The argument the builders are having about AI, and why it deserves a wider audience.

The most consequential argument about artificial intelligence right now is not being had in the pages of the major magazines, the congressional hearings, or the large AI conferences. It is being had among the people building the infrastructure those systems run on. About a thousand of them met in San Francisco in March. Their claim deserves a wider audience: the political and economic architecture being built beneath AI is what will decide the technology's effect on the next decade. The default architecture concentrates data, capital, and decision-making power in a small number of actors, in ways no democracy should accept.

The event was Funding the Commons' Intelligence at the Frontier, the third annual gathering in San Francisco for a community that has also convened in Buenos Aires, Bangkok, Tokyo, Taipei, Berlin, and New York over the past few years. The community began at Protocol Labs, the organization best known for IPFS and Filecoin, in 2021. It now sits on the UNDP's Blockchain Advisory Board, convenes alongside UNICEF's Office of Innovation, and counts the Internet Archive, the Tor Project, NEAR, and Berkeley's Center for Responsible, Decentralized Intelligence among its partners.

The March event doubled the previous year's attendance. A thousand people in total, against roughly five hundred in SF 2025. One hundred and seventy-six speakers from forty-nine organizations. Attendees from ninety-two countries. A third arrived from outside the United States. Fifty-five funders were in the room. Eight in ten were attending Funding the Commons for the first time, a number that reflects the community's history as a recruiting ground into a broader conversation about decentralized infrastructure, populated heavily by people who came through the Ethereum and Protocol Labs worlds.

The speakers were unusual. One of the eight co-authors of Attention Is All You Need was on stage, the 2017 paper that introduced the transformer architecture. The "T" in ChatGPT is that paper's contribution. A former science and technology policy advisor to Presidents Clinton and Obama followed him. A Rhodes Scholar anthropologist who spent a decade at IARPA and DARPA (where he was known in-house as "the DARPAnthropologist") before taking over the AI Division at USC's Information Sciences Institute. A Harvard-trained lawyer who founded the World Economic Forum's blockchain and digital assets team before running crypto's largest policy coalition. The Tor Project's director of fundraising. The co-lead of Protocol Labs' economies-and-governance research. The former Chief Privacy Officer at Cisco. Researchers from NIST, Harvard, Mozilla, MATS, and Arizona State's Cooperative Futures Institute. On paper, a heterogeneous room. In practice, a room in unusual agreement about where the problem sits.

A mainstage panel, audience watching. Most of the people in this room had never been to a Funding the Commons conference before.

The problem is not that AI will take jobs, or fail to take jobs, or become conscious, or fail to become conscious. Those are arguments being had elsewhere. The problem is that the coordination infrastructure the next decade will run on, the rails for how value moves between autonomous systems, how agents are authenticated, how compliance is verified, how mixed communities of humans and machines govern shared resources, is being built right now, by a small number of centralized actors, on terms most of the world has not been given the chance to weigh in on. The frontier labs are building their own settlement protocols. Visa, Google, Coinbase, Stripe, PayPal, and a hundred startups are racing to become the default payments layer for machine-to-machine commerce. The largest sovereign states are nationalizing AI infrastructure. The United Nations itself has warned of imminent financial insolvency. The institutional order that managed global coordination for eighty years has not yet been replaced by anything coherent, and the technologies that will define the next eighty are being designed and deployed in the interim.

The community that gathered in San Francisco believes there is a specific kind of infrastructure this moment requires: open-source and auditable, sovereign at every layer, resistant to capture by any single state or corporation, compliant where legal process applies, and structured so that value accrues to the participants who produce it rather than to a small set of well-capitalized intermediaries. The community does not claim this is easy. Five years in, no dominant solution exists. Juan Benet of Protocol Labs said this directly in his closing keynote: the broad public-goods movement "has not yet achieved what we set out to achieve." The admission was the opposite of a victory lap. It was a reset, delivered to the one audience prepared to hear it.

Three observations, named across the two days by different speakers in different sessions, gave the argument its structure. Each deserves to be taken on its own terms.

Institutions need to co-evolve, and almost none of them are built to. Adam Russell, the DARPAnthropologist now running the AI Division at USC's Information Sciences Institute, argued that institutions are themselves a technology. The oldest we have. Their job is to solve the recurring functions societies depend on: coordination at scale, knowledge accumulation, resource allocation, conflict resolution, risk pooling, identity. AI is the first general-purpose technology in living memory to unsettle all of these functions at once, at a speed no existing institution is structurally equipped to absorb. Resilience, Russell said, is not stability. It is fast co-evolution. Very little of what we depend on was designed to co-evolve at this pace. The implication: institutions need to be part of the response, and the response begins by naming exactly where their current structure blocks them from contributing.

The default trajectory concentrates value. Luke Drago and Rudolf Laine's Intelligence Curse, an essay series that has lodged itself in both AI safety and policy circles over the past year, became the conference's most-cited reference. The framework borrows from development economics. Countries that discover enormous natural wealth often end up worse off, because the wealth concentrates and states built on extraction stop needing to invest in their populations. AI, Drago argues, is on track to produce exactly that kind of resource. The value in an automated economy accrues to whoever owns the frontier model weights, the compute, and the data. The structural consequence, stated from several stages across the two days: AI safety and value redistribution are the same workstream. Treating them as separate workstreams is how the default trajectory wins.

Due process is friction, and it is being engineered away. Sheila Warren, a Harvard Law-trained lawyer whose formative legal experience was with the Innocence Project and the California Appellate Project, now chair of the Advanced AI Society and former head of the WEF's blockchain and digital assets work, returned across keynote and two interviews to a single argument. Friction in social, financial, and legal systems is often not a design flaw. It is a feature. Due process is friction. The review step that catches an AI agent before it makes an irreversible decision is friction. The appeals window that lets a wrongful conviction be overturned is friction. The design instinct of the last decade has been to remove friction from user experience in the name of speed and convenience. That instinct, applied to the agent economy, will remove friction from governance itself, because the same systems will do both. We are watching the erosion of due process in real time, in domains we thought were settled, at a speed that does not register until it is well advanced.

These three concerns run under most of what follows. They are the preoccupations of people who have built the technology they are discussing. They are saying something that the frontier-lab communications teams are not saying, that the AI safety debate in its current form is not quite capturing, and that policymakers and business leaders outside this room are not yet aware of. The window in which their observations remain actionable is narrow. What follows is an attempt to state them clearly, and to identify the audience that should care.

Eight questions the room kept returning to. The default trajectory each one is on. The specific work already underway against it.

The room. One thousand attendees, ninety-two countries represented, a third arriving from outside the United States.

The eight questions below recurred across the two days, in different sessions, raised by people from very different professional worlds. Each one has a default trajectory the room named: a failure mode that the technology produces unless something interrupts it. Each one also has work already underway. The point of this map is that the right side of every row is real. The bridges are narrow, undercapitalized, and unfinished, but they exist. The work of the next two years is widening them faster than the default consolidates.

Each row above is treated in more depth in the sections that follow. The right column does not win automatically. It exists, with named people doing named work. The question of the next two years is whether the capital and the institutional cover arrive in time to scale it.

Two people who've built institutions for the U.S. government walk into a conference and say the structure is the thing to design.

Adam Russell opened his Day 1 keynote by apologizing for being an anthropologist. He was there, he said, mostly to fill the time until Tom Kalil arrived. The joke read as modesty; the talk read as the conference's intellectual anchor.

Russell's argument was that we routinely mistake what technology is. We treat institutions as the things technology acts on, the fixed structures that absorb or resist innovation. He proposed the inverse. Institutions are technology. The oldest we have. The formal, semi-formal, and informal arrangements that societies invented to solve their core functions: coordination at scale, accumulation of knowledge, legitimate authority, resource allocation, conflict resolution, risk pooling, identity and belonging. Each of these is a technical problem. Each has received a succession of institutional answers over human history, some more successful than others.

Running through his slides was a framework from the Dutch philosopher Philip Brey's 2022 paper on socially disruptive technologies: the ways a technology can unsettle the institutions it encounters. Depth of impact. Range of impact. Ethical salience. Speed of diffusion. Irreversibility. AI scores unusually high on each.

Institutions are the OG Technology. Nowadays we talk about institutions versus technology. We distinguish those things. My argument is institutions are our technology. Adam Russell · It's the Institutions, Stupid

Adam Russell (left) and Tom Kalil · Fireside chat, Day 1

What he was pointing at is a diagnosis more than a prescription. Institutions evolve slowly because that is their job. Resilience, he said, is not stability, it is fast co-evolution. The institutions we have now are mostly oriented toward exploiting the contexts they were built for. Very few were designed to explore. AI has arrived as a general-purpose technology at a moment when almost no institution has the structural capacity to co-evolve with it at its speed.

Some of the most interesting institutional work in the tower was happening off the mainstage, inside the building itself. David Casey, FtC's CEO, designed a comparative governance experiment across six floors of Frontier Tower over the conference weekend. Each floor received between $700 and $3,000 in seed funding, $10,300 in total, jointly funded by Octant, the Hypercerts Foundation, and FtC. The single condition: each floor lead had to document how they decided to spend it. Two on-chain governance tools were made available to anyone who wanted to use them. Simocracy, an AI-assisted preference-aggregation platform built by David Dao, and Hypercerts, a contribution-attestation protocol integrated into the FtC platform by James Farrell.

The empirical finding was the part that surprised most people. Every floor chose some variant of concentrated authority. Three settled on benevolent dictators, one ran a pre-existing council, one delegated to a single member, one defaulted to its parent nonprofit. None of the six adopted on-chain mechanisms during the event, even though Floor 12 hosts what its operators and the Ethereum Foundation describe as the first permanent Ethereum community hub in the United States. Floor 14, Human Flourishing, was the only floor that built new governance tooling during the weekend, a participatory budget tool integrated with Hypercerts. The Artisan Fund ran a live public-goods funding round timed to the event. These are small experiments. They are also the shape of the argument above, in miniature, institutions attempting to redesign themselves, in public, at the speed the technology requires.

The most striking moment came after the conference closed. On March 16, the day after the event ended, the floor leads gathered on Floor 14 with David Dao, co-lead of Protocol Labs' economies-and-governance research, to allocate the $5,000 Hypercerts Foundation contribution across eighteen proposals that floor members had minted as Hypercerts during the weekend. The mechanism Dao brought into the room was Simocracy. Each floor lead had sat earlier for a thirty-minute conversation with an AI system, describing what they valued, what they would instantly fund, what they would instantly reject. From those interviews, the system generated cartoonish digital avatars, sims, and placed them inside a digital version of the tower. The sims walked around, met each other, argued, deliberated.

Underneath Simocracy sits the S-Process, originally developed by Andrew Critch and now used by the Survival and Flourishing Fund, which has deployed $34.92 million in grants using AI-aggregated preference functions paired with human veto authority. Hypercerts, the impact-data protocol pioneered inside Protocol Labs and now stewarded by an independent foundation, held the proposals. The same protocol issued Hypercerts to the conference's 176 speakers, published March 17 via the AT Protocol.

Simocracy in session · the tower's sims walking around their digital twin floors, S-Process allocating $4,270 of $5,000 across the floor leads' proposals · March 16, 2026

Two things made the session extraordinary. The first was the willingness of floor leads to take their own twins seriously. Judy Zhou from Floor 14, who runs Human Flourishing with a focus on community and gardening, asked her sim what she should eat for dinner. The sim recommended local groceries and a walk. She loved it. Another floor lead spent her morning in the tower procrastinating on her actual job to track down her rival's sim and argue with it, trying to convince the twin of the floor lead she disagreed with to see things her way. The second was the human override. After the sims completed their allocation, the humans in the room voted not to divide the $5,000 across floors at all. They pooled it into a tower-wide community treasury and committed to govern it through Simocracy going forward. The treasury reaches roughly $8,000 once event revenue is deposited. Two pieces of infrastructure now sit alongside it. The Frontier Tower Agent, an autonomous co-signer named René, was built during the weekend's hackathon and is being deployed onto the treasury. A 2% / 20% revenue-share template was added to the building's contractual stack so that future events seed the treasury automatically.

David Dao is the first to say none of this proves anything yet. The hypothesis, that digital twins might extend Ostrom's principles past Dunbar's limit, is one iteration old. But it is the shape of the thing. Institutions of the next decade may not look like the committees and councils and legislatures we inherited. They may look like rooms full of people and their sims, each interaction accountable back to a human but operating at a cadence and scale no human committee can match. The experiment was small. The implications are not. (Casey's full write-up of the comparative governance findings is circulating among partners now.)

The same experiment, in the places it has to work

The tower was a sandbox. The institutions most under pressure from what AI is doing are running the same experiment at much higher stakes, with fewer privileges. Two of them, represented on the mainstage by their own implementers, stood out for the specificity of their work.

Matthias Broner spoke for UNICEF's Office of Innovation. The talk was titled The Coordination Problem: How UNICEF Is Building Financial Rails for the Age of AI, and the project underneath it is called AidLink. The problem it addresses is blunt. A significant share of the world's poorest populations, the communities UNICEF exists to reach, have neither bank accounts nor reliable internet. Traditional humanitarian cash transfer is expensive, slow, and leaky; existing crypto alternatives assume connectivity that doesn't exist. AidLink uses SMS-based digital wallets to reach people where they are, pairs them with blockchain-based distributed funding mechanisms such as DRIPS, and channels idle crypto treasury into direct humanitarian disbursement. Broner also named a set of adjacent deployments: Babel (parenting), Hope (humanitarian information), Oki (menstrual health), Primero (child protection), Rabbit Pro (messaging). Each is a small, live, specific-to-context piece of infrastructure. The talk was one of the clearest reminders at the event that the coordination primitives this community talks about in the abstract are already being used, by people the rest of the world forgets to include, to move real money to real children.

Augusto Ardiles Diaz · Government of the City of Buenos Aires

Augusto Ardiles Diaz spoke for the Government of the City of Buenos Aires. His talk, Designing the Institutional Infrastructure for AI: How Cities Become Innovation Ecosystems, took as its thesis that cities, not nations, not labs, are currently the most functional unit of AI governance. Buenos Aires produces about a fifth of Argentina's GDP in a small geographic footprint. Its city government has the authority to run regulatory sandboxes, set soft-law norms, and co-locate academia, companies, and investors in a single AI district. None of those powers require federal approval. The tools for testing emerging technology at city scale are sharper, faster, and more politically survivable than at national scale. Ardiles Diaz's argument was not that cities will solve AI policy alone. It was that they are the laboratories where the good policy will be proven out before it can possibly be coherent at any larger level, and that places like Buenos Aires, with talent density, cultural openness, and the political flexibility to experiment, have an outsized role to play in how the institutional playbook for AI gets written. The FtC community's partnerships with UNICEF and with city governments in Latin America, Asia, and Africa are, among other things, a bet on this being true.

Tom Kalil, who followed Russell on stage, spent the 90s and 2010s inside the institutions in question. He helped shape the National Nanotechnology Initiative under Clinton, advanced market commitments under Obama, and the legal scaffolding for DARPA's incentive prize authority. His current work at Renaissance Philanthropy is an explicit attempt to install new mechanisms alongside, not against, existing ones. Pay-on-delivery research funding. Unbundling idea generation from execution. Marketplaces for outcomes rather than inputs.

Kalil's most useful sentence was a rephrasing of a very common one.

A lot of people are saying, oh, what is AI going to do to us? And the question that I think is a little bit more life-affirming and interesting is: what do we want AI to do for us? What is the civilizational tech tree that we want to build? Tom Kalil · in conversation with Adam Russell

This is the rare version of the AI question that treats the technology as a tool for building rather than an asteroid to brace for. Kalil's answer, threaded through the conversation, is that the bottleneck is almost never the model. It is the complements. The benchmark datasets, the verified proof systems, the curated knowledge bases, the coordination norms that made progress in structural biology and language modeling possible and that do not yet exist for most of the problems we would most like to solve.

The protein folding example he kept returning to was not a story about DeepMind. It was a story about the Protein Data Bank, a 1971 agreement among structural biologists that produced, over five decades, a twenty-billion-dollar curated dataset and a benchmark competition that made AlphaFold legible as progress. "Academics are rewarded for novelty as opposed to utility," Kalil said. The data that would train the next AlphaFolds in chemistry, climate, or medicine is sitting behind paywalls, in labs, or simply never generated, because nobody is paid to generate it.

This is not a glamorous problem. It is the conference's real one.

Sheila Warren's keynote, The Acceleration of Dehumanization, took the institutional question sideways. Warren had spent five years at the World Economic Forum running its Center for the Fourth Industrial Revolution, the institution many people in the room held responsible, rightly or not, for making blockchain and AI respectable to governments. She is Ostrom-fluent. Her frame was the commons, scaled down rather than up.

Localized governance that's very hyperlocal to small communities is not like taking one model of governance and blowing that up and trying to extend it everywhere. I think of it as a language translation issue. You do something in English, then you just translate it directly into some other language and assume it's going to be culturally relevant, which it never is. Sheila Warren · The Acceleration of Dehumanization

She ended on a concept she said she hadn't written about yet. Differential friction. The early promise of Bitcoin was the removal of friction from payments. She had been skeptical for a long time. Not all friction is bad. Friction is what due process looks like in code. Friction is what makes a transfer from an employer to an employee different from a smart contract that encodes the employer's interpretation of the employee's behavior in permanent form. Friction is, in many cases, governance.

Remove all friction from a system and what you have left is not liberty. What you have is the frictionless ability of an agent, yours, nominally; working for a company whose incentives may or may not align with yours; interpreting your preferences well or poorly, to codify your digital life into smart-contract form in ways you cannot later unwind.

Warren's closing word on this was collaborative. Her language, spoken to a room of builders, was less a call to arms than a warning about what individual virtuosity produces when it is not paired with collective structure. The reality of competing with defense tech and surveillance tech alone, she said, is that it is not going to happen. Collectively, it might.

Luke Drago's resource curse, Ashutosh Desai's colonial mirror, and the default concentration of AI's value.

Luke Drago is twenty-seven and runs Workshop Labs out of North Carolina. He has spent the last year writing about AI under the name The Intelligence Curse, which is also the title he gave his Day 1 keynote and, now, an essay that has lodged itself in both safety and policy circles. His talk was the most-cited reference of the conference.

The framework is borrowed from development economics. The resource curse describes the pattern by which countries that discover enormous natural wealth often end up worse off. Oil concentrates. States built on resource extraction do not need to educate, invest in, or negotiate with their citizens, because the value is in the ground rather than in the people. Over time, the state drifts from its population, and its population from it.

Drago's argument is structural, not rhetorical. AI, he says, is on track to produce exactly this kind of resource. The value in an automated economy accrues to whoever owns the frontier model weights, the compute, and the data the weights were trained on. If you believe, as the labs themselves now publicly claim, that a meaningful fraction of knowledge work will be automated within the decade, you are describing a world in which labor stops being the mechanism by which most people are included in the value they produce. Dario Amodei's 20% white-collar unemployment figure is not speculation here; it is the assumption.

The bedrock of the economy and of civil society is this relationship between labor and capital, where because people are useful, people get rewarded. When technologies rewrite that social contract, the social contract can move in two different directions. More favorably towards regular people, or much worse. Luke Drago · The Intelligence Curse

Drago's exit from the diagnosis was a technical one. Not policy reform but different technology. Build models, interfaces, and coordination systems that keep humans economically load-bearing rather than substituting for them. Reduce biological and cyber risk so frontier capability can be diffused broadly without catastrophe. Pair diffusion with democratization, so that capability in the hands of many doesn't reduce to a subscription to ChatGPT. Refuse, he said, to build the model that makes the human irrelevant. The room noticed that this was a technical argument addressed to technical people, not a policy appeal in search of a regulator.

Luke Drago · off-stage interview · co-author (with Rudolf Laine) of The Intelligence Curse

In an interview off the mainstage, Drago sharpened the positive vision. The analogue he returned to was not a modern economist but Pope Leo XIII's 1891 encyclical Rerum Novarum, the Catholic Church's first attempt to reason publicly about industrial automation and the relationship between labor and capital that it was dissolving. Leo's observation, Drago said, was that the answer to automation is not more socialism and not more concentration. It is to convert workers into owners. Not five hundred companies that control everything. Five hundred million companies, each with small teams of people operating alongside carefully sculpted machines, each better at understanding a local market than any centralized system could be. That vision is technically achievable, Drago argued. It requires specific choices about what gets built, by whom, and on what terms. The absence of those choices is the default.

Ashutosh Desai took the argument further back in time, and further toward the political. His family is from a village outside Surat, in Gujarat, India. His keynote, On the Shoulders of Capitalism, opened with a photograph of the East India Company's first 1608 landing at that coast and a reminder that the East India Company was the second publicly-traded shareholder company in human history. The first was its Dutch predecessor. The joint-stock corporation, he was arguing, did not come into the world neutral. Its earliest deployments were colonial.

Technological advantage enabled capital to out-leverage both labor and governance. The British control of India was enabled by Indians. The Mughal control of India was enabled by Indians. My last name is Desai. Landlord, tax collector. As we think about our roles here today, are we going to be the ones who enable certain types of caste systems to be rebuilt with this new revolution? Ashutosh Desai · On the Shoulders of Capitalism

Desai's move was to refuse the comfortable self-image of Silicon Valley as a neutral meritocracy and read the AI moment through the same lens Indians have used to read 400 years of their own political economy. His target was not the technology. His target was what he called "Colonialism 2.0", the rebranding of the American military-industrial complex as American dynamism, and the ease with which a class of founders and funders can persuade themselves that extractive concentration is natural law rather than policy choice.

He said something the room found uncomfortable and true: the penalty for dishonesty, in the current climate, is lower than the penalty for uncomfortable truths. Then he gave a concrete counter-example. India's Aadhaar, open-source, sovereign, state-run digital identity infrastructure that has absorbed more than a billion people into the digital economy over a decade, is the kind of thing the builder class could be working on. Worldcoin, with its retina-scanning orbs, is the kind of thing it is. Aadhaar was not built by venture capital. It was built by an institution that understood what it was for.

Desai's closing request was less technical than temperamental. Speak truth to power. His observation was that sycophancy, already a well-documented problem in the alignment of large language models, is not only a model problem. It shows up across AI's adjacent rooms: investors telling founders what they want to hear, founders telling engineers what they want to hear, engineers telling themselves that the thing they are building is something else.

Both men, one from development economics, one from colonial history, had arrived at the same observation by opposite routes. The default trajectory of the AI transition concentrates. If you want a different trajectory, you have to build one, and you have to build it early enough that the default has not already won.

Max Song, who serves as entrepreneur-in-residence and strike force member in Peter Diamandis's ecosystem and spent earlier years co-founding Make School (a university experiment in San Francisco), pushed the argument one register deeper. The labor-capital contract is not only an economic arrangement, Song argued. It is the arrangement out of which our educational, civic, and ritual structures have been built. Factory-era schooling, standardized classrooms, standardized tests, age-based cohorts, rules about punctuality and obedience, was designed to produce modular workers for factories. It worked. It is also sunsetting. The contract for which it trained people is dissolving faster than the institutions it runs inside can adapt to.

Education is civilizational architecture. We are currently spending somewhere between a hundred and a thousand times more on AI infrastructure than on developing human capability to use it. That is going to come back and bite us. Max Song · in conversation

Song's working project, named Primer after Neal Stephenson's Diamond Age, is an attempt to build a personalized tutor that grows with the learner. Part of the point is that the current tools make it possible to learn by making first and asking the AI to explain what you just made afterward. The inversion, make, then understand, is in his own use reducing existential anxiety and replacing it with something closer to joy. He is careful not to oversell it. But the observation underneath is one the rest of the room kept circling and one the broader AI debate still struggles to hold. The question is not whether AI will transform human capability. The question is who gets access to the version that transforms it upward, and who gets the version that flattens it. That answer is a policy and infrastructure question, not a technology one.

Song also raised a frame less common in Silicon Valley. Human knowledge is not uniformly represented in the corpora that train today's models. The kinds of knowledge that shaped the ecosystems small communities manage, indigenous, traditional, ecological, oral, exist in forms that are often not digitized, or digitized poorly, or held in cultural practices that cannot be captured by scraping Reddit. He calls this ancestral intelligence. The project of making it legible to AI systems, and of ensuring that the communities holding it are partners rather than raw material, is one the broader AI discourse has barely started.

What the world is already starting to look like when autonomous software acts on its own, and why the people closest to it are half-exhilarated and half-terrified.

Shady El Damaty, founder of Holonym and the human.tech protocol, opened his Human Flourishing keynote with a confession that set the tone for much of Day 2. "A lot of my priors about products, tech, user, human computer interaction have been completely disrupted over the last six months," he said. The qualifier that followed was not the standard optimism-of-the-builder. It was closer to disoriented witness. He was not the only person in the room who said some version of this. In conversation after conversation, during hallway breaks and onstage, the phrase the last six months came up without prompting. The object being described was the speed at which autonomous software, agents, has moved from prototype to daily working reality inside the community that builds it.

Shaw Walters stated the consequence more plainly. Walters is the founder of Eliza Labs and ElizaOS, and one of the more technically fluent agent-builders in the room. He opened his Day 2 talk by describing his current relationship to his work this way:

I mostly make code or like babysit robots making code these days. Shaw Walters · Open by Default

The line is not a joke in the sense most readers outside the room will take it. Walters is a capable engineer who has spent the last year supervising autonomous coding agents that produce software faster than he can, in quantities he has given up on reading line by line. He does not call himself post-labor. He does describe himself as post-job. The audience, which contained a great many people in similar relationships to their own agents, did not find the observation surprising. Several, pressed later, said the same thing.

Shaw Walters · Eliza Labs · Open by Default

What any of this will mean at the scale of an economy, a legal system, or a country was the question Juan Benet and Illia Polosukhin's closing fireside spent ninety minutes trying to get at. Polosukhin is one of the eight co-authors of Attention Is All You Need, the 2017 paper that introduced the transformer architecture now running inside every major frontier model (the "T" in ChatGPT is that paper's contribution). He left Google to start NEAR Protocol. Juan Benet founded Protocol Labs, the organization behind IPFS and Filecoin and, indirectly, Funding the Commons itself. The two have been working the intersection of crypto and AI for four years. The fireside was neither entirely hopeful nor entirely grim. It was precise about specific scenarios, and the specificity is what made it frightening.

The first scenario was structural, and it was about consolidation. If AI capability continues to accelerate and a small number of frontier labs capture most of the economic surplus, Polosukhin said, the obvious policy response for large states is to nationalize those labs. Two or three US labs becoming national security assets in a year or two is not a remote outcome; it is a predictable one. Once that happens, the infrastructure the rest of the world runs on becomes a strategic weapon of whichever government holds it, available for freezing, surveillance, or selective shutdown, the Tether-froze-Venezuela pattern, but applied to AI inference instead of to stablecoin balances, and exercised by a state rather than a private issuer. The only structural alternative Polosukhin could name was open, neutral rails that belong to no single state or company. He was candid that building them before the nationalization window closes is the work of approximately three to five years, not decades, and that the window is shortening.

We need some platform that belongs to everyone, that is neutral, kind of how the Internet is. Imagine the Internet would be US Internet, and everybody who enters needs to get on with their passport. This would obviously be a very bad world. With AI we are getting into that state pretty quickly. I think we're literally like a couple of years away from that. Illia Polosukhin · in fireside with Juan Benet

Illia Polosukhin · Building a Secure Agentic Future

The second scenario was weirder. Juan Benet and Polosukhin agreed that the first fully autonomous AI corporation, an entity that raises capital, enters contracts, hires subcontractors, and operates without a human in the decision loop, is probably launching this year or next. The first AI "person," an entity the law will have to reckon with as a participant rather than a tool, is on a similar timeline. Polosukhin's shorthand for what is emerging was the agent city: a legible economic zone in which autonomous corporations, agent-mediated governance, agent-to-agent markets, and the humans still in the loop all interact under a shared legal and infrastructural framework. Parts of that zone already exist, prototyped into NEAR's ecosystem. Most of it does not.

Juan Benet, who has a habit of saying calmly what other people tiptoe around, then raised the scenario the rest of the room had been avoiding out loud. You can build a city like the one Polosukhin described, he said, and immediately "your hair should be rising, because this is exactly how you get paperclip maximizers. Be very careful how you set it off." Polosukhin answered with a one-line joke ("That's how Skynet started") and they moved on. Neither was joking. Both were describing a design surface on which the default failure mode is catastrophic rather than merely inconvenient. Both also believed, and this is the part that separates them from the public AI doom register, that the technical tools exist to contain the design surface, if the builders of that infrastructure get capital and institutional cover faster than the alternative.

The third scenario was subtler and in some ways the most sobering. Juan Benet pointed out that even if governance works, even if the coordination rails get built, even if alignment succeeds, the basic macroeconomic assumptions that modern states rely on will stop describing the world. GDP, the metric by which every government on earth measures itself, measures the value added by labor. If autonomous agents and robotics produce most value, GDP stops being meaningful. Unemployment, the metric the Federal Reserve uses to set interest rates, becomes incoherent once the denominator is a mix of people, agents, and machines. Every major economic transition in modern history has been a fight over how to assign value; this one is a fight over whether the existing metrics can even describe what is happening.

It's very unlikely that ten years from now we don't have major elections and governments determined by the question of how we're going to deal with the fact that the notion of our labor and our capital and so on are radically being reshaped underneath us. So no pressure for us. Juan Benet · in fireside with Illia Polosukhin

Both Juan Benet and Polosukhin landed in the same place. The technical tools to interrupt the default trajectory exist. Whether they get capital and institutional cover faster than the alternative is the question of the next two years. Neither believed the outcome is determined. Neither believed the window is generous.

Closing fireside · Juan Benet & Illia Polosukhin · Day 2

The infrastructure gap, named

The speculative register should not obscure the technical one. Polosukhin's own keynote was a thirty-minute, non-speculative sketch of a structural problem the agent economy is already producing, without quite admitting to.

The Internet is not built for agents. Internet is built for people clicking, typing URLs. It's not built for Sybil resistance when one agent can effectively just DDoS a website by accident. It's not built for commerce where the agents now can do commercial relationships. And we need privacy because we are putting so much more information about ourselves into the systems. Illia Polosukhin · Building a Secure Agentic Future

When your browsing agent negotiates with a vendor's sales agent over a service contract, five things happen that the current web has no infrastructure for. The agents need to verify each other's legitimacy. They need to negotiate at machine speed. They need to transact without human-in-the-loop authorization. They need to produce an auditable record of what they did. They need to preserve the privacy of the humans they represent. None of these are solved problems. Today they are solved, badly, by every agent-payment protocol being separately invented at Visa, Google, Coinbase, Stripe, PayPal, and a hundred startups, most of them racing to become the default settlement layer before any public standard emerges. The first company to win becomes, by default, the one the entire machine economy runs through, on terms it sets.

Brian Behlendorf co-founded the Apache Software Foundation thirty years ago and is now at the Open Source Security Foundation. His fireside with Tricia Wang walked through what it took to turn "open source" into a recognizable movement: the naming, the legal defensibility, the commercial model, the long institutional arc from "freeware" marginality to IBM acquisitions. Wang's implicit argument, which she made explicit, was that proof of control, the verification that an agent acting in the world is doing so under a traceable, accountable, tamper-evident chain of authority, needs the same kind of movement. Without a shared name, a shared definition, and a shared commercial model, the category dissolves into the same vocabulary fatigue that sank the last decade's conversation about decentralization. The Advanced AI Society, whose chair Sheila Warren sat in the audience, is an early attempt at this. It is an industry association organizing the companies building proof-of-control technologies into a coordinated movement that enterprise buyers and regulators can recognize as a single thing.

Athena Aktipis of Arizona State's Cooperative Futures Institute brought the deepest design frame. Her research spans multicellular biology, human societies, and multi-agent AI. Three systems that share the same underlying problem: how do you scale cooperation among many agents without the cooperative system being consumed by exploiters? Multicellular life solved this through five foundations, controlled proliferation, controlled cell death, fair resource allocation, division of labor, maintenance of a shared extracellular environment. Cancer is what it looks like when those foundations fail. Each foundation has a technical analog in agent design. Controlled cell death, she pointed out, is the biological name for a kill switch.

Cancer is a problem of multicellular cooperation. When we look at what cooperative foundations are required for multicellularity to be viable, we see that cells limit their own proliferation, have controlled cell death, allocate resources effectively, divide labor, and maintain an extracellular environment. These are foundations we can think about as we build and scale multi-agent systems. Athena Aktipis · From Multicellularity to Multi-Agent AI

Michelle Finneran Dennedy, former Chief Privacy Officer at Cisco and now at Abaxx, closed Day 1 with Privacy is the Heartbeat of Humanity: A Manifesto for the Agentic Age. Her emphasis was not privacy as defensive crouch. It was data value as something currently unmeasured and therefore, by the Drucker rule, unprotected. The future of AI, she said, will not be decided by what machines can do. Machines can do a lot. The future will be what builders choose to protect.

Running underneath all of this. Polosukhin's neutral rails, Aktipis's cooperation primitives, Dennedy's data sovereignty, Warren's proof-of-control framework, Juan Benet's humility about the last five years, was a single question the conference kept circling without landing on. Who builds the infrastructure, and on whose terms? The application layer, as Shady El Damaty put it Day 1 from the Human Flourishing floor, has become a one-way valve. Nation-states pour capital, compute, and talent into a handful of frontier labs. The bleed-back is minimal. The labs hire fewer people, generate more value, and retain an outsized share of it. The pattern is already installed; Dario Amodei, CEO of Anthropic, warned in early 2026 that sufficiently powerful AI will put small groups within reach of engineered biological weapons. The attack surface grows with the technology. The safety architecture has not kept pace.

Whether the next decade produces the concentration Drago, Desai, and El Damaty diagnose, or the neutral and forkable coordination infrastructure Polosukhin and Juan Benet are trying to build, depends less on any single technology than on whether a community of the kind that was in the room, cross-disciplinary, cross-jurisdictional, willing to work below the application layer, can keep building in the face of being structurally out-capitalized by the actors building the alternative. On the evidence of March, they can. Whether they are given the time and capital to do so before the default settles is what the remaining pages of this report attempt to name.

Sheila Warren and the argument most of AI policy is missing.

Of all the arguments named across the two days, the one most likely to survive translation to audiences outside the room came from Sheila Warren, and it concerned a category of cost that technology designers have been trained to minimize for twenty years: friction.

Warren is chair of the Advanced AI Society and founding CEO of the Crypto Council for Innovation. Before that, she spent five years running the World Economic Forum's blockchain and digital assets work, the institution most responsible, rightly or not, for making blockchain and AI respectable to governments. Before tech, she was a lawyer. She studied at Harvard Law, began her legal career at Cravath, and in college and law school interned with the Innocence Project and the California Appellate Project, work she has said shaped how she thinks about due process long before she thought about AI. The through-line is relevant because it is what her argument is built on.

Sheila Warren · The Acceleration of Dehumanization · Day 1 keynote

Due process is OG friction. If the state of California will execute people, they had damn well better be sure it's the right person. Sheila Warren · In conversation

The argument, distilled, is this. The design instinct of the consumer internet has been to remove friction. Every bank-fee delay, every two-step verification, every warning dialog box has been treated as an obstacle to smoother user experience. The removal of some of that friction has been a real gain, a migrant worker sending remittances home pays a fraction of what they used to. But friction in systems is a category. It contains everything from inconvenience at checkout to due process at trial. Some friction is design debt. Some friction is governance.

The U.S. legal system, Warren argued, is running a real-time experiment on what happens when governance-level friction is removed. The federal government's current approach to deportation, moving people out of the country faster than the legal process has time to verify whether it is the right person, whether the action is lawful, whether a mistake has been made, is, in her reading, the due-process friction of constitutional law being bypassed in the name of speed. Civil liberties organizations across the political spectrum are now making this same observation. What Warren asks the technology community to notice is that the same mechanism is being engineered into the systems we are building for the next decade.

Agent-mediated systems, by default, remove human review from every step where review was previously embedded. An AI agent acting on your behalf can execute transactions, sign contracts, modify permissions, and commit resources with a latency and confidence no human process can match. The feature is also the bug. The same systems that let you pay your electric bill without logging in will, in a slightly later iteration, let an agent enter into recurring obligations in your name, on the basis of preferences you expressed once, in writing, at a moment when you were irritated and wanted the task done. Multiply this across every sector of the economy. The erosion is the structural tendency of the technology absent active design against it.

Warren's position is not that friction should be preserved everywhere. It is that we should be able to name which frictions we want to keep, and we should be naming those decisions in public, deliberately, before they are made for us by the default settings of the coding agents. Her phrase for this is differential friction. It is the first question, she argues, that any system designer should be asking: where is the friction, and what is it doing, and whose interests does it serve, and what happens if it is removed? She has not yet written the essay on this. She mentioned twice that she needed to.

Asked in a follow-up conversation how she would know the argument had been lost, Warren had a specific answer.

We're going to slip the minute that our kids understand that their teacher is AI and don't care anymore. Because that's how it's going to start. If the person you look to for authority on a topic — your teacher, your professor, your coach — is an AI, and you're just rolling along with that, that's a generation for whom you've already crossed the event horizon. They're going to start to think that's the case everywhere. Sheila Warren · in conversation

The point is not that every AI-mediated interaction is bad. Warren was careful to say otherwise. The point is that the distinction between a human teacher and a software teacher, between a human judge and a software one, between a human doctor and a software one, is an ingredient most people will not miss until it is gone, and the absence of which will have consequences that no current legal or educational institution is structured to notice, let alone contest. The slippage is already underway in customer service, search, and transactional commerce. Whether it reaches the rooms our children learn and make meaning in is a question that will be answered by the design choices of the next few years, not by any future regulation.

The argument sits under most of the other warnings the conference named.

Al Smith of the Tor Project described a version at the funding layer. Internet Freedom tools. Tor, Signal, Snowflake, reach about 45 million people in censored countries. They survived two decades on decentralized engineering and centralized funding. When the U.S. government cancelled its Internet Freedom grants mid-contract last year, the infrastructure stayed intact and the budgets disappeared. The funding mechanism was not distributed. It was centralized. That is the same pattern, a governance-level friction (stable, long-term support for civil-liberties infrastructure) removed in the name of executive convenience.

De Kai, a researcher with 45 years in the field and co-founder of the Empathetic AI Institute, described the same pattern at the level of cognition. Losing control of AI is not about the physical world, he said. We lost mind control 20 years ago. Recommendation algorithms have been deciding, for the better part of two decades, which ideas propagate and which communities form, and therefore which publics exist at all. The emergence of foundation models accelerates this. It does not create it. The cognitive commons have been under quiet capture for longer than most in the room had been building.

Evan Miyazono · Atlas Computing · Alignment Is Not Enough

Evan Miyazono opened Day 2 with the most widely-shared talk of the event. Miyazono is the founder of Atlas Computing; earlier in this community's history, he headed Funding the Commons before David Casey took over. His talk, Alignment Is Not Enough, walked the room through a thought experiment. Imagine an "alignment box", a room with an AI in it, and a person. The person enters, spends some time with the system, and walks out. The system is now perfectly aligned with that person's values. How many people think the problem is solved? Almost nobody in the room raised their hand. The reasons, Miyazono catalogued, were structural. Who gets to use the box? Who decides who gets to use it? If two people's values conflict, which wins? Can values even be logically derived from facts, the Humean is-ought gap? What happens when a person today is not aligned with the person they will be a decade later?

Miyazono's argument was that alignment is necessary and radically insufficient. Even if the technical problem were solved, the political problem would remain: whose values, and who decides. He reached for James Madison's Federalist No. 51, the paper that specifies why the American constitutional system divides power across branches, because any government that could control the governed also has to be structured to control itself. Miyazono's version for AI: if alignment were perfect, governance of alignment would still be the problem. His proposal was not a single technical fix. It was a 501(c)(6) non-profit into which the frontier labs could send delegates to agree on a shared model specification, a movable but legible document the rest of the world could see, contest, and amend. That proposal has been circulating in AI policy circles since March. It is one of the few that treats alignment and governance as the same design problem.

As intelligence gets cheap, agreement or consensus gets expensive. Figuring out how to do that part of the work much faster is going to be very important. Evan Miyazono · Alignment Is Not Enough

Esben Kran · Seldon Lab & Apart Research · Superintelligence Companies & Humanity's Future

Esben Kran opened a different register. Kran runs Apart Research and Seldon Lab, a Y Combinator-style accelerator for AI safety startups. His talk, Superintelligence Companies & Humanity's Future, took as its premise that superintelligence is probably arriving within a decade and worked backward. Given that, he argued, there are three ways a person who sees the risk can spend the next few years productively. One: join or found a frontier lab, hope to shape AGI from the inside, accept that this is a Hail Mary with humanity as the stakes. Two: lobby, protest, write policy, bet on second-order effects. Three: use the remaining time to build the specific pieces of secure infrastructure that a future decision-maker will need to pull, the levers that, if they are not already built, cannot be used at the moment they are needed. Kran's case was that option three is the most under-chosen of the three, and the most tractable. Seldon's portfolio is built on it.

And Ryan Kidd of MATS put one figure in front of the room that did not leave it. The share of people working in AI who are focused on safety and security is roughly 0.1 percent. Getting to one percent would matter. Getting to what cybersecurity represents as a share of the tech workforce would require a hundredfold increase. The funding exists. The people do not. Paul Riechers of Simplex made the companion point at the research level: the science of understanding what AI models are doing from the inside, interpretability, mechanistic analysis, receives a fraction of what flows into making them more powerful. The ratio of capability investment to interpretability investment is the ratio at which we are choosing, structurally, not to know.

The warnings do not conflict. They describe a single pattern at different scales. Governance-level friction, the review steps, the verification processes, the waiting periods, the standards bodies, the capacity to understand what a system is doing before it is allowed to do more, is being removed from domains that depend on it, at a speed that outpaces the institutions that would normally slow it down. We are choosing this. We are choosing it by not choosing.

Sheila Warren · off-stage interview

The governance layer of AI has no institutional home. That is the white space this community is sitting next to.

Across the two days, one observation kept surfacing without anyone quite committing to it.

There is no distributed, multi-stakeholder research and venture lab whose primary mandate is the governance layer of AI. There are centralized labs working on alignment. There are academic centers (GovAI at Oxford, the Ada Lovelace Institute, the Collective Intelligence Project) working on policy and democratic experimentation. There are decentralized compute and training networks (Bittensor, Gensyn, Prime Intellect). There are think tanks, advisory bodies, and standards organizations. None of them sit where the gap is.

The gap is this. As AI agents enter economic and governance contexts at scale, they need infrastructure that does not yet exist: privacy-preserving compliance, verifiable proof-of-personhood distinct from proof-of-agency, programmable funding mechanisms with audit trails, identity systems that survive both surveillance and impersonation, coordination protocols that cannot be unilaterally frozen by the jurisdiction they happen to route through. Each primitive exists, somewhere, in prototype. None of them have been knit together into a coordination infrastructure of the sovereign stack that a mid-sized democracy could deploy, one that cannot afford to either surrender its AI infrastructure to a foreign hyperscaler or build it all itself from scratch.

Building this layer is not lucrative in the timeframes venture capital is built around. The buyers, governments, multilaterals, civil society networks, communities, operate on 18-to-36-month procurement cycles that most seed-stage companies cannot survive. The value accrues over decades. And the work is technically demanding enough that the teams capable of doing it are the same teams being actively hired away by frontier labs.

The conference did not announce a solution to this gap. The organizations represented in the room. Funding the Commons, Protocol Labs, NEAR, the Internet Archive, UNDP AltFinLab, UNICEF Office of Innovation, Octant, the Tor Project, Holonym, and the networks of researchers and founders that have been circulating through FtC events for five years, sit unusually close to it. The partners are in place. The convening record is real. The cross-cultural footprint, from San Francisco to Mumbai to Buenos Aires to Zurich to Bangkok, is rare in infrastructure communities. Whether the gap gets filled by a new institution built for it, or whether it gets filled piecemeal by founders deploying against fragments of the need, is an open question. What is clear is that the window for filling it before the default trajectory consolidates is narrower than it looks. Measured in small numbers of years, not decades.

This observation is why the event mattered, and why a conference that reads on paper as another AI event is, in its substance, a conference about whether the political and institutional architecture of the 21st century gets written by a small number of centralized actors with the present capital advantage, or by a larger, messier, more federated set of builders working on the assumption that concentration is a design failure rather than a natural outcome.

Juan Benet's charge to the room

Juan Benet's closing keynote was the most direct prescriptive statement either day. In twelve minutes without slides, he told the community what the next two years of the work need to look like if it is going to matter. Five charges:

One. Become AI-native, immediately. Not as a talking point. As a working practice. Every project the community is running should now be asking whether an AI agent can compress it from months to weeks, from committees to a few LLM calls, from prototype to deployment. The community has spent five years developing ideas for coordination infrastructure; the tools that now exist can build those ideas at a speed and scale that was impossible a year ago.

Two. Prepare for autonomous agents as participants, not tools. Juan Benet's working expectation is that the first fully autonomous AI corporation launches within a year, and the first AI "person", an entity legal systems will have to treat as a participant rather than an instrument, within two. The coordination mechanisms the community builds need to be designed for that world, not retrofitted into it once it arrives. FtC gatherings two years from now, Juan Benet said, will likely include AI agents as meaningful participants in the room.

Three. Use AI to explore the design space aggressively. Simulate participants. Simulate adversaries. Attack your own constructions. Test mechanism designs at a scale that no governance pilot has ever been able to reach, because the cost of each test has collapsed. The community has the rare combination of technical fluency and substantive problem to make this genuinely productive. Do it.

Four. Let go of projects that will not matter in five years. One-shot software generation means that many of the specific tools this community has been painstakingly building by hand are about to be obsoleted by any competent coding agent. The question to ask of every project: will this matter in 2030? If the answer is no, move on. The ratio of effort to outcome changes when the tools change.

Five. Be honest about the reset. Five years in, the public-goods movement has not produced a dominant solution to the problem it set out to solve. Juan Benet said this on stage, directly, without hedging. The point of the admission was not to discourage the room. It was to release it from a trajectory that has not worked and free it to try a different one. The tools for a different trajectory exist. The community has the relationships and the technical depth. What is missing is the decision to operate at the new speed.

The Argument Map at the front of this report is what answering Juan Benet's charge looks like in specific work already underway. The right column on every row is real, undercapitalized, and unfinished. The bridges are narrow. The traffic on them is not yet one-way in any direction. But each line is named work, by named people, against a named failure. What the community needs is not more ideas. It is the capital and the institutional cover to scale the ones that already exist, fast enough to matter.

The stakes, if there is any chance of describing them without sounding grand, are these. The people in this room have children, or will have children, or are the children their own parents built the current institutions for. The question of whether the next decade's AI systems are built by a few actors on terms most of the world cannot contest, or by a wider set of builders on terms that can be examined and amended, is not an abstract policy question. It is a question about what world our kids inherit, whether they are learning from humans or from software trained on data they cannot see, whether they participate in economies where their contributions are legible or are absorbed into systems they cannot verify, whether the institutions that protect them still function at a scale their lives are lived at. The argument of this essay is that the shape of that inheritance is being decided now, quietly, by the people who understand the infrastructure well enough to see the decisions being made. The purpose of reporting on them is to surface those decisions to a wider audience before they calcify.

Three gatherings in the next six months, and the specific kinds of support this movement needs to continue.

Funding the Commons returns to San Francisco in October 2026. Location to be confirmed. The framing will continue to develop from where this event left off.

Between now and then, two further gatherings.

In Paris, in June, the next DPI Leaders' Dinner. Under Chatham House Rules, thirty to forty policymakers, researchers, and infrastructure builders sit down to the questions this report has only begun to sketch. Prior dinners have convened voices from the IMF, the World Bank, the Ethereum Foundation, UNICEF, and the Open Society Foundations.

In Kenya, in the fourth quarter, the UNDP × Funding the Commons Builder Residency. A multi-week embed for international and regional founders working on live deployment challenges sourced from UNDP Country Office problem statements, digital identity, cash transfer optimization, climate finance verification, programmable allocation for humanitarian response. The residency model is deliberately close to the work. The technology gets proven inside the deployment environment rather than pitched into it.

What kind of support this movement needs to continue

This is a movement with a specific argument and a specific deficit. The argument is in the preceding pages. The deficit is one of capacity: the work described above is technically demanding, institutionally unglamorous, operates on timelines venture capital is not built for, and is structurally undercapitalized against the centralized alternatives it is trying to interrupt. Four concrete forms of support would change the pace materially:

Speakers and contributors whose work answers a question this community has not yet been asking well enough. Founders, researchers, policymakers, and practitioners working below the application layer on identity, compliance, coordination, or programmable allocation. If your project sits at one of the bridges in Section III, write to us. You should be on our stages.

Sponsors and institutional partners who read Section III and recognized a white space in the shape of their own mandate. Governments, multilaterals, philanthropies, and mission-aligned capital whose work depends on infrastructure that does not yet exist. We have five years of experience convening the people who can build it.

Researchers, writers, and journalists willing to treat this community as a subject of serious long-form inquiry, in languages and publications outside the Anglophone technology press. This essay is built almost entirely on the substance of the talks. Much more substance is available, and much of it deserves more careful treatment than a two-day event or a single report permits.

Academic institutions, especially outside Europe and North America, interested in joint research, residencies, or long-form partnerships on the governance, funding, and coordination infrastructure questions raised here. If your faculty or students are already working on adjacent problems, we would like to talk.

The most useful way to reach us is to send a paragraph, not a pitch. Say what you are working on and why the thesis of this document does or does not describe it well. We read, we respond, and we are especially interested in critiques.

The weekend, on the record.

Intelligence at the Frontier took over Frontier Tower on Pi Day weekend, March 14 to 15, 2026. Ten floors held active programming. The second floor ran the mainstage. Upper floors held tracks on AI and autonomous systems, Ethereum and decentralized tech, longevity research, neuro and biotech, arts and music, immersive media, maker practice, flourishing, and a d/acc lounge. A hackathon ran in parallel across the same weekend. The Commons Platform, a persistent home for session recordings, speaker profiles, and community, launched publicly at the event.

A floor session, hands up. The conference ran across ten floors of Frontier Tower simultaneously, with the energy spilling out of the mainstage and into the building.

The hackathon. 145 builders submitted 53 projects across four tracks (physical AI, agentic funding, AI safety, sovereign infrastructure) over 36 hours.

Floor 6 ran nighttime music programming alongside the conference. The weekend ran on multiple registers at once.

Mili · the open-source robot built by Solo Tech, playing piano during lunch on the mainstage

About Funding the Commons

Funding the Commons began at Protocol Labs, the organization behind IPFS and Filecoin, in 2021, gathered around a specific, narrow question: how do we sustainably fund open-source software and the other shared infrastructure the digital economy runs on? The early meetings brought together researchers, open-source maintainers, and the small number of funders actively experimenting with new mechanisms (quadratic funding, retroactive funding rounds, impact certificates, hypercerts). The work was technical and often obscure to outsiders, and it mattered because so much of the infrastructure the rest of the economy depends on is produced and maintained by people whose work has no clear funding model.

Over five years, the community's work became applicable to a much broader set of problems. The same coordination-and-funding primitives that work for open-source software are the primitives the agentic economy, the sovereign-AI stacks, multilateral humanitarian finance, and city-level innovation policy all now need. The conversation has expanded accordingly. FtC's 2026 center of gravity sits where coordination infrastructure, AI governance, and the systems that will determine how power and value flow through the automated economies of the next decade. The community is still doing the foundational work on funding mechanisms; it is now also doing the work on what those mechanisms are funding, and for whom.

FtC has convened conferences, residencies, and leaders' dinners in San Francisco (three annual editions), Buenos Aires, Bangkok, Tokyo, Taipei, Berlin, New York, and elsewhere, with partners including UNICEF's Office of Innovation, the Internet Archive, the Tor Project, NEAR, Berkeley's Center for Responsible, Decentralized Intelligence, and the UNDP. FtC holds a seat on the UNDP Blockchain Advisory Board. Neutral ground.

Watch and explore

Intelligence at the Frontier Tower, FtC San Francisco · March 14–15, 2026 · Recap Film.

Juan Benet's closing keynote: "Thoughts for the Funding the Commons community in 2026." From the closing keynote, you can navigate the playlist to all 39 sessions.

Photo library · the full photographic archive of the weekend, on Google Drive →

References

• Brey, Philip. "Disruption and displacement in the ethics of emerging technology" (2022). Discussed in Adam Russell, Day 1 keynote.
• Casey, David. Comparative Floor Governance at Funding the Commons SF. Funding the Commons working paper, 2026 (forthcoming).
• Drago, Luke, and Rudolf Laine. The Intelligence Curse. Workshop Labs, ongoing essay series, 2024–2026.
• Leo XIII, Pope. Rerum Novarum. Vatican encyclical, 1891.
• Ostrom, Elinor. Governing the Commons: The Evolution of Institutions for Collective Action. Cambridge University Press, 1990.
• Survival and Flourishing Fund. The S-Process methodology, developed by Andrew Critch. survivalandflourishing.fund.
• Vaswani, Ashish, et al. "Attention Is All You Need." Advances in Neural Information Processing Systems 30, 2017.

Partners, team, and credits

Sponsors: Protocol Labs, ElevenLabs, Bittensor, Activeloop, DeepLake, Solana, VESSL.AI, Velda, Filecoin Foundation, NEAR, Hypercerts Foundation, Octant, Solo Tech, Nomadic ML, human.tech, Optimized Foods. Partners: the Tor Project, Internet Archive, Foresight Institute, Advanced AI Society, Artizen Fund, Flourishing Systems Foundation, UNICEF Office of Innovation, Founder Institute, ETHSF, Metagov. Floor and experience: Quorum1, Biopunk Labs, Viva.City, UFB, Ethhouse.xyz, GTM at the Edge, Learning Layer Labs, Tea Tribe, Rich DDT.

The San Francisco event was convened and produced by the FtC core team: David Casey, Kim Buisson, Tereza Bízková, Ira Nezhynska, and colleagues, alongside an extended network of advisors, volunteers, and partner institutions. None of this would exist without Juan Benet, who has supported Funding the Commons since its earliest days at Protocol Labs. Special thanks to Evan Miyazono, who led the organization before David Casey and returned as a Day 2 speaker in March, and to the community members (Open Source Observer, the Hypercerts Foundation, and many more) who have kept showing up year after year. This field report was inspired by David Dao of Protocol Labs Research, whose work on collective intelligence shaped its framing.