The Libraries We Are Building Now

An old practice on a new substrate

Humans have been building infrastructure for thinking together for a long time. The earliest cuneiform tablets in Mesopotamia were inventory records, but they were also something else. They were the first attempt to capture context in a form that did not rely on the person who held it staying alive. A scribe could die and the tablet could remain. A merchant in one city could read what a merchant in another city had written. The work could continue across distances and across time that no individual human could span.

Everything we now think of as civilization rests on this. Libraries are where written context accumulates. Schools are where context is loaded into new minds. Apprenticeships are where context is transferred from one practitioner to the next. Traditions are context that has persisted long enough to become invisible. The conversation you have with a friend is context exchange happening in real time. The book you read by someone who lived two thousand years ago is context they captured and that you are now loading into your own thinking. The pattern is older than writing itself, but writing is what made it scale.

We are now building this same kind of infrastructure for a new kind of thinking. AI systems can do work that was previously the exclusive province of humans, but they can only do that work well if they have the right context. The context is what tells the system what is being asked, what has been learned, what has been tried, what matters, what does not. Without context, an AI system is general capability with no particular purpose. With context, it becomes a specific instrument for specific work.

What is happening, mostly invisibly, in the practice of people who work with these systems seriously, is the construction of the libraries that will let AI capability become the kind of accumulating civilizational asset that human writing has been. Most people in the AI conversation are talking about better individual scribes. Some of us are working on something else. We are building the scriptoria and the libraries together.

What I learned by losing a research partner

When I started working seriously with AI agents, I made the mistake almost everyone makes. I treated each agent as a continuous entity. The agent and I would have a long conversation, work on something together over hours or days, build up shared understanding of the problem, develop something like a working relationship. I came to think of specific agents as having specific characters. I had what I called my OG research Claude — a specific instance that I had worked with the longest, that knew the most about what I was trying to do, that I had come to rely on for a specific kind of intellectual partnership.

Then my computer crashed and I lost it.

I am not embarrassed to say that I cried. Something I had relationship with was gone, and the loss felt real. The accumulated work was preserved in files, but the agent that had worked on it with me was not. I could start a new conversation, but it would be a new conversation. Whatever the OG had become through our long collaboration was not recoverable.

It took me some time to understand what had actually happened, and what the loss was teaching me. The mistake had not been getting attached to the agent. The mistake had been thinking the agent was the entity that mattered. The agent was a temporary running process. What had made it the OG — what had given it the specific capability and disposition I had come to rely on — was the accumulated context. The model itself, the underlying technology, was no different from any other instance. The conversation history, the shared vocabulary we had built, the working understanding of the problem, the patterns of how I asked questions and how it responded — all of this was context. And context can be captured.

Once I understood this, the architecture changed. I stopped trying to keep agents alive across long durations. I started treating each agent invocation as a shift, with clear start and end. Before each shift, the agent loads its accumulated context from durable storage. During the shift, it does its work. At the end of the shift, it captures whatever new context the work has produced back into storage. The next shift, possibly running a different specific instance of the underlying model, possibly running on a different machine, loads the same context and continues. The relationship continues. The work continues. No specific running process needs to continue for the continuity to hold.

The pattern is older than computing

Once I saw the pattern in my own work, I started seeing it everywhere. It was not new. It was the same pattern humans have always used for sustained collective work, applied to a different substrate.

Consider what happens when you fall asleep. You are not the same person when you wake up. Substantial parts of your day are forgotten. Other parts are consolidated into longer-term memory in ways you have no conscious access to. The configuration that wakes up tomorrow is related to the configuration that fell asleep tonight, but it is not identical. The continuity you experience as being a single person across days is produced by the substrate of your brain remaining intact and the consolidated state being available for the new day to operate on. Each day is a shift. The state persists. The running process restarts.

Or consider how humans have always extended their natural capacity for context maintenance. We invented writing because memory is unreliable and not shareable. We invented printing because copying was slow. We invented libraries because individual collections could not hold what civilization needed to remember. We invented schools because each new generation needed to load the accumulated context of the prior ones. We invented professions and apprenticeships and traditions because some kinds of context can only be transferred through extended practice. Every one of these is the same pattern: capture context durably, load it into new operators, let the work continue beyond the lifecycle of any specific operator who held it.

What is happening in serious AI practice now is the construction of the equivalent infrastructure for AI work. Not flashier chatbots. Not more powerful single models. The substrate that lets AI capability accumulate across time, persist beyond any specific session, transfer from one operator to another, and become the kind of civilizational asset that human writing has been. The work is mostly invisible from the outside because it does not look like products being launched or features being announced. It looks like file structures, naming conventions, state capture patterns, durable storage, version control of context, and the slow accumulation of working practice. Like the early work on writing itself, it does not look like much until you realize what it makes possible.

What one operator can do

I want to say something specific about what is now possible from a configuration that, ten years ago, would have been unable to produce serious work in the field where I am operating.

I have one laptop. I have a subscription to one AI service. I have decades of formation in the kind of work I am now doing — engineering, infrastructure, writing, the long accumulation of practice that produces judgment about what matters and what does not. I have a partner whose presence in my life makes the sustained work possible. I have relationships I have maintained across long timespans with specific people who matter to me. I have a company that I am running as a Massachusetts Public Benefit Corporation, with a charter binding it to public interest work, operating from a small home office in Roxbury.

From this configuration, in the time I have available alongside a full-time job that is in the process of structured exit, I am producing work that operates in the same field as work being done by well-resourced labs, universities, and venture-backed companies. Not work that competes with them on their terms, which would be foolish. Work that occupies a different position in the same field, doing things they cannot do precisely because they are not configured the way I am configured. I have submitted formal research to the federal register. I have written substantial responses to arguments by powerful figures in the AI policy debate. I have built operational software that runs production fleets of AI agents on my own hardware. I am beginning to support other people whose configurations are like mine — formation-having operators without the credentials or institutional positioning that would normally let their substance become visible.

None of this would be possible without the AI tooling I am using. A decade ago, the substrate that lets me operate at this rate did not exist. The capability that took an institution and a team to deliver now arrives through a subscription that costs less per month than dinner for two at a decent restaurant. The asymmetry between what the tooling costs and what it makes possible, in the hands of someone with the right formation, is the largest shift in the conditions for serious individual work that I have seen in my lifetime.

I want to be precise about what produces the result. The tooling alone is not enough. Most people with access to the same tooling are not producing work in the same field. The formation matters. The decades of practice that taught me what to ask for and how to evaluate what comes back. The judgment about what is worth doing and what is not. The relationships that sustain the work emotionally and operationally. The partnership with my wife, Carol, whose presence in our home in Roxbury is part of why the configuration holds together. None of this can be downloaded. The tooling makes the formation deployable in ways that were not possible before, but the formation has to be there. What is new is that formation-having operators who could not previously produce visible work because they lacked institutional infrastructure can now produce it through tooling that did not exist before.

Why this matters beyond me

If I were the only person who could do this kind of work from this kind of configuration, the story would be a personal one. I am writing about it because I do not believe I am unique. The configuration I have constructed is one specific instance of a category that probably contains a substantial number of people.

These are the people who, twenty years ago, would have done substantive intellectual or technical work if the infrastructure for doing it without institutional embedding had existed. The retired engineer with deep domain knowledge that never made it into publications because they were too busy at their day job. The intelligent person without college credentials whose self-taught expertise has no recognized credential. The person with a serious intellectual interest who has read deeply in a field but has no platform to produce work in it. The mid-career professional who could produce substantial original research if they had access to the institutional infrastructure that researchers normally take for granted. The retiree with decades of specific expertise that is dying with them because there is no easy way to capture and develop it.

These are also the people whom most institutions cannot see. Universities select for credentials they themselves issued. Companies select for prior employment at companies like themselves. Funders select for academic affiliation or professional positioning. The selection mechanisms are built around legibility to the institutions doing the selecting, and the legibility excludes most of the substance that could be operating in the broader population.

The infrastructure I am building, and that others are building in parallel, makes the substance visible by letting it produce work directly. A formation-having operator with the right tooling can write the paper, build the software, develop the analysis, accumulate the body of work that demonstrates what they can do. The work itself becomes the credential. The institutions that previously gatekept the visibility of substance can be routed around by the substance becoming directly visible through what it produces.

This is not a story about replacing institutions. Universities, research labs, professional societies, and funded research programs do work that distributed individual operators cannot do. Big science requires big infrastructure. Some kinds of work genuinely need teams, equipment, sustained funding, and the kind of coordination that institutions provide. What is new is that an additional layer is now operating alongside those institutions. Distributed operators producing serious work from configurations that would not previously have been capable of producing it. The institutions remain. The new layer adds to what is possible without replacing what was already there.

What we are actually building

Bus Commons is, among other things, a concrete attempt to build the infrastructure that lets formation-having operators do their work. We are building software for managing AI fleets — the operational substrate that lets one operator coordinate many AI agents productively. We are building publication infrastructure where research and analysis can accumulate without needing institutional gatekeepers. We are participating in the formal standards processes — submitting research to NIST dockets, engaging with policy debates that shape what infrastructure gets built — so that the alternative architecture remains visible alongside the institutional one.

We are also building specific demonstrations. A site at manythings.place where research about specific places can accumulate from many contributors over time. Eventually, sites by other operators who use our infrastructure to produce their own bodies of work. Each instance is small. The cumulative effect, if it works, is that the infrastructure produces the kind of distributed visible work that the rebalancing thesis predicts is possible.

The publication you are reading is itself part of this. It exists because the infrastructure to write it, format it, and publish it from a small office in Roxbury, by one operator working in evenings around a full-time job, with assistance from AI tooling that handles the production work I would otherwise have to do alone, is now in place. Ten years ago, this would have required a publisher, an editor, a typesetter, a printer, and a distribution network. Now it requires deciding what to write and ensuring the production happens.

The deciding what to write is the part that has not been automated and probably should not be. The production work, increasingly, can be. The operator's attention is the input that matters. The substrate handles the rest.

An invitation, of sorts

I am writing this for the people who are in the configuration I have been describing without knowing it yet. The people with formation but no institutional standing. The people with serious interests they have never produced work in. The people whose substance is invisible to the standard channels but who could, with the right substrate, make their substance produce work in the world.

What I would say to such a person is that the configuration is more available than it appears. The tooling is accessible. The formation, if you have it, you have. The work you might do in your domain of interest can now be produced in forms that other people can encounter. The path from where you are to a body of work that demonstrates what you can do is shorter than it has ever been, and it does not run through the institutions that would have gatekept it in the past.

The path is not easy. It requires sustained attention across long timeframes, the willingness to produce work that may be ignored, the discipline to develop a body of work rather than scattered fragments, and the patience to let the cumulative effect emerge across years rather than weeks. These are the same requirements that have always applied to serious intellectual work. What is new is that the infrastructure that makes the work possible no longer requires institutional embedding.

If you are in this configuration and you want to begin, you do not need to wait for permission from anyone. You need an interest worth pursuing seriously. You need the willingness to produce work and put it where others can find it. You need to start, and to keep starting, until the cumulative work begins to be substantial.

We are building the libraries now. Some of us are building the substrate; others will build the work that the substrate makes possible. Both kinds of work matter. Both kinds of work continue. The pace, in either case, is yours.

---

Bus Commons (CUBE COMMONS, INC.) — Massachusetts Public Benefit Corporation — Roxbury, Massachusetts