Constellations of Borrowed Light

The Inheritance

Human knowledge is never contained in one person. It grows from the relationships we create between each other and the world, and still it is never complete.

• Paul Kalanithi, When Breath Becomes Air

The knowledge to save your life can exist and still fail to reach you.

When I was six, I saw the same doctors twelve times with the same symptoms. The diagnostic triad for my brain tumor had been in the literature for decades. The knowledge existed. The system couldn’t carry it. I survived, but the failure wasn’t medical — it was structural. I watched the same thing happen to others in my family — an uncle, my grandmother. My first heroes were the doctors who saved me. The more I learned, the more I realized: these people were brilliant and tireless, doing the best they could inside a system designed for a different era. Drugs that work sitting in papers that never reach clinicians. Materials breakthroughs trapped in notebooks that engineers will never read. The people aren’t failing. The system between discovery and reality was built for a world that no longer exists.

The real gap in modern science is between knowing and arriving. We have not even delivered on most of what science already knows. AI is about to transform scientific discovery — for the first time, the compiler that Vannevar Bush dreamed about in 1945 actually exists. But AI is arriving into that same outdated system, and if we don’t build the ecosystem first, it amplifies the defects at machine speed. The greatest opportunity of our generation is rebuilding the ecosystem of science itself: shared state, execution infrastructure, and an open network that lets knowledge compound instead of scatter — and that opens science to everyone who wants to contribute.

Look up at the night sky and you are seeing ghosts. Some of those stars collapsed millions of years ago. Their light is still traveling. Knowledge moves the same way: across time, across distance, across hands that never meet. Every doctor inherits from patients she never saw. Every scientist inherits from experiments she never ran. We live by borrowed light, or fail to.

We can build the systems that let the light travel.

The Pattern

My case was not unusual. That is precisely what makes it typical.

Only a fraction of pediatric brain tumors are diagnosed within the first month of symptoms. Many children are misdiagnosed on their first visit. Even simple examinations that could surface the pattern are often not performed. The literature is not empty. The problem is that the knowledge does not reliably move from where it is stored to where it is needed.

Medicine is only one instance of a larger pattern.

For more than twenty years, Alzheimer’s research converged around a single hypothesis: amyloid plaques drive the disease, so clearing the plaques should stop the decline. Trial after trial failed to produce the clinical outcome that would justify the confidence. Each Phase III failure meant years of enrollment unwound, patients and families who had organized their lives around a hypothesis that quietly died in an interim analysis. Alternative targets remained underexplored. The field did not lack papers. It lacked a record that could force failed trials, failed replications, and changing confidence back into the live state of the question. *

The knowledge existed, and so did the corrections. Nothing carried them.

Now imagine something different. A researcher entering the field in 2015 opens the compiled state of amyloid therapeutics and sees, in one view, which targets have been tested, which have failed, under what conditions, and where the frontier is genuinely thin. Failed paths are inherited as structured knowledge. The next team starts from the real edge instead of the published consensus. That world is buildable. The tools to build it exist now. What follows is an argument for how.

The same thing happens with failed experiments. At Haverford College, a materials science lab had years of failed synthesis attempts for vanadium selenites sitting in notebooks — data that would normally be discarded. When they finally fed those failed reactions into a machine learning model, it predicted outcomes better than human intuition could. The failures contained real knowledge. It was just trapped in notebooks that no one else would ever read. *

This is the normal condition of science. On average, it takes seventeen years for research evidence to reach clinical practice. Only 14% of original discoveries are ever integrated into the work of the people who could use them. The field pays for the same lesson twice, and most of the lessons are never shared at all. *

Most scientific knowledge never becomes a paper at all. The published literature is only the visible fraction: what survived selection and format. Negative results vanish. Tacit judgment vanishes. Retractions often fail to reach the work built on top of them. We treat this as friction around the edges of science. Much of the time it is closer to the center. *

Science still treats the paper as its unit of record. That is the mistake. A paper is a rendering of science: compressed into narrative, stripped of much of its operational context, and frozen at the moment of publication. The paper was never the atom of science; it is a lossy rendering.

A system that cannot carry failed paths, contested claims, lower-layer context, and revision is a system that loses contact with its own learning. Knowledge is hard won, and too much of it is won and then allowed to scatter.

Something has to change about the medium itself.

The Substrate

For most of history, the transmission problem was terrible but stable. A doctor in 1950 could reasonably expect that much of what she learned in training would remain current for years. A scientist might spend a career in one field without the literature outrunning the pace at which a person could still read and absorb it.

That stability is gone. The literature now doubles faster than any person can read it, intelligence is getting cheaper by the month, and the economics of structure have changed. *

For eighty years, the dream has been clear. Vannevar Bush imagined associative trails through knowledge in 1945. Many efforts since then tried to make science more structured, linked, and machine-readable. Most failed for the same reason: they required scientists to re-author the world into a new schema before they delivered enough value in return. The vision is old. The incentives were wrong. The compiler is new.

Language models can now do some of the restructuring work. They can extract candidate findings from prose, draft links between ideas, surface contradictions, and turn a pile of documents into the beginnings of a structured map. For the first time, the compiler that Bush imagined actually exists, and it is getting better fast.

AI is entering science fast, and it will transform it. We have already seen what happens when AI meets a well-structured scientific problem: AlphaFold predicted the three-dimensional structure of nearly every known protein, solving in months a challenge that structural biologists had worked on for fifty years. Two hundred million structures, released freely, overnight making accessible what would have taken the entire field another century of crystallography. That was one model applied to one problem with clean, structured data beneath it. *

Now imagine that across every field. AI agents designing experiments, analyzing results, generating hypotheses, managing entire research workflows — not for one well-curated dataset, but across all of science. The major discoveries ahead of us will be made with AI. But right now, these systems are arriving into a medium that was never built to hold scientific state. They can search documents. They cannot inherit a field’s structured knowledge, reason over its contested claims, or write their findings back into a shared record that other agents and researchers can build on. AlphaFold worked because protein sequences were already structured data. Most of science is not. That is the gap the substrate has to close.

Software offers the clearest contrast. Code compounds because it lives in a medium designed for compounding. Git gave software memory — a way for code to be inherited across distance, across hands that never meet. Compilers made source executable. GitHub and package ecosystems made that state networked and reusable. Agents arrived on top of that substrate. They did not create it.

AI writes code because this infrastructure exists. It is operating in a world with versioned state, executable runtimes, and shared networks. It can inherit from prior work. It can propose a change against a system that knows how to remember, test, merge, and distribute it. *

Science has no Git. It has no GitHub. It jumped straight to agents, and the ecosystem beneath them is still sand. The light has no medium to carry it.

Whoever shapes the knowledge substrate shapes what intelligence believes is true. If the substrate defaults to closed, proprietary document silos with AI wrappers on top, that is what science inherits — and unwinding it later is much harder than building it right the first time.

Science does not lack intelligence so much as a medium that can hold intelligence in public.

Science needs what software already has: a shared state layer, an execution runtime, and a network that lets both compound.

The state layer is the record of what a field currently knows in machine-operable form: findings, evidence, provenance, revision, dissent, and the calibration and lineage beneath them. Not documents that mention those things. State that preserves them directly.

The runtime is the layer that forces ideas back into contact with reality: protocols, experiments, trials, instruments, and the systems that capture what actually happened.

The network is what makes the first two layers cumulative rather than local: shared identifiers, open protocols, interoperable formats, and a grammar of connection broad enough for one institution’s work to become another’s starting point.

Without state, science cannot remember properly.
Without runtime, it cannot learn properly.
Without network, it cannot compound properly.

The missing architecture is epistemic, not merely informational.

The goal is to build a medium in which what is known remains tied to how it was measured, where it came from, how it changed, and what it should change next, rather than making science more legible while leaving it detached from reality.

An operating system for science would begin to provide exactly that: not a single product or interface, but a substrate — the medium that lets the light arrive instead of scattering.

Everything else people want from AI in science sits on top of that. The next question is what the first layer looks like.

The Constellation

The first visible surface of that substrate is the constellation.

As a first layer, the constellation makes a field legible to itself. Each finding is a point of light — some bright and well-replicated, some dim and contested, some connected by lines of evidence, some isolated in the dark. The structure of a field is not a list but a sky.

Today, if a clinician, researcher, or agent wants to understand the state of a question, they reconstruct it from scattered containers. They read papers. They follow citations. They search for reviews. They guess which caveats still matter and which have already been superseded. They build a map in their head. I have done this myself, and felt the particular frustration of knowing the map would not survive the project. Then, too often, that map disappears when they leave.

The constellation is what happens when that map stops living only in private memory.

It turns scattered knowledge into terrain: shared scientific state in navigable form. Findings can be addressed directly, linked to the evidence beneath them, connected to what they support or contradict, revised when new evidence arrives, and traversed by both people and machines. It is the place where a field can see, in one view, not only what has been claimed but how those claims hang together.

Think about those failed reactions at Haverford. In the current system, the knowledge they contained was invisible — trapped in one lab’s notebooks, inaccessible to anyone else working on similar compounds. In the constellation, every failed synthesis, every boundary condition, every result that narrowed the search space would be part of the shared record. A researcher at another university querying the same chemical space would see what had already been tried, under what conditions, and why it failed. The terrain would be visible instead of hidden.

The same applies across every field. A climate scientist could see the full state of competing estimates for a parameter, with the evidence and corrections attached, instead of reconstructing it from scattered papers. A drug discovery team could see which targets have been tested and where the frontier is genuinely thin. The decisions are still difficult, but the knowledge is no longer buried.

What changes is the unit of record itself. The paper does not disappear, but it stops being the atom. The finding takes its place — situated inside a deeper record that preserves what the paper alone cannot.

A finding, in this world, is a claim with its support attached: what was measured, under what conditions, with what evidence, and with what uncertainty. Some findings support others. Some contradict them. Some quietly depend on assumptions that may later fail. Those relationships are part of the knowledge, not commentary around it.

Corrections become structural rather than rhetorical.

Today, when a paper is retracted or a result fails to replicate, the update often moves by rumor, review article, or the luck of who happened to notice. In the constellation, a correction is not a note on the side of the record, but an event inside it. If a finding weakens, the things that depended on it should know.

Some of these stars have already collapsed, but in the current system we cannot tell — the light is still traveling. Disagreement, contested replication, shifting confidence: these are part of the frontier, and the constellation preserves them rather than papering over them. This is also the state layer that AI agents need to reason over a field — without it, they are reading the same scattered papers everyone else reads.

Shared scientific state does not eliminate judgment, expertise, or interpretation. It makes them inheritable. A pharmaceutical company, an academic lab, and a regulator may still look at the same frontier and weight it differently, and they should. The goal is to let them inherit from the same underlying record rather than rebuild incompatible private maps from the same pile of prose. The constellation is not a centralized authority that decides what is true. It is a shared map on which truth claims, evidence, dissent, and revision can appear in the open, so that a field can see where it already stands.

The Gigafactory

The constellation is the map. The gigafactory is the part that changes the world.

A field does not advance only because it can see itself more clearly. It advances because some part of that clarity gets forced back into contact with reality: an experiment is run, a protocol is executed, a patient is enrolled, a material is synthesized, a measurement is taken, a result arrives, and the record changes. That is the runtime — the layer where ideas meet the world and return as evidence.

We already have a proof of concept. When COVID hit, many countries ran trials. Most were fragmented: hospital by hospital, protocol by protocol, each site carrying its own administrative weight. The UK did something simpler and therefore much more powerful. RECOVERY created shared execution infrastructure: one protocol, one ethics path, lightweight enrollment, integration with existing records, a system any hospital could join. The first patient was enrolled within days. One in six hospitalized COVID patients in the UK entered the trial. Within 100 days, RECOVERY had produced a result that changed care worldwide. Dexamethasone, a cheap generic steroid, reduced mortality in ventilated patients and has since saved an extraordinary number of lives. *

What mattered there was execution structure as much as intelligence. The system made participation easy, learning fast, and discovery something that could arrive. What one hospital learned became light that reached every other hospital within days.

RECOVERY showed what becomes possible when execution infrastructure is shared. The question is how to make that the default condition of science, not the exception.

In science, execution is the scarce loop. Ideas are not the scarcest thing. They may soon be the cheapest thing in the system. Models can generate hypotheses faster than cells grow, faster than patients enroll, faster than assays complete, faster than atoms settle into structure. What remains scarce is contact with reality itself.

This is why the runtime matters so much. A better map without a better execution layer gives you cleaner thought and the same bottleneck. A better execution layer without shared state gives you local productivity and the same fragmentation. The point is the loop between them.

That is what a scientific runtime should do more generally. It should make it easier for ideas to become measured encounters with reality, and for those encounters to flow back into shared state as a byproduct of the work itself.

The Haverford lab proved this at a small scale: failed experiments contain real knowledge, and when that knowledge enters a shared system, it changes what the next researcher can do. Now imagine that at the scale of an entire field. A materials scientist opens the compiled frontier and sees every compound that has been tested, every condition that shaped each result, every failure that narrowed the search space. She designs her experiment starting from the real edge instead of reconstructing the map from scratch. The runtime parses her protocol, schedules the synthesis, tracks the sample through every step, and captures the result as it happens.

The synthesis fails. In the current system, that failure dies locally — in a notebook, a memory, a folder no one else will ever read. In the better system, it enters shared scientific state directly: this compound, these conditions, this protocol, this measured outcome, this uncertainty. A researcher in Osaka, querying the same space that evening, sees the frontier move and does not spend the next six months rediscovering the same dead end.

The map improves execution because it shows where the frontier is thin, contested, or overconfident. The runtime improves the map because every experiment, trial, and failed attempt can return to the shared record in structured form. Better maps choose better experiments. Better experiments build better maps. The cycle compounds.

Results have to enter the record before they are polished into narrative. Protocols have to be machine-operable. Sample identity has to survive handoffs. Measurements have to remain attached to calibration, uncertainty, and lineage. Otherwise the runtime is only speeding up one local workflow inside the old medium.

The gigafactory is what makes that allocation possible: not a factory in the narrow industrial sense, but a system that turns latent possibility into repeated, measurable contact with the world. The constellation makes the frontier visible. The gigafactory moves it.

If that loop is going to matter broadly, it cannot harden into private maps. A runtime that compounds only inside one institution is just a faster local workflow. The whole point is that the loop feeds back into shared state — and shared state requires shared protocols.

Science is inherently federated. No one lab, company, hospital system, publisher, or country sees the whole picture, and the system only becomes more valuable as different parts can connect.

There will always be private data, and there should be. The argument is for keeping the grammar of connection shared.

What must remain open is the grammar, not every institution’s data. If each powerful actor builds its own private scientific state layer, the field does not become a network. It becomes a set of private constellations, each visible only from inside its own walls — and the light cannot travel between them.

We have seen versions of this before. Healthcare digitized records without insisting strongly enough on interoperability, and paper silos became digital silos. Science could easily repeat the mistake at a larger scale.

There are other ways to get this wrong. A state layer controlled by one company becomes a chokepoint, not infrastructure. Compilation without curation could scale confident error faster than careful truth. The same tools that make knowledge legible could make it gameable, and premature standardization could freeze the wrong abstractions into place. Honesty about these risks is part of building well.

The ideal settlement is the same one that made earlier infrastructure epochs work. The protocol layer stays open. Value is created above it. Builders will have to prove the state layer in one workflow, one field, one repeated pain point at a time. From the beginning, though, they should emit public primitives: shared IDs, portable formats, verifiable records, open interfaces.

The larger promise is participation. Think about what happened with code. Thirty years ago, contributing to software meant working at a company that owned the codebase. Then Git, GitHub, and open-source ecosystems dropped the barrier until a teenager with a laptop could ship a library used by millions. The infrastructure did not just accelerate the people who were already coding. It invited an entirely new population into the work.

Science is still behind that curve. The expertise required is real and hard-won — that will not change. But AI is about to change what a single expert can do, and what a motivated newcomer can contribute. When AI agents can run literature synthesis, design protocols, analyze results, and manage experimental workflows, the bottleneck shifts from “do you have twenty years of training” to “can you ask the right question and direct the investigation.” The substrate is what makes that shift possible. Without structured state to reason over, AI in science is just a better search engine over bad records. With it, AI becomes a genuine co-scientist — one that can inherit the full state of a field, propose the next experiment, and write its results back into the constellation.

That is the unlock. The state layer gives AI something to think with. The runtime gives it something to act through. The network lets every discovery compound globally instead of staying trapped in one lab’s notebook. Together, they make science accessible the same way Git and GitHub made code accessible — not by removing the need for skill, but by giving skilled people (and their AI collaborators) infrastructure that multiplies what they can do.

A student who has never set foot in a research lab can open the compiled frontier and see where the open questions are. A small team with AI agents and shared protocols can run investigations that once required a full department. A researcher in Nairobi or São Paulo directs a fleet of AI co-scientists through the same substrate a lab in Boston uses, because the infrastructure is shared.

That is what science abundance means. Not just faster discovery by the same small group, but an expansion of who gets to discover — human and AI together, building on each other’s work, compounding across every frontier at once.

The Sky We Leave Behind

Now imagine every field had what RECOVERY had.

Imagine the frontier of any scientific question was as navigable as a codebase — forkable, searchable, correctable, alive. Imagine a world where a breakthrough in materials science in Toronto compounds with a measurement in Osaka and a correction in São Paulo overnight, because the infrastructure carried it, because the constellation held it, because someone built the layers that let knowledge compound instead of scatter.

A graduate student opens the compiled frontier of her field and starts from the real edge instead of spending a year reconstructing scattered maps. A clinician sees the live state of a treatment landscape, with corrections attached, instead of guessing which review is still current. A failed experiment enters the shared record before the notebook closes, and the next team starts from where the last one stopped.

That world is closer than it looks. The compiler exists. The architecture is clear. The first generation of builders is here.

Science has always been humanity’s longest bet — the wager that what one generation learns, the next can inherit and extend.

For most of history, we have made that bet with paper and memory. We can build something better: a shared sky where every finding has a place, every correction propagates, and every failed path saves someone the trip.

A constellation, not a pile of documents.

Come help place the stars.