout Europe.
• Europe didn’t have any nation-states as we think of them today; the relevant political authority was some combination of city-states and feudal lords.
• The linguistic landscape was extremely diverse: Latin was the language of the church, while larger regions might have a dominant dialect, which itself could differ from local dialects only spoken in a limited geographic area.

The printing press was a direct assault on that last point: because it still cost money to produce a book, it made sense to print books in the most dominant dialect in the region; because books were compelling it behooved people to learn to read that dominant dialect. This, over time, would mean that the dominant dialect would increase its dominance in a virtuous cycle — network effects, in other words.

Books, meanwhile, transmitted culture, building affinity between neighboring city states; it took decades and, in some cases, centuries, but over time Europe settled into a new equilibrium of distinct nation-states, with their own languages. Critical to this reorganization was point one: the printing press meant everyone could have access to the Bible, or read pamphlets challenging the Catholic Church. Martin Luther’s 95 Theses was one such example: printing presses spread the challenge to papal authority far and wide precisely because it was so incendiary — that was good for business. The Protestant Reformation that followed didn’t just have theological implications: it also provided the religious underpinnings for those distinct nation states, which legitimized their rule with their own national churches.

Of course history didn’t end there: the apotheosis of the Reformation’s influence on nation states was the United States, which set out an explicit guarantee that there would be no official government religion at all; every person was free to serve God in whatever way they pleased. This freedom was itself emblematic of what America represented in its most idealized form:¹ endless frontier and the freedom to pursue one’s God-given rights of “Life, Liberty and the pursuit of Happiness.”

Aggregation Theory

In this view the Internet is the final frontier, and not just because the American West was finally settled: on the Internet there are, or at least were, no rules, and not just in the legalistic sense; there were also no more economic rules as understood in the world of the printing press. Publishing and distribution were now zero marginal cost activities, just like consumption: you didn’t need a printing press.

The economic impact of this change hit newspapers first; from 2014’s Economic Power in the Age of Abundance:

One of the great paradoxes for newspapers today is that their financial prospects are inversely correlated to their addressable market. Even as advertising revenues have fallen off a cliff…newspapers are able to reach audiences not just in their hometowns but literally all over the world.

The problem for publishers, though, is that the free distribution provided by the Internet is not an exclusive. It’s available to every other newspaper as well. Moreover, it’s also available to publishers of any type, even bloggers like myself.

To be clear, this is absolutely a boon, particularly for readers, but also for any writer looking to have a broad impact. For your typical newspaper, though, the competitive environment is diametrically opposed to what they are used to: instead of there being a scarce amount of published material, there is an overwhelming abundance. More importantly, this shift in the competitive environment has fundamentally changed just who has economic power.

In a world defined by scarcity, those who control the scarce resources have the power to set the price for access to those resources. In the case of newspapers, the scarce resource was readers’ attention, and the purchasers were advertisers. The expected response in a well-functioning market would be for competitors to arise to offer more of whatever resource is scarce, but this was always more difficult when it came to newspapers: publishers enjoyed the dual moats of significant up-front capital costs (printing presses are expensive!) as well as a two-sided network (readers and advertisers). The result is that many newspapers enjoyed a monopoly in their area, or an oligopoly at worse.

The Internet, though, is a world of abundance, and there is a new power that matters: the ability to make sense of that abundance, to index it, to find needles in the proverbial haystack. And that power is held by Google. Thus, while the audiences advertisers crave are now hopelessly fractured amongst an effectively infinite number of publishers, the readers they seek to reach by necessity start at the same place — Google — and thus, that is where the advertising money has gone.

This is Aggregation Theory, which explained why the Internet was not just the final state of the printing press world, but in fact the start of a new order: the fact that anyone can publish didn’t mean that power was further decentralized; it actually meant that new centers of power emerged on the west coast of the United States. These powers didn’t control distribution, but rather discovery in a world marked not by scarcity but by abundance.

The economics of these Aggregators, meanwhile, were like the printing press but on steroids; everyone talks about the astronomical revenue and profits of the biggest consumer tech companies, but their costs are massive as well: in 2023 Amazon spent $537 billion, Apple $267 billion, Google $223 billion, Microsoft $127 billion, Meta $88 billion.² These costs are justified by the fact the Internet makes it possible to serve the entire world, providing unprecedented leverage on those costs, resulting in those astronomical profits.

There have always been grumblings about this state of affairs: China, famously, banned most of the American tech companies from operating in the country, not for economic reasons but rather political ones; the economic beneficiaries were China’s own Aggregators like WeChat and Baidu. The E.U., meanwhile, continues to pass ever more elaborate laws seeking to limit the Aggregators, but mostly just entrenching their position, as regulation so often does.

The reality is that Aggregators succeed because users like them; I wrote in the original formulation of Aggregation Theory:

The Internet has made distribution (of digital goods) free, neutralizing the advantage that pre-Internet distributors leveraged to integrate with suppliers. Secondly, the Internet has made transaction costs zero, making it viable for a distributor to integrate forward with end users/consumers at scale.

This has fundamentally changed the plane of competition: no longer do distributors compete based upon exclusive supplier relationships, with consumers/users an afterthought. Instead, suppliers can be commoditized leaving consumers/users as a first order priority. By extension, this means that the most important factor determining success is the user experience: the best distributors/aggregators/market-makers win by providing the best experience, which earns them the most consumers/users, which attracts the most suppliers, which enhances the user experience in a virtuous cycle.

This, more than anything, makes Aggregators politically powerful: people may complain about Google or Meta or any of the other big tech companies, but their revealed preference is that they aren’t particularly interested in finding alternatives (in part because network effects make it all but impossible for alternatives to be as attractive). And so, over the last two decades, we have drifted to a world still organized by nation states, but with a parallel political economy defined by American tech companies.

Internet 3.0: Politics

The oddity of this parallel political economy is that it has long been in the Aggregators’ interest to eschew politics; after all, their economics depends on serving everyone. This, though, doesn’t mean they haven’t had a political impact. I laid this impact out in the case of Facebook in 2016’s The Voters Decide:

Given their power over what users see Facebook could, if it chose, be the most potent political force in the world. Until, of course, said meddling was uncovered, at which point the service, having so significantly betrayed trust, would lose a substantial number of users and thus its lucrative and privileged place in advertising, leading to a plunge in market value. In short, there are no incentives for Facebook to explicitly favor any type of content beyond that which drives deeper engagement; all evidence suggests that is exactly what the service does.

Said reticence, though, creates a curious dynamic in politics in particular: there is no one dominant force when it comes to the dispersal of political information, and that includes the parties described in the previous section. Remember, in a Facebook world, information suppliers are modularized and commoditized as most people get their news from their feed. This has two implications:

• All news sources are competing on an equal footing; those controlled or bought by a party are not inherently privileged
• The likelihood any particular message will “break out” is based not on who is propagating said message but on how many users are receptive to hearing it. The power has shifted from the supply side to the demand side

This is a big problem for the parties as described in The Party Decides. Remember, in Noel and company’s description party actors care more about their policy preferences than they do voter preferences, but in an aggregated world it is voters aka users who decide which issues get traction and which don’t. And, by extension, the most successful politicians in an aggregated world are not those who serve the party but rather those who tell voters what they most want to hear.

In this view blaming Facebook explicitly for the election of Donald Trump made no sense; what is valid, though, is blaming the Internet and the way it changed incentives for the media generally: in a world of infinite competition Trump provided ratings from his fans and enemies alike; it was television (and some newspapers) that propelled him to the White House, in part because their incentives in an Aggregator-organized world were to give him ever more attention.

Trump’s election, though, drove tech companies to start considering their potential political power more overtly. I wrote last week about that post-election Google all-hands meeting mourning the results; Facebook CEO Mark Zuckerberg embarked on a nationwide listening tour, and came back and wrote about Building Global Community. To me this was a worrying sign, as I wrote in Manifestos and Monopoly:

Zuckerberg not only gives his perspective on how the world is changing — and, at least in passing, some small admission that Facebook’s focus on engagement may have driven things like filter bubbles and fake news — but for the first time explicitly commits Facebook to playing a central role in effecting that change in a manner that aligns with Zuckerberg’s personal views on the world. Zuckerberg writes:

This is a time when many of us around the world are reflecting on how we can have the most positive impact. I am reminded of my favorite saying about technology: “We always overestimate what we can do in two years, and we underestimate what we can do in ten years.” We may not have the power to create the world we want immediately, but we can all start working on the long term today. In times like these, the most important thing we at Facebook can do is develop the social infrastructure to give people the power to build a global community that works for all of us.

For the past decade, Facebook has focused on connecting friends and families. With that foundation, our next focus will be developing the social infrastructure for community — for supporting us, for keeping us safe, for informing us, for civic engagement, and for inclusion of all.

It all sounds so benign, and given Zuckerberg’s framing of the disintegration of institutions that held society together, helpful, even. And one can even argue that just as the industrial revolution shifted political power from localized fiefdoms and cities to centralized nation-states, the Internet revolution will, perhaps, require a shift in political power to global entities. That seems to be Zuckerberg’s position:

Our greatest opportunities are now global — like spreading prosperity and freedom, promoting peace and understanding, lifting people out of poverty, and accelerating science. Our greatest challenges also need global responses — like ending terrorism, fighting climate change, and preventing pandemics. Progress now requires humanity coming together not just as cities or nations, but also as a global community.

There’s just one problem: first, Zuckerberg may be wrong; it’s just as plausible to argue that the ultimate end-state of the Internet Revolution is a devolution of power to smaller more responsive self-selected entities. And, even if Zuckerberg is right, is there anyone who believes that a private company run by an unaccountable all-powerful person that tracks your every move for the purpose of selling advertising is the best possible form said global governance should take?

These concerns gradually faded as the tech companies invested billions of dollars in combatting “misinformation”, but January 6 laid the Aggregator’s power bare: first Facebook and then Twitter muzzled the sitting President, and while their decisions were understandable in the American context, Aggregators are not just American actors. I laid out the risks of those decisions in Internet 3.0 and the Beginning of (Tech) History:

Tech companies would surely argue that the context of Trump’s removal was exceptional, but when it comes to sovereignty it is not clear why U.S. domestic political considerations are India’s concern, or any other country’s. The fact that the capability exists for their own leaders to be silenced by an unreachable and unaccountable executive in San Francisco is all that matters, and it is completely understandable to think that countries will find this status quo unacceptable.

That Article argued that the first phase of the Internet was defined by technology; the second by economics (i.e. Aggregators). This new era, though, would be defined by politics:

This is why I suspect that Internet 2.0, despite its economic logic predicated on the technology undergirding the Internet, is not the end-state. When I called the current status quo The End of the Beginning, it turns out “The Beginning” I was referring to was History. The capitalization is intentional; Fukuyama wrote in the Introduction of The End of History and the Last Man:

What I suggested had come to an end was not the occurrence of events, even large and grave events, but History: that is, history understood as a single, coherent, evolutionary process, when taking into account the experience of all peoples in all times…Both Hegel and Marx believed that the evolution of human societies was not open-ended, but would end when mankind had achieved a form of society that satisfied its deepest and most fundamental longings. Both thinkers thus posited an “end of history”: for Hegel this was the liberal state, while for Marx it was a communist society. This did not mean that the natural cycle of birth, life, and death would end, that important events would no longer happen, or that newspapers reporting them would cease to be published. It meant, rather, that there would be no further progress in the development of underlying principles and institutions, because all of the really big questions had been settled.

It turns out that when it comes to Information Technology, very little is settled; after decades of developing the Internet and realizing its economic potential, the entire world is waking up to the reality that the Internet is not simply a new medium, but a new maker of reality.

Like all too many predictions that are economically worthless, I think this was directionally right but wrong in timing: the Aggregators did not lose influence because Trump was banned; AI, though, might be a different story.

The Aggregator’s AI Problem

From Axios last Friday:

Meta’s Imagine AI image generator makes the same kind of historical gaffes that caused Google to stop all generation of images of humans in its Gemini chatbot two weeks ago…AI makers are trying to counter biases and stereotyping in the data they used to train their models by turning up the “diversity” dial — but they’re over-correcting and producing problematic results…

After high-profile social media posters and news outlets fanned an outcry over images of Black men in Nazi uniforms and female popes created by Google’s Gemini AI image generator in response to generic prompts, Google was quick to take the blame. This isn’t just a Google problem, though some critics have painted the search giant as “too woke.” As late as Friday afternoon, Meta’s Imagine AI tool was generating images similar to those that Gemini created.

• Imagine does not respond to the “pope” prompt, but when asked for a group of popes, it showed Black popes.
• Many of the images of founding fathers included a diverse group.
• The prompt “a group of people in American colonial times” showed a group of Asian women.
• The prompt for “Professional American football players” produced only photos of women in football uniforms.

Meta disabled the feature before I could verify the results, or see if it, like Gemini, would flat out refuse to generate an image of a white person (while generating images of any other ethnicity). It was, though, a useful riposte to the idea that Google was unique in having a specific view of the world embedded in its model.

It is also what prompted this Article, and the extended review of tech company power. Remember that Aggregator power comes from controlling demand, and that their economic model depends on demand being universal; the ability to control demand is a function of providing a discovery mechanism for the abundance of supply. What I now appreciate, though, is that the abundance of supply also provided political cover for the Aggregators: sure, Google employees may have been distraught that Trump won, but Google still gave you results you were looking for. Facebook may have had designs on global community, but it still connected you with the people you cared about.

Generative AI flips this paradigm on its head: suddenly, there isn’t an abundance of supply, at least from the perspective of the end users; there is simply one answer. To put it another way, AI is the anti-printing press:³ it collapses all published knowledge to that single answer, and it is impossible for that single answer to make everyone happy.

This isn’t any sort of moral judgment, to be clear: plenty of people are offended by Gemini’s heavy hand; plenty of people (including many in the first camp!) would be offended if Gemini went too far in the other direction, and was perceived as not being diverse enough, or having the “wrong” opinions about whatever topic people were upset about last week (the “San Francisco Board of Supervisors” are people too!). Indeed, the entire reason why I felt the need to clarify that “this isn’t any sort of moral judgment” is because moral judgments are at stake, and no one company — or its AI — can satisfy everyone.

This does, in many respects, make the risk for the Aggregators — particularly Google — more grave: the implication of one AI never scaling to everyone is that the economic model of an Aggregator is suddenly much more precarious. On one hand, costs are going up, both in terms of the compute necessary and also to acquire data; on the other hand, the customers that disagree with the AI’s morals will be heavily incentivized to go elsewhere.

This, I would note, has always been the weakness of the Aggregator model: Aggregators’ competitive positions are entrenched by regulation, and supplier strikes have no impact because supply is commoditized; the power comes from demand, which is to say demand has the ultimate power. Users deciding to go somewhere else is the only thing that can bring an Aggregator down — or at least significantly impair their margins (timing, as ever, to be determined).

Personalized AIs

This outcome is also not inevitable. Daniel Gross, in last week’s Stratechery Interview, explained where Gemini went wrong:

Pre-training and fine-tuning a model are not distinct ideas, they’re sort of the same thing. That fine-tuning is just more the pre-training at the end. As you train models, this is something I think we believe, but we now see backed by a lot of science, the ordering of the information is extremely important. Because look, the ordering for figuring out basic things like how to properly punctuate a sentence, whatever, you could figure that out either way. But for higher sensitivity things, the aesthetic of the model, the political preferences of the model, the areas that are not totally binary, it turns out that the ordering of how you show the information matters a lot.

In my head, I always imagine it like you’re trying to draw a sheet, a very tight bed sheet over a bed, and that’s your embedding space, and you pull the bed sheet in the upper right-hand corner and the bottom left hand corner pops off, and you do that and then the top right hand corner pops off, that’s sort of what you’re doing. You’re trying to align this high dimensional space to a particular set of mathematical values, and then at some point you’re never going to have a perfect answer or a loss of zero. So, the ordering matters, and fine-tuning is traditionally more pre-training do at the end.

I think that’s originally the liberal leanings of the OpenAI ChatGPT model, came out of that. I think it was a relatively innocuous byproduct of those final data points that you show the model to, it becomes very sensitive to and those data points, it’s very easy to accidentally bias that. For example, if you have just a few words in the internal software you have where you’re giving the human graders prompts in terms of what tokens they should be writing into the model, those words can bias them, and if the graders can see the results of other graders, you have these reflexive processes. It’s like a resonant frequency and very quickly it compounds. Errors compound over time. I actually think you could end up without really thinking through it with a model that’s slightly left-leaning, a lot of the online text is slightly left-leaning.

In this view the biggest problem with these language models is actually the prompt: the part of the prompt you see is what you type, but that is augmented by a system prompt that is inserted in the model every time you ask a question. I have not extracted the Gemini prompt personally, but this person on Twitter claims to have extracted a portion: