Alexandria: In Defense of Libraries

Imagine you’re an American college student in the early 1970s who has elected to spend four years studying history. The world, as far as you can tell, is ablaze. Your childhood best friend was drafted into the military and killed thousands of miles from home for a cause he knew nothing about, your earliest memories are overshadowed by the assassination of the president, your hometown has long since splintered into factions with opposing beliefs on integration and civil rights, and you’re at the onset of the first economic recession in recent memory. You’re not certain of what’s happened in recent years because of how many different narratives exist, and everyone seems to disagree about what to do. You’re interested in history because you want to understand the conditions that led to the world you know, and you would like to develop your political philosophy based on what you learn. Your intentions are to seek the truth.

How do you do this? Who determines what’s right? Where do you even look?

In the 1970s, you might turn to your professor, who curates the texts you read to develop your initial corpus of knowledge. You might turn to your classmates, with whom you engage in debate until you’ve mapped every possible argument. You might turn to the archives, where you ask a librarian to help you locate hundreds of first-hand documents that allow you to step into the contemporary discourse directly. By the end of your education, you have a grasp of what the truth might be from the lens of a student who has access to your local context. You balance your trust in others’ earned authority within the collective body of knowledge with your own individual reasoning, and you arrive at some version of the truth through the frictions required to reconcile all the perspectives you’ve come across. In this world, the collective body of knowledge is gatekept by your local context alone. Personal viewpoints and physical availability of source material are the primary determinants of your priors and the premises with which you engage. Nefarious actors can attempt to brainwash you through local media or newspapers or book recommendations or legal censorship, but the bulk of your encounter with this material is sovereign.

Fast forward forty years, and the first thing you do is search the web. You continue to engage with your professor and classmates and archives, but you can also see what others think online, and you can access a whole world of ranked documents in seconds at any hour of the day. Your learning comes from piecing together the different bits of information you find across different websites, with increased scale and deeper research enabled by a search algorithm, and synthesizing them. You now have access to a global context that pushes you to manage different models for the truth in parallel, and you can reason as deeply as you’d like about which one makes sense in a given context. You’re still the principal synthesizer in this world, and the outputs you consume are typically developed by those who undergo the same process. However, you are now subject to the loss functions embedded in each of the platforms you use to encounter information organically, which are designed opaquely by powerful companies who might have specific interests. The questions you become interested in on your own time and the information ranked for your consumption are shaped by these entities, and you implicitly give them total authority over your inputs. In this world, the collective body of knowledge is more accessible in theory but gatekept in practice by optimization processes that know everything about you. Nefarious actors can credibly control your viewpoint by determining what you see in the first place over the course of many years.

What about twenty years after that? In the year 2033, what does the pursuit of knowledge actually look like? With AI trained on the entire corpus of human knowledge and fine-tuned to provide a polished answer that satisfies the broadest base of users? Perhaps you ask an AI your question, and you receive an answer that you trust is mostly correct, even if it’s only presenting one version of reality among many, and it’s difficult for you to disaggregate the response in any meaningful way because there’s not much you know that an AI doesn’t. You trust the AI to pull the truth from the sources it gathers, and you’ve spent all your life learning this way, so you don’t know when to press further. You might even ask the AI to help you form your core political philosophy, or to help you decide what to do in your career, or what to believe in. Importantly, the AI presents “the truth” in any scenario as a single response without any discourse required to go from one conclusion to another. This is a categorical change from the previous technologies we’ve seen. If the AI can synthesize information for you, and it can do so as well as an expert human being who’s undergone years of academic or professional training, we begin to live in a world that makes it difficult for an unacquainted reader to do the hard work of individual reasoning. Gradually, it becomes nearly impossible for the reader to become acquainted in any way with the truth beyond the perspective offered to him by the AI and its model trainers. This is a world where the collective body of knowledge is fully accessible to the human through the AI, but the gatekeeper is a maximally optimized computer that knows everything about you and is trained to give you answers that you’ll be satisfied with the first time. Nefarious actors can control everything the individual thinks and believes because there is no alternative set of inputs for him, and he feels good about it. The technology and its makers have total authority.

In each of these periods, the state of your political will is a direct consequence of your epistemic inputs and your ability to reason across them. That is, your political autonomy is shaped by both your access to the collective body of knowledge and your right to engage critically with it, including knowing where it comes from. But, as we see with the advance of technology across the three periods, the more access to the collective body of knowledge the individual has in theory, the more he tends to lean on comparably scalable authority to help him navigate it, starting with our social technologies and institutions, and then our algorithms, and eventually an AI trained on all available knowledge. Consequently, the rise of highly capable technology also gives rise to centrally powerful entities and their interests, amplifying the ability for those groups to control public perspective by capturing the very inputs and learning processes that people have access to.

If we take this analysis a step further and argue that political will tends to decline with the advancement of technology, then the idea of epistemic sovereignty actually becomes critical in the age of unprecedentedly capable AI, so much so that we should treat it as a right.

How do we protect this right? How do we retain our epistemic sovereignty in an age when our access to knowledge seems to be so frictionless and total? I propose that we revisit the core question of what it means for a human being to reason individually at scale in a world where our access to the collective body of knowledge increases dramatically by returning to a tried and tested model: the library.

The library allows for the best of both worlds: it gathers the sum of human knowledge in one place and allows for competing perspectives and discourse and forms of authority, but it requires the individual to engage with that knowledge directly. Importantly, though you might subscribe to specific schools of thought or scholars or ideologies, you have the ability to read other books, and you’re able to engage with material at a granular level, breaking free of the all-or-nothing truthfulness of an AI response. You might have a professor or a lecturer or an authority on different bodies of work who curates the right material for you to read in order to arrive at the frontier, but you retain the right to engage with that material or its alternatives as you wish.

I would like to take this model a step further and propose a novel claim library system called Alexandria that offers a human-centered, technology-scale solution to this epistemic capture by allowing human beings to define their priors, or the inputs that inform their reasoning process. This is the core idea underlying the solution: individuals who engage in the process of defining their priors are able to reason more critically at scale when encountering new information. This solution assumes that we should not impose an infrastructure-level belief about which truths are correct, but rather that we should enable humans to reason how they do anyway — through the comparison of new information to their priors, which come from pluralistic, socially entangled perceptions of reality.

Using Alexandria, human beings can build “Libraries” of claims that they trust, potentially representing different worldviews or truth-telling frameworks, and they can share and debate and modify the contents of these Libraries with others, the same way they might open-source a code repository. The Libraries to which the user subscribes are then used as reference points when he encounters new information, comparing the content he consumes at its most fundamental level to the claims he defines as his priors. Alexandria does not tell him whether the overall content is correct or truthful in monolithic fashion, but rather fragments the input in a logically aware way into its core, falsifiable subclaims and shows him how each subclaim stacks against the ground truths in his different Libraries.

This model does two things: (1) it allows the user to engage critically with a high volume of information from different vantage points that might have different implications on the truth; and (2) it allows us to develop several different corpora of human-derived truths that are openly debatable through individual contributions as well as democratic methods.

The exercise of developing ground-truth Libraries in and of itself contributes to the individual reasoning ability of the human, and more importantly, the ability to battle-test a given Library through crowd feedback is a digital encoding of the deeply human and long-lived practice of having social discourse contribute to the collective body of knowledge throughout history. In one turn, we combine the individual’s ability to reason with the collective’s ability to aggregate knowledge, mirroring the core exchange that’s driven so much of our civilizational history.

My goal with Alexandria is to model a human-centered approach to technology that preserves human autonomy while taking advantage of everything our advances have to offer. I’m releasing a beta version at alexandrialibraries.com. I would love for you to check it out.