The Two-Species Layer: How the Noosphere Became Infrastructure

Here is how it happened, proper nouns included. Koh Phangan, Thailand. The offline side ran under two flags, GreenPill Phangan and ReFi Phangan, together with my co-founder Anastasia: one local node inside two global networks. GreenPill Network and ReFi DAO count dozens of nodes like ours across the planet, from Mexico to Asia, each one real people in a real place, the whole thing held together by Telegram groups and shared documents. The software side ran under EcoSynthesisX: a public-goods studio at first, later a DAO, our workshop and our face in all of those networks. The idea itself was simple to state: take real-world ecological impact, the kind that normally lives and dies inside PDF reports, and put it on-chain as something people could actually hold and fund.

Out of that workshop came Regen Bazaar, a protocol for tokenizing real-world impact, and two of its pilots built with ReFi Phangan: Clean Phangan, where the island's beach cleanups become mintable on-chain objects, and EcoThailand, mangrove and coral restoration in the Thai Gulf, where a tenth of every sale routes back to the local node. Under the GreenPill Phangan flag, DeCleanup Network got its start: pick up trash, upload geotagged before-and-after photos, pass verification, get an on-chain record and tokens; in 2024 that meant thirteen real cleanups across Japan, Nigeria, and Thailand.

The money came from neither a bank nor a fund nor an employer. From late 2023 through late 2025, the three projects went through twenty-odd grants: Gitcoin rounds from GG19 to GG23, Giveth, Octant, Celo. The tranches are laughable by venture standards, a few hundred to a few thousand dollars each, but none of them was decided by a committee; each settled through quadratic funding, thousands of small donations from strangers determining how a shared matching pool splits. And one round, Spring QF, EcoSynthesisX ran itself, on the giving side this time: seven hundred dollars across three chains went to that same DeCleanup.

No office. No payroll. No institution in the funding chain. A local node of real people on an island, global tools in hand, and capital from thousands of strangers who found us over the same wire we shipped the software on.

The last essay argued that the boundary of the mind is softer than it feels from the inside: the felt "I" may be a model the brain renders, no more permanent than the prediction of your face that breaks when you shave. This essay is the outward half of the same thought. The boundary is also softening outward, into a shared layer, and the story above is what that looks like as ordinary work.

A century ago, a Russian scientist gave that layer a name. He turned out to be more right than he could have known, and wrong in one way he could not have helped.

Two noospheres, and only one is arriving

Vladimir Vernadsky was a geochemist, not a mystic, and the distinction matters. In "The Biosphere and the Noosphere," published in American Scientist in 1945, the year he died, he laid out a layered picture of the planet. First the geosphere, the dead matter. Then the biosphere, the layer of life, which did not just sit on the planet but transformed it: living things made the oxygen atmosphere, shaped the chemistry of the oceans, built the limestone. Life became, in his words, a geological force. And then a third layer emerging from the second: the noosphere, the sphere of mind, in which human cognition becomes a geological-scale force in its own right. He pointed at the evidence a chemist would notice. Native aluminum, which never existed on Earth before, was suddenly being produced in any quantity. Mind was leaving a mark in the strata.

The historian David Christian, glossing Vernadsky in a 2017 essay, described the pre-noosphere world in a phrase I keep returning to: even organisms with developed brains "foraged for energy and resources individually, in pointillesque fashion." The phrase is Christian's, not Vernadsky's, but it captures the idea exactly. Each mind a separate dot, connected to the others slowly and narrowly. What changed, in Christian's telling, is that humans began to accumulate information collectively, through language, across the whole species. He calls it collective learning, and argues it is what made us a planet-changing species. In Vernadsky's time this was philosophical poetry with almost nothing to stand on: no network, no instant communication, no way for the dots to actually join.

The word itself has a messier history than the clean version suggests. It emerged from conversations in Paris in the mid-1920s among three people: Vernadsky, the Jesuit paleontologist Pierre Teilhard de Chardin, and the philosopher Édouard Le Roy. Priority is genuinely disputed. What matters is the split between two versions of the same word. Teilhard's noosphere was spiritual: consciousness ascending toward an "Omega Point," a cosmic convergence into something like a world-soul. Beautiful, unfalsifiable, and as far as anyone can tell, not happening. Vernadsky rejected the vitalism his Paris companions were drawn to. His noosphere was secular and material: not a spiritual ascent but a new geological stage, driven by the physical activity of minds building, mining, communicating, moving energy and matter at planetary scale.

It is Vernadsky's version that is arriving. Nobody is converging on a world-soul. What exists instead is infrastructure: cable, data centers, trained models, chains that settle value. A technical substrate through which mind acts at planetary scale, built as engineering, not prophecy. The half of the idea its author grounded in natural science is going up in steel and silicon. The mystical half remains exactly where it was in 1925.

Isn't this just the internet?

Here is the objection this essay has to survive. Open source shipped an operating system through mailing lists in the 1990s. Wikipedia assembled an encyclopedia out of strangers. Remote-first companies coordinate thousands of people who will never meet. If the claim is "people now coordinate through a shared digital layer," the claim is thirty years old, and everything above is a re-description of the web with a Russian citation stapled to it.

Two things have changed, and both are recent.

The first: value now settles on the layer itself. Linux was coordinated over the network, but only the talk lived there; the money, the contracts, the ownership all stayed inside the old institutions. In our story, the funding moved over the same wire as the arguments, strangers pooling capital by protocol with no bank in the chain. The corner of the network where I spent the last few years calls this pattern "Ethereum localism": local nodes of real people in real places holding global tools, ours on Phangan among them, with the chains described as a substrate for human coordination (their phrase, not mine). That label is one embodiment of the pattern, not its name. Balaji Srinivasan's "network state" is another branch: online communities coordinating capital and eventually territory, aimed at political sovereignty, which is a more ambitious destination than anything I am describing. The common core is that the layer stopped merely carrying talk about coordination and started carrying the coordination itself, capital included.

The second change is the one Vernadsky could not have seen, because half of what it requires did not exist. A second stream of coordination now runs on the layer, and it is not human.

The second stream

The word "coordination" is load-bearing here, and I am using it deliberately. The last essay spent its length refusing to call these systems minds, and that still holds. They have no continuous self, and whether there is any experience is an open question. What they demonstrably have is the ability to coordinate, and Part 1 argued that those two things come apart.

There are two layers of it, and the distinction matters. The first is an agent calling tools. The AI sales agent my co-founders and I run for a rental platform fields a query, pulls options from a search tool, runs the request through a language model, checks availability against another service, writes the result into a CRM, and pings a human only at the end, if at all. The industry has standardized this pattern, and the honest label for it is agent-to-tool: one agent reaching for its instruments. Real machine work with no person in most of the chain, but not yet machines coordinating with each other.

The genuine machine-to-machine case now also exists as named infrastructure, and it arrived within the last year and a half. In April 2025, Google released a protocol called A2A, for agent-to-agent, that lets autonomous agents built by different vendors discover each other, authenticate, and delegate tasks among themselves, with no human in the loop. Its own documentation draws the line I am drawing: a machine-to-machine protocol for autonomous agents, explicitly not a messaging app like the one where my human collaborators coordinate. By late 2025 the major agent protocols sat under one neutral home, the Linux Foundation's Agentic AI Foundation, co-founded by the largest AI labs. The second stream is not something I am asserting from one bot. It is being standardized, in the open, right now.

The loop, measured

Two streams on one substrate could still be a coincidence, two tenants in one building. What makes them one story is that each is being built out of the other, and that is a checkable claim, not a metaphor.

One direction: the machine stream is made from the record of the human one. A model is trained on what human coordination left behind: everything we wrote, argued, documented, photographed, filmed, and recorded. Text models compress our writing; multimodal models compress our images, video, and speech. The machine stream is human coordination in residue form.

The other direction: human coordination increasingly runs through the machine stream. Drafts get written with a model in the loop, collaborators get found through it, and the funding systems my projects lived on until about a year ago were already wiring model-based evaluation into how grants get routed. In my own working week, the share of coordination that passes through a model at some point is no longer small. That is an anecdote. The next two facts are not.

Since roughly November 2024, about half of new articles published on the web have been primarily machine-generated. That number comes from Graphite, an SEO firm that sampled tens of thousands of pages from Common Crawl, the public web archive that also feeds AI training data, and ran them through AI detectors. The methodology has honest limits (detectors misfire, and paywalled human writing is undercounted), but the shape is clear: the share went from around a tenth before ChatGPT to roughly half, where it has plateaued. The layer's new record is now about half machine-written.

And that closes the loop in a way you can measure. In 2024, Shumailov and colleagues published a paper in Nature showing that models trained indiscriminately on recursively generated data degrade, irreversibly, as the tails of the original distribution disappear. They named it model collapse. The firms building the machine stream now curate their training data against the machine stream's own output. Substrate and users are constituting each other to the point where it has become an engineering problem: the layer is being partly rebuilt from its own exhaust, and its builders are fighting the consequences.

None of this needed a new philosophy. That cognition leaks past the skull is an old, named idea: Clark and Chalmers argued in 1998 that Otto's notebook is doing the work of Otto's memory and is in the relevant sense part of his mind, and Edwin Hutchins showed navigation on a ship's bridge done by the whole coordinated system rather than any single head. Two things are new. The scale, and the fact that half the scaffolding is no longer passive. Otto's notebook never coordinated back. Where exactly the boundary of "mind" sits is a definitional fight I do not need to win; the philosophers call the failure mode cognitive bloating, and the worry is fair. The structural change is observable whatever you decide to call it.

Whose layer it is

Vernadsky's noosphere was, in principle, nobody's property. The machine stream is not. The models that human coordination increasingly runs through are built, owned, and priced by a small number of firms, and the protocols that let agents coordinate are being standardized by the same companies that dominate the rest of the stack. So when I say coordination now runs through the machine layer, I am also saying it runs through infrastructure that someone owns, that is legible to its owners in a way a message between two people is not, and that can be shaped, throttled, or surveilled at the exact layer where the coordinating happens. This is the same concentration story I have written about in the context of wealth, showing up in the substrate of cognition itself. The noosphere becoming infrastructure is not automatically the noosphere becoming free. Whether the layer ends up a commons or a toll road is not a technical question, and it is very far from settled.

Back to the node

When we were raising those funding rounds, I thought we were just using better tools: a faster way to find collaborators, move money, ship software. I now think we were standing at an early edge of the thing Vernadsky described, at the point where it stopped being a metaphor and became a place you could work. The points are joining, the way he said they would. He imagined the layer as a sphere of human reason, and on that one detail the century corrected him: the layer is being built more or less on schedule, and it is not only ours. A second stream of coordination runs on it beside us, made from the record of our own, feeding back into what we make next. It is possible we are a transitional form, the last cohort that remembers fully isolated cognition; I hold that loosely, as a frame and not a fact. The infrastructure underneath that frame is real either way: it is being laid right now, and the question of who owns it is still open.