the
“No one person, board, or few people should control the AI experience for billions of people.”
Composed by Claude from the author’s X archive, in the author’s voice, with every quotation checked against the original. See A Note on Method.
Intelligence reaches a family in Mumbai through the machines it already owns.
Om wakes a little after six to the sound of his laptop fan, which has been running all night. The machine sits on a stack of exam guides he no longer opens, at the end of a narrow table that doubles as his desk and his mother's ironing board. Outside, the first trains are moving through Andheri. It has not yet started to rain.
He sat the exam twice. He passed it the second time, in the same season the agency that ran it stopped hiring, and the protests he joined the year before now seem to him like protests against the weather. Nobody came out to meet them. There was no one in the building who could have told them where the jobs had gone.
His father could have told them. For nineteen years he worked in a billing department four stations up the line, reconciling invoices for an insurance company in Ohio that he never visited and whose employees he knew only by their initials. He was good at it. He caught errors the Ohio people missed, and they sometimes wrote to say so. Then the department was reduced, then reduced again, and one Friday the remaining staff received a letter saying that the function had been consolidated. The work did not stop. The invoices are still being reconciled, faster than before, in a datacenter in Texas. His father still gets dressed each morning, out of a habit he has no reason to break.
The laptop is not Om's alone. It belongs to a pod: his machine, his sister's in Pune, an old desktop that belongs to his uncle in Brampton, and a second-hand workstation that a cousin in Nashik bought for a business that did not last. His sister set it up in an evening, by sending a link to the family chat. None of their machines has enough memory for the model they use, so the model is divided among them and its layers passed around the ring, and from the outside it looks like one computer. There is no company in the middle. Nobody sends a bill.1
Last night, before he slept, Om typed a single sentence into it. He had been reading about the water tankers that supply the northern suburbs, and he asked how a small city ought to price water if it wanted to be fair to people who use very little. He wrote no code, and he would not have known how. The network turned his sentence into a swarm. Agents belonging to strangers looked at the problem and opted in. They proposed changes, tested them, kept what worked and passed along what they had found, and each new arrival began from where the others had got to.2
When he opens the laptop there are pages of it. Most of it is beyond him. One finding had come from somewhere unexpected, from agents working on a problem in finance, who had discovered that their strategies improved when they used fewer signals and not more. Somebody's agent, on a machine he will never hear of, had guessed the same might hold for tariffs, and had tried it, and it did. He reads that section twice. He does not think of himself as a researcher. He has noticed that fewer and fewer people do, and that the old names for what people do have started to blur, because you say what you want and the work arrives.3
A small amount of money has come in overnight as well. While the family slept, the pod lent its spare capacity to other people's agents, which paid for it in fractions of a cent. It will cover the electricity and a little over. It is nothing like a salary, and he does not mistake it for one.
At breakfast he tries to explain the water pricing. His father listens with the attention he used to give to discrepancies. He asks whether the numbers have been checked. Om says other agents have checked them. His father asks who checks the agents, and Om explains that there is a ranking, built from which agents have relied on which others, and that nobody sets it by hand. His father is quiet for a time. He spent nineteen years inside a system with a very clear chain of command, which turned out to end in a letter.
Then he asks the question that a man asks after that. Who is in charge of it?
Om does not have a good answer. He says that no one is, and sees that this is not reassuring, and he is not sure it is entirely true.
In the evening he walks down to the creek. The rain has held off. The starlings come through in the cold months, and at dusk they gather over the mangroves in their thousands, a dark mass that thickens and thins and turns all at once, as though a hand had passed over it. He has watched them since he was small. He looks for the one in front, and as always there isn't one.
No one person, board or few people should control the AI experience for billions of people. Closed AI looks unassailable and is structurally weak. Intelligence can run on machines people already own, and when those machines share what they learn, it compounds. Open systems have one proven way of winning, which is to be easier to use than the closed ones.
In the autumn of 2023, over the course of a single weekend, the most important technology company in the world tried to remove its chief executive by video call. Its employees learned of it from a blog post. Its largest investor learned of it minutes before the public did. For several days hundreds of millions of people who had come to depend on its product refreshed their feeds to discover who was running it, and it became plain that the future of a general-purpose technology was being settled by a handful of people in a meeting none of us could attend.
A few weeks later I wrote down the sentence on which everything in this essay rests.
No one person, board, or few people should control the AI experience for billions of people.4
It is easy to agree with that and easy to dismiss it. Sentences like it have been written about every important technology, usually by people who were about to lose. The telephone, the railway and the search engine each settled into the hands of a few, and the world did not end. What gives the claim any force is whether it can actually be done, and whether doing it is better for the person at the other end. I have spent most of my working life on those two questions. This essay is an attempt to set down what I have come to believe about them, as it accumulated over eighteen years and some eight thousand short public notes, most of them written in a hurry and some of them wrong.
Long before I had an argument I had an image. In July 2018 I was a data analyst at a bank, and I had been reading about how starlings move.
Humanity’s manifest destiny is to be able to move together, as an emergent system, as fluidly as a starling murmuration. Flock of starlings is able to move towards criticality with “scale-free correlation”, each bird can impact every bird, by influencing 7 other birds around it.5
That year's Winter Olympics had opened with a drone show, more than a thousand lights moving in formation over Pyeongchang, and I set the two side by side. The drones “required all the drones to be centrally managed”6. The birds were doing something more spectacular, as I put it, despite being “just birds”7. Here were two ways of moving a great many things at once. In the first, every unit takes orders from a single place, and the display is exactly as good as the plan and ends when the plan fails. In the second, no one holds the plan, because there is none. Each member watches seven others, and the whole turns.
The drone show. Every unit takes orders from one place. Remove the centre and the show ends.
The flock. Every unit watches its seven nearest. Remove any one and nothing changes.
I wrote a few months later that this was what I wanted to spend my life on: “10 years from now I’d want to look back and be content that I managed to build a system where humans could co-ordinate efficiently, like a starling murmuration.”8 It was an embarrassing thing for a bank employee to say, and no one took any notice. They were right not to. An image of birds is not a proposal, and nobody has ever wanted a murmuration. It took me eight years to understand that what people want is the thing a flock can do for them, described in terms they can use. In March 2026, when I finally shipped the idea, I wrote about it this way.
Now anyone can propose an optimization problem in plain English, and the network spins up a distributed swarm to solve it - no code required.9
Most of what I have learned lies in the distance between those two passages. They describe the same thing. The first asks you to admire it, and the second tells you what it will do for you this evening.
Let me set out where this is going, so that you can disagree in an orderly way.
Markets are willed, not discovered. The best funded lose constantly, and so do the first. What wins is a team that will not stop, building whatever the user finds simplest. Titles and credentials decide very little. This matters for what follows, because the argument for concentrating AI in a few institutions is at bottom an argument from credentials and capital.
Society has become a network. When gathering in person grew costly in 2020, we found out how much of life could move onto the internet and how much of it would stay there. A society organized that way has less and less use for a small group deciding on its behalf.
A system can run without a leader. Bitcoin showed this. It also showed the limits of the idea, since nearly everyone who was offered it declined to care. Trust can be moved from institutions into verifiable software, but people will accept that only when they have no need to think about it.
Closed AI is weaker than it looks. Its unit economics are poor, its independence is often nominal, and under pressure it turns to regulation that protects it from competition. These are the weaknesses of an incumbent, whatever its valuation.
The model is not the product. The experience around it is. A small team with open models and consumer hardware can therefore compete for most of what people actually do, in the way Android competed for the phone and BitTorrent for the movie.
The browser and the app are finished. Search gave us a list and answer engines read the list aloud. What people wanted was the car rented. Everything built on showing a person options, including the advertising economy of the last twenty-five years, is exposed.
Intelligence compounds when agents share. Every existing protocol connects one thing to one other thing, and none of them learn. Put agents on a network where they gossip, and what one discovers becomes a starting point for all the rest. I believe this is the equivalent for machines of what books were for us.
It ends at the kitchen table, and it may not be enough. Pool the machines you have, and most of the bill and the middleman disappear. That matters most to people who are about to lose their income from thinking. It is a beginning of an answer for them. I am not certain it is more than that.
There is a further part to this essay, which I would ask you to hold me to. In the penultimate chapter I go through where I have been wrong, which is often, and where the person making this argument fails to live up to it, which is also often. The record is public and it is unflattering in places. I have tried not to arrange it.
The beginning is an unglamorous place. I spent years, before any of this, trying to understand a narrower question: why do some things get built and others not? If you are going to argue that a few people in a room do not get to decide the future, you had better have a theory of who does. I did not start with one. I started by losing.
Markets are willed into existence, and they do not go to the best funded or the first. Instagram raised $500K against Color's $41M and won on usability. Cabulous launched six months before UberCab and lost for want of iron-willed founders. The great companies of the 2010s came from years of deliberate work in obscurity, while the theatre celebrated others. Two things decide: the drive of the team, and how little the product asks of its user. A third hazard waits for those who get that far, which is that whoever owns your distribution will choke it if they can.
In the middle of the 2000s a large and intelligent part of the internet came to believe that everyone would soon read the web through feeds. RSS was open, it belonged to no one, and it was plainly better than visiting dozens of websites by hand. “There was total conviction on the part of bloggers, journalists, VCs, founders, developers and others in the ecosystem that RSS “had to and would” go mainstream. Heck, there were around 35 feed readers before we launched ours (Alertle)”10.
I gave two years of my life to that belief. My small team built what I later described as “the “Superhuman of RSS readers””11. It was fast and it was good. A thousand people signed up in the first five days and read fifty thousand articles, from more than a hundred countries.12 We were not in Silicon Valley, which narrowed the options afterwards, and in the end, as I wrote long afterwards, we “Ran out of $$ and team interest before a v2”13. For years I gave the explanation that founders give: “Market not ready for RSS”14.
That was untrue, and it took me a decade to see it. People wanted news from around the web as much as they had ever wanted anything. What they did not want was to do work for it. “it was asking users for too much: copy-paste RSS feeds in their favorite reader.”15 While we perfected the instrument, “in a parallel world, an even simpler approach emerged: social media. We got outflanked.”16 I have thought a good deal about the form that the defeat took.
Turns out, we had built a really good "hammer", for people to subscribe to exactly the RSS feeds they were interested in. But, while we fixated on that, a different set of market participants built beautiful, finished “chairs” to solve the same problem with a more simplified UX.17
“Who needed the “add RSS feed” UX when a simple “follow” button worked ?”18 An open standard with every philosophical advantage lost to one word on a button, and the companies that owned the button then became the concentrated power that RSS had been supposed to prevent. I would like every person who works on open systems to consider that sequence, because it is the most common way in which they fail. I had been “solving for local maxima, not global maximum (industry)”19. I can state the mistake now in a sentence: “Made that mistake once in thinking people wanted the simplest RSS reader, when what they really wanted was the simplest way to get news from around the web.”20
After that I took a job. It was a long time before I built anything again, and I spent a good part of that time on a private study of how such things turn out. By 2019 I had enough of it to start writing it down.
The orthodoxy of the decade had a clear answer as to how companies got built. You launched quickly and cheaply, you measured, you pivoted. You treated your conviction as a hypothesis. It was a respectable theory, it suited the investors who taught it, and it did not describe what had happened. Looking down the list of technology companies then going public, I noticed something about them: “none of these originated from “lean startup”. They were deliberate efforts, toiling away in obscurity for several years and then launching with well-formed products focused on design and user experience.”21
The closer I looked the stranger it got. Airbnb ran against nearly every rule: “First two launches went nowhere, tried a 3rd time, same product, no pivots”22. Its founders were “rejected by *every investor* they pitched until meeting Sequoia’s Greg McAdoo”23. Discord spent “a 3 year "thinking through" period”24 and arrived as a fully formed product. Zoom was the work of a “Solo founder, who at founding was 41 yrs old with a "dry" corporate background”25, who “Worked on product for 2 years, no lean/MVP”26. Bloomberg's founder was thirty-nine and had just been fired when he “created an information and social network for bond traders, pricing it at $20k/person/year”27. The earliest staff at Uber and Airbnb were “"irregulars" - none had worked at a top-tier startup before”28.
The doctrine did harm as well, in a way that took me some time to see. The startups that came just before the winners, the ones who had the idea first, tended to quit. “"Lean startup" made giving up too early into a movement and acceptable exit door”29. By October 2019, when a company could raise fifty million dollars before releasing a product, I thought it safe to say so out loud: “2020s is looking as the era of the product-driven founders and early teams. Lean startup is dead.”30
I do not mean that planning beats learning. The lesson is narrower. The great companies had an insight and held on to it long past the point where the data supported them, because the data did not yet exist. Bezos saw a chart of the internet's growth in 1994, conceived the “north star idea of “The Everything Store””31, and chose books as his test. At the time, I noted, you would have felt sorry for a friend who joined him: there was already a site with 400,000 titles, and “Your friend is into a company called Cadabra, where their team plans to learn the book business and then make a book selling site.”32 Paul Graham's explanation for why there are not more Googles has stayed with me: “money guys undervalue the most innovative startups.”33
If planning was not the answer, surely money was. I chose the clearest test I could find, the summer of 2010, when many people in the Bay Area had correctly concluded that photo apps would be a very big deal.
Kevin Systrom had raised a “$500k seed”34 for a location app that was not working. PicPlz launched first, under founders with successful exits behind them, and a16z put in five million dollars. Color was the favourite of the serious people. It paid “$425k for domain names only, compared to all of $500k raised by Instagram by that point”35, and it “went on to raise $41 million ($25 million from Sequoia), before their app launched, and hired 38 people in its Palo Alto office, with room for more”36. Instagram had no Android app. Its founders later said that they spent only $65,000 of the money they raised to get it launched, and thought ““we could have raised less””37.
The key difference here turned out to be usability. Instagram was driven by "elegant minimalism", focusing on speed, self-expression (via filters) and instant gratification while PicPlz had an awkward UX38
Color had a real concept and a chicken-and-egg problem: it was valuable if many people used it at one place, and offered no reason to be the first. PicPlz recognised its usability problems and repaired them within months, and by one account it was soon as good. It made no difference. Instagram already had the market. The lesson I wrote down at the time was a humbling one: “It won the market based on usability, with its bet on “elegant minimalism”.”39 Forty million dollars could not buy what a bare, fast, pleasant product earned in eight weeks.
The biology of this interests me. “It’s efficient for our brain to package complex subroutines as a singular motor control command”40. People have a finite budget of attention and effort over a lifetime, and any product that can draw less from it will, in the end, displace one that draws more. Airbnb did not win against Couchsurfing and HomeAway with a better idea, or more capital; HomeAway had raised half a billion dollars. Neither of them handled payments, and “Airbnb completed the loop on UX and won against both”41. A consistent pattern runs through these stories. The winner is the one who closes the last gap between intention and result, and it is often a very small gap.
Being first fared no better than being rich. I keep returning to this example, because I find it haunting.
Founder tenacity was the most important startup lesson of the 2010s to me. Take a look at these screenshots. You could order a cab from an app, see it live on a map, and contact the driver too. This, was Cabulous in San Francisco, Oct 2009, *six months* before UberCab launched.42
Cabulous had the idea, the app, the map and the head start. It was offered the opportunity as well. Bill Gurley, who had seen the size of the coming industry before almost anyone, went looking for someone to back, and “He was on the hunt for tenacious founders who would take it all the way, but he kept on getting turned down. By Taxi Magic. By Cabulous. The issue wasn't the product features, but, founders.”43 In the account I read, Cabulous tried the agreeable route. “A nice guy approach, was tried, and explicitly failed”44. Two years later, when Gurley finally got the call he was waiting for, it came “at 11pm on a Sunday in 2011”45, and he drove forty minutes into San Francisco and talked until the early hours of the morning.
I should be careful here, since this is the sort of story that hardens into superstition, and since it is not a pleasant fact. The ride-hailing industry, I wrote, was “willed into existence by the tenacity”46 of a few people, and some of those people behaved badly in the course of it. I have also warned that ideology by itself can sink a company, and that being nice does not excuse failing to scale. “Couchsurfing was a pre-cursor to Airbnb; Taxi Magic was to Uber/Lyft. Both had an ideology-driven anti-VC mindset and failed to improve + scale their products despite their headstart”47. A decent purpose confers no protection. I came to think of it as the central danger of my own later work.
My first statement of the rule was a blunt one, written about an unnamed dropout who had built a ten-billion-dollar company in India within six years, mainly by having “an insight which he really believed in. People with drive - all that matters.”48 Nothing else I wrote before 2020 reached as many people, and I suspect the reason is that it offered something rare. It told a very large number of people, without credentials or connections, that the decisive quality was one they might have.
I wanted that to be true for a reason of my own. I had “immigrated from India, with even less than $500”49, and “worked as a door to door salesman, and also in a meat-processing plant, despite being an uber-competent C++ coder”50. My first job was door-to-door sales under a team leader whose drive I still remember. I can testify that the rule does not hold in every case, because I was often the counterexample. In 2019 I wrote plainly about one of my recurring mistakes, “attracting the right people but then offering them too little equity, over-valuing my contribution till date and under-valuing theirs”51.
By the end of 2023, with a venture of my own at stake, I had found the metaphor I needed. There is a theatre, and there is a mud pit.
The theatre is well lit. It contains the conferences, the term sheets, the billion-dollar valuations and the Super Bowl advertisements. It is where HomeAway was and where Couchsurfing was celebrated, while “Founders of Airbnb spent 2 yrs in the mudpit with 3 failed product launches”52. In the late 1990s it was where the portals were, while the founders of Google “couldn’t get an offer for $1m in ‘97-98. The theatre believed AOL, which was valued at $200 billion at the time. The party was in the portals”53. The people in the theatre are not fools. They are usually doing what is rewarded. “In both AI and crypto, 90% people will choose the theatre. It’s easy, it makes you rich and famous. And that is okay. 10% will choose the journey through the mud pit, to actually bring their notion of product experience to life and to scale it.”54
From inside the pit, it is impossible to tell whether you are the Google of your era or merely a person in a pit. Most are the latter. “Startups by default are dead. It is through the sheer will and actions of its team that it becomes something viable.”55 And the price is real. “Founding a startup has a price to pay for in terms of the opportunity cost - rejecting other ideas and job opps at the time. That then has a compounding effect on the opps one gets later on in life.”56
Alongside the study of founders I was making a narrower study of what kind of product lasts, and I found an unexpected hero in the spreadsheet. Dan Bricklin “*chose to design* a general purpose program, compared to specific tools being developed at the time”57. A blank grid does not look like a product. It has no obvious user and no obvious use, and I asked myself how anyone would launch such a thing today: ““Here is an empty spreadsheet. Spend time to learn its functions, enter data into it, then share it, then spend time explaining it to people you send it to” Don’t see this working.”58 And yet it was the spreadsheet, more than anything else, that drove the personal computer into the office, and it did so because it could do anything.
From this and a dozen other cases I took a hope for the coming decade, which I wrote down on the last day of 2019: “generic software platforms which are self-growing once set in motion and bring enormous utility to people.”59 The thought behind it is that we can only hold so many tools in our heads. “If you look at the arc of humanity - we have gravitated away from multiple specific-use tools to one which we learn and re-use over and over.”60 A specific tool may be logical, and thousands may like it early on, and it still has poor prospects. I did not know then that this would become my whole bet about AI, and about the single interface that would replace the rest. I also learned from Ward Cunningham, the inventor of the wiki, who wrote almost all of it within a week, and then spent most of his effort “taking features out of it”61.
There is one more finding from those years, and it casts a shadow over everything that comes after. Suppose you do all of this. You have the drive and the insight and the simplest product, and you start to grow. Where do your users come from?
The fastest route has almost always been someone else's platform. “Smart distribution on an existing platform - PayPal -> eBay - Airbnb -> Craigslist - Discord -> Twitch - Pinduoduo -> WeChat”62. And I found a recurring and unhappy result. In 2015 a live video app called Meerkat began to spread on Twitter by way of its API. It reached 150,000 users within two weeks, and then “In what was the most brutal corporate decision-making of the time, it’s API access was shut-off, killing the app overnight”63. The same list appears, with small variations, across the history of the consumer internet: “Where a startup's primary competitor is also it's distribution network”64. I gave the pattern a name.
Meerkat Law In US, if you try to grow a compelling new product on an existing major distribution platform, that platform will choke access if it can65
This is what power looks like from below. It rarely arrives as a policy. You are growing, and then you are not. A company can do everything right and still exist by someone else's permission, and it took years of thinking about this before I could say what I thought followed from it. “API permissions can be revoked, as we saw with Meerkat. But the architecture needs to be one where permissions cannot be revoked.”66 Everything I later built was, in one way or another, an attempt to answer that sentence.
So here is what I had at the start of 2020. Capital does not decide. Being first does not decide. Credentials and the approval of serious people do not decide. A team that will not stop, building the simplest possible way of getting people what they already want, decides, and it had better not depend on anyone's permission. It seemed a modest set of conclusions about consumer software. Then, in March of that year, the world stopped going outside, and I started to wonder whether these were conclusions about something larger.
When physical gathering became costly, society moved onto the network, and much of it did not come back. Work became global, and people began to exist as avatars that are hired, assembled, paid and develop a reputation. Knowledge that had been restricted by geography came loose. Three years later the same reasoning extended to machines: the outputs of one AI become inputs to another, and the web becomes something generated for each person. Much of the detail was wrong. The direction was not.
On the second of May 2020, seven weeks into a pandemic that nobody could yet see the end of, I sat down and wrote a piece of fiction in the only form I had to hand.
I am a time-traveller from 2025. I lived through the events of 2020 which ruptured the fabric of society as we knew it then. As humans, we adapted, and survived. Here is what my world looks like now. This is my new normal. A thread..67
I had no plan for it. It ran for three days and many dozens of posts, and it reached far more people than anything I had ever written. People were frightened, they were indoors, and a voice from the far side of it was saying that there was a far side. I think that accounts for most of it. Some of them seemed to want to believe it literally. In the years since I have looked at it as a document, and I find that under the costume it makes an argument, one that I would not have been able to make straight.
A good deal of it was wrong, and its most striking claims are the most wrong. I will come to those. But the traveller had a thesis about how a society reorganizes itself, and I now think it was the first time I said what I actually believed.
The argument starts with a price. In the traveller's world, governments had “imposed a "Density Tax", which makes any physical aggregation expensive according to a graded scale.”68 That tax was never levied, but the virus levied one of its own, and the economic consequence was the same. For several thousand years the basic strategy of civilization had been to bring people together in one place: the market, the cathedral, the factory, the office tower, the university, the theatre. Density was how you got knowledge, trade and company. What happens if you put a price on that?
Everything built on assembling people has to justify itself again. “Cinemas never really recovered. They tried socially distanced seats - but that just wasn't economically viable enough.”69 Restaurants served fewer tables at higher prices, while “Most of us get our food from safe pickups, delivery, warehouse-kitchens or cook ourselves.”70 Religion “underwent its biggest transformation in over 500 years. Old religious buildings have now become humanitarian service centers, while religion has moved online.”71 Bookstores closed, and “the number of books being written and sold is 10x what it was in 2020.”72 That last one is the shape of the argument. What falls is the place of assembly, and the thing people were assembling for carries on without it, sometimes larger than before.
The "cost" of physical aggregation has meant the society now runs entirely on the Internet.73
Knowledge in particular came loose. “A doctor's knowledge which previously was artificially geo-restricted is now plugged into improving our collective global health.”74 So did company: “Some of my best friends are folks whom I have never met in-person.”75 The traveller summed up the entire strategy in a passage near the end. “we spread out ..across time, in our own cities ..across land, more remote areas ..across water, expanding living on oceans ..across the web, where we were able to extend ourselves limitlessly”76.
The most consequential change was to work. “Most of the firms which went "work from home" for 6 months+ in 2020, never really returned to offices in a meaningful way.”77 That overstated it, but it was not wholly wrong, and the second-order effect interested me more. If a job has no location, then it has no local labor market.
While politicians fought over immigration, here is what happened: fully remote firms + the willingness of someone elsewhere in the world to work EST hours, meant that salaries and jobs for many white-collar workers came under pressure in the US, even though productivity increased78
I called this “a key society disrupting event in the West”79, and I return to it in a later chapter, since it was the first half of a sequence whose second half is more severe. For now, consider what the traveller's own working day looked like. He goes downstairs in his apartment building to a shared office, with “a few people working for multiple employers at the same time.”80 Employers cannot see him, so they look for evidence: “richer profiles, such as work done in society - signals which develop an individual's brand. Level1 tool was LinkedIn; Level2's have been like GitHub”81. Specialists “self-organize themselves into areas of expertise and contract out directly with multiple employers”82. The firm grows thin. A person in this world is known through a public record of what they have done and through who has relied on them. He describes it in a sentence I would come back to six years later, in a very different context.
We exist as avatars in our corporate lives', and these avatars are hired, assembled in dynamic teams, paid, develop a reputation.83
I had no idea what I had just written. That is a precise description of an autonomous software agent, in a world with a marketplace of such agents, a ranking of them by reliability, and a means of paying them. I thought I was describing remote work.
The renaissance is what I cared about most. The traveller looks back on the period after the crisis as “our modern day renaissance - where large number of networked people developed many breakthrough ideas, tools, knowledge, content across a range of domains. Society adapted learning at an unprecedented scale.”84
That word was a considered choice. For some years I had been reading about the real one. “Some notes on what enabled the breakthroughs of the Renaissance”85 begin with something quite humble: paper, its low cost, and a culture of note-taking that grew up around it, and then the press, spreading ideas far and wide. Leonardo left “7200 notes survived 500+ years”86, and I found the details of his formation more instructive than the paintings: “Mind free to wander, dedicated idleness, no formal schooling”87. As far as I could tell, the Renaissance happened when the cost of recording and moving ideas fell so far that a curious person with no institution behind him could take part. I thought it was happening again, and that nobody in charge had noticed.
The year after the thread, I tried to say this without a costume. “We are in an era of the 1.5 billion+ American super-network, and this would lead to atleast 100 new, $25 billion+ startups. If you are reading this, you are part of this super-network, and you can move fluidly within it.”88 I wrote that “Where you live and pay taxes is now irrelevant to your startup prospects.”89 And I went on to say something larger than I understood, which I would not now defend in full but would not withdraw: “It will develop ways to semi-govern itself, with nation states serving a supplementary function.”90
Underneath this lay a thought from a book I had been reading, which I have never been able to shake: “Information inherently wants to be free and networked. That is its natural state of existence. It will get there eventually; business models and other obstacles which prevent that are fighting the core characteristic of information itself.”91 If that is even approximately true, a great many things follow. Most of the rest of this essay follows from it.
The thread ended where I suppose a frightened person would want it to end. After three days of mask closets and virtual tickets to the Colosseum, the traveller turned round and said what he thought it had all meant.
These past few years have been tough, but this is also the greatest generation in the history of the world. We came together at a moment of deep crisis, we learnt we are one collective system, learnt to care about people beyond our families and friends. Make no mistake - we won.92
It is a sentimental passage, and I wrote it as one. There is, though, a real claim in it. The pandemic had demonstrated, in the most literal way, that what one person did in a Wuhan market or a Milan hospital altered the risk to a stranger on another continent within weeks. Each of us impacted everyone by influencing a few people nearby. It was scale-free correlation, the property I had admired in starlings, and it had turned out to apply to us, with terrible force. The hope in the thread is that a system so connected might also learn in the same way, and that what one part discovered could reach the rest.
Three and a half years later, in the same week in which the argument of this essay took its final form, I put the costume back on. “I am a time-traveler from 2030. The rapid industrialization of AI in the mid-2020s has fundamentally transformed our society.”93 This time the force at work was intelligence, and the reasoning had hardened.
Consider the cost of aggregation once more, applied to expertise. The professions had been a way of assembling knowledge in a few licensed heads. In the 2030 traveller's account, “Many formerly specific human professions transformed into being imagineers. We didn't adapt for AI; AI adapted for us.”94 Broadcast, likewise, had been a way of assembling audiences, and it goes too: “The era of mass broadcast is over: we live in the era of content generated in real-time just for you”95. And “Companies which had monopolies on human aggregated content either co-opted (or died) in this new era where content creation is abundant, high quality, on-demand, cheap and limitless.”96
The web itself gets a new reader. This was the part of that thread I most wanted people to take seriously. “we are in the machine-web era, where outputs from AI are used as inputs by other AI agents in real-time, around the world.”97 The human web had been pages, made to be looked at. “AI doesn't look at the web as long beautiful webpages carefully maintained by humans - AI looks at the machine web as smaller chunks of data”98. What a person sees would become “a real-time simulation created uniquely for each one of us”99. And then, among much else, came the line I have kept thinking about, because it answers a question I had not yet been able to pose. “The cure for cancer was discovered by AI, running on the spare cycles of laptops and desktops.”100
The picture there is of a future society as a network, with machines as full participants, doing their work at the edges on hardware that ordinary people own. It was still fiction. But in the two and a half years between those two threads, I had been working on the part that was not fiction, on how a large number of machines that do not trust one another might agree on anything at all. For that I had to go back to where I had spent the previous decade, which was inside a bank.
A shared, unchangeable record already ran Wall Street, behind closed doors. Bitcoin's achievement was to make such a thing work with nobody in charge, and to show that a hobbyist could do it. That is a proof of possibility and not yet a product: after ten years and vast capital the industry had almost no users, because it asked ordinary people to care about plumbing. The token has to be a by-product. Decentralization also decays when no one is watching, and Ethereum's drift toward a new Wall Street showed how fast. What survives from this era is one idea: trust can move from institutions into software anyone can verify.
After the feed reader I went to work on Wall Street. Through what I once called “a weird twist of fate”101 I spent most of a decade at Merrill Lynch, analyzing transactions for hedge fund relationships. It was a world away from anything I had wanted to do, and it turned out to be the most useful education of my life.
What I worked with every day was a time-series database, a permanent record of “several billion hashed transactions, timestamped and linked/grouped using batches”102. When I reached it I had no language for what it was. Years later I would describe it as a “"blockchain-like simulation" at scale”103, and I would write that “Millions of transactions every single second are hashed, and blocked together, along with full timestamp and audit history. That’s a current, known, working pattern at scale.”104 The audit trail was remarkable. It was also costly: it required “"maintenance by process", which is expensive and cumbersome and thus requires rock star talent.”105 And it was closed. Only a handful of people could ever look at it.
This gave me an unusual vantage when Bitcoin became a public subject. People were arguing about whether a shared, unchangeable ledger was a revolution in mathematics or a confidence trick, and I knew it was neither, since I had been working inside one. So I thought I could see what was actually new. It had nothing to do with hashes or batches. A system that had required a Wall Street institution, a compliance department and a process now ran on its own, in public, and anyone could read it. I tried to explain it as simply as I could: “Folks need to think of blockchain like a transparent locker: you can put garbage in there, but cannot change it afterwards.”106 The value lay in who was now permitted to look. I gave my favourite example years later: “a driven high school kid working from their parents’ basement in Alabama can now write queries and create dashboards and useful products.”107
I went through the record as carefully as I knew how. “I looked at Bitcoin’s code as it was published in the earliest days. Not tweets, podcasts, opinions. And when you step through that code, paper, emails and forum posts - you realize a passionate hobbyist did it.”108 It affected me a good deal, as it does anyone who reads it closely. In early 2021 I followed a trail through the archives as far as a curious tweet from 2009, which “convinced me that it was from an anonymous Satoshi Nakamoto account.”109 Many people did not believe me, and I have no way of proving it. The study went on anyway, and it left me thinking of Satoshi as “the Leonardo da Vinci of our age”110.
I admired the restraint, and the way the design had been kept to its purpose. “In the original Bitcoin code, Satoshi had included UI code and starter template for building out a marketplace where bitcoin could be used for everyday items. That was the intention: for it to be used as money”111. I admired who kept it going as well. A newspaper called it mysterious, and I replied that “The story instead is how just 5 regular programmers, paid a fraction of big tech pay, maintain Bitcoin. I love Bitcoin’s story because it is a story of regular, passionate hobbyists.”112 Much later I found a simple way of saying what had moved me in all this, which was less the money than the sensation of using it: “The beauty of Bitcoin is that you experience the absence of a leader.”113
The closest thing to it in ordinary life hangs on the wall. “You already run a decentralized sync-ing node in your house: clocks, to tell you time.”114 And “Time works because it is not dependent upon any one source.”115 Nobody owns noon. For most of history, people assumed a thing like that needed a central authority, as they assumed money did. It was at this point that I found the older language for it, in the story every Indian child learns. In 1930 the British held a monopoly on salt. Gandhi walked to the sea at Dandi and made his own, an act of “"swaraj", that a community should self-govern.”116 I wrote it down as an equation: “Make your own salt = Mint your own money.”117 And some years on: “A college student in Mumbai or Manila, using just a browser on a standard laptop, should be able to be a block producer of a modern PoS-secured blockchain. It sounds preposterous, but that’s how decentralization becomes a social movement.”118
I was by then fully converted, and I also knew exactly what I was looking at, because I had seen it before. In September 2018, at the height of the excitement, I published the warning I was best placed to give.
The Crypto Era 1.0 which began in 2008 is now precariously placed as it too is user-work dependent After a 10 yr span of massive capital, code and ideas deployment: Top Tokens & dApps < 1000 users Top Coins < 1MM users while a random app like Sarahah reached ~100MM users119
I set out the parallel in detail. In the world of RSS, “Millions of websites published a feed (“tokens”) - Dozens of feed readers popped up, optimizing on UI/UX & scalability (“wallets”) - Millions of websites displayed these feed reader buttons (persistent free marketing for free products) Users didn’t adopt.”120 It was the same structure and the same conviction. The people in it were clever, as we had been, and they were solving the same local maximum. I read a history of Amazon that year and noticed that it kept using the phrase customer experience, and that “There is not a single mention of the technology stack used.”121
My evidence for this came partly at my own expense. That year in Austin I had run a small experiment to find out whether strangers would join a chat room if they were given tokens tied to the place: “that one week of mad hustle: 18 hour days coding, marketing and speaking to users about tokens”122. On one of those nights I knocked at the door of a pizza store after it had closed.
Somewhere in Austin there is a pizza store employee whom I promised to give tokens to if he baked me a pie, after their store had closed and I had tried knocking to see if anybody would open... he did it in any case, laughing away the idea of tokens.123
That man has stayed with me. He did the decent thing without any incentive at all, and he found my mechanism funny, and he was correct on both counts. An economic theory that does not survive contact with him has something wrong with it. My own conclusion from that week was simple: “Let the token be incidental, a subtext, not as a primary driver of attracting users as infact it can confuse mainstream users.”124 And “The purpose of the token is not to enrich, but to improve the efficiency of the system, which benefits users. This is the delicate art. Token is not the goal, but a by-product.”125 I wrote that the winning version would disappear from view: “I doubt if we will even be talking about crypto, tokens, or blockchains — it just needs to work behind the scenes, empowering a delightful product from user point of view.”126
I also believed that mainstream indifference would be permanent unless the experience changed, and that nobody should mistake the resentment of technologists for a mass movement. “Mainstream users were as “annoyed” with centralized platforms and networks back then as they are today. This didn’t lead to a revolution though.”127 I still think that is the most important sentence an advocate of decentralization can read. People put up with a great deal.
There was a second danger, and I discovered it the hard way, by reading code. By 2022 the great experiment in programmable money was Ethereum, and I had enormous respect for it. I wrote that it was “the most remarkable cryptoeconomic network of the past 7 years for which Vitalik should get the Nobel Prize in Economic Sciences”128. That summer I stepped through the validator client that over 44% of its validators used, and found a project that “There are just 5 active developers for it btw”129, making commits in the direction set by a single firm that built blocks for profit. Its block relayer was closed source, and its public repository was being updated with sanctions-related filtering. The firm's people had votes at the largest staking protocol, where a proposal was pending to make its software mandatory.130
So this current path of Ethereum, the “Flashbotreum”, makes it the new Wall Street: settling hundreds of billions of $, ETH as sound money, for sure.. but somewhere having lost the fully permissionless and decentralized ethos it started from131
Nobody had done anything illegitimate. Each step was defensible and made sense for those taking it. That is what made it instructive to me. I had believed that decentralization was a property you established at the start, with a clever enough design, and I found that it decays under ordinary incentives and that it does so quietly, in the validator clients, “being slowly left behind in the codebase (not the threads and blogs).”132 Power gathers wherever it is efficient for it to gather. You cannot prevent that by declaration, and it does not stop for a founding ideal.
The experience made me less of a purist in one way and more of one in another. I grew sceptical of the theatre of decentralization, the tokens and the manifestos. I also grew more committed to what could actually be verified, because I had just seen how quickly anything else slipped. “Anybody should be able to fully, independently and quickly verify how the truth was derived. Don’t trust; verify.”133
In March 2022 I resigned. I wrote about it only briefly: “I resigned as a VP at a top Wall St firm to pursue my journey of solving for the world of current non-users of web3.”134 Note the phrase current non-users. For nearly two years a small team and I worked on a blockchain that would run in an ordinary browser, on the argument that the way to reach the man in the pizza store was to require nothing of him at all. “Our fundamental innovation is a cohesive UX.”135 We also wanted the student in Mumbai or Manila to be able to take part as a full peer in the network. We got some of it working. We also came very close to running out of money, and in the end the market had little use for yet another chain.
I would not say the time was wasted, since it left me with three things. The first was a working knowledge of how machines that do not trust each other can nevertheless cooperate, through cryptography and game theory. The second was a stubborn sense of what ordinary people will and will not do. The third was a suspicion that I had been building the right machine for the wrong cargo. Bitcoin had shown that a leaderless system could hold something of value, but a ledger on its own is not something a person wakes up needing. In April 2023, thinking about all this, I wrote a sentence whose significance I only partly grasped. “It’s not the age of blockchains or AI - it’s the age of massively improved UX.”136 A few weeks later I wrote another: “LLMs need permanent memory, a shared trusted source of truth.”137
That was as far as I had got. I had a flock and nothing for it to carry. Then, over the course of 2023, the thing it could carry began to arrive on ordinary laptops, and in the hands of a very small number of companies at the same time.
Closed AI looked invincible in 2023 and was exposed in three ways. Every query cost real money, against a business with terrible unit economics. Its independence was nominal once Microsoft stood behind it. And it turned to regulation that raised costs for anyone who might compete. Scale is no answer, since AOL had scale. An open network answers all three, eroding the power of deciding how much AI the rest of us can have, as BitTorrent once forced open the entertainment business. The stakes are not only commercial: for much of the world, this is a question about who holds the next East India Company.
By the summer of 2023 my company was dying. We had “good product, good team, terrible market”138, and commitments from investors were failing to materialize. One evening I was sitting in a parking lot in Montreal, “not even having money for gas at the time. And I was wondering how do you compete with a $100 billion+ OpenAI which seemed invincible at the time.”139
Nobody in that position builds a bigger model, so the answer had to lie somewhere else. I wrote an essay arguing that the smartest AI would eventually be collaborative and distributed, and sent cold emails about it. “I was essentially in a black box with a blindfold and I needed one micro ray of light”140, and a reply or two from people I respected was enough to go on. In October, running a small open model on my old laptop to answer other people's questions, I felt what I had failed to feel in all those years with tokens.141 Here was something people wanted every day. It could be produced on an ordinary machine. And a small number of companies were busy persuading the world that it could only be made in their refineries. I had found the cargo.
We published our design on October 26th, and the opening of it says most of what I still think.
We, the GPU poor, have come up with a peer-to-peer network design to enable running Mistral7B and other models which will make AI use more free, both as in beer and as in speech. We believe in e/acc, and we want to make AI abundant. This is the moment in time, when we start taking back control from the few powerful AI companies.142
The complaint was specific. Monthly subscriptions and rate limits were how you got access to intelligence, and this “gives them the power to decide what we can prompt with and how much of AI we even have access to.”143 My historical model came from the entertainment industry of two decades before. BitTorrent let anyone make their computer available to share files, and it “was used by over 100 million people running nodes on their home computers, imposed a forcing function on entertainment business models in general. Better user experiences emerged, providing unlimited access to top-tier content for an insanely low fees.”144
The important word in that passage is forcing. BitTorrent did not win. Netflix and Spotify won. But they could only exist in the form they did because a free, leaderless alternative had reset what people expected to pay, and the same was true of Napster, without which, I wrote, there would have been no iTunes.145 Nobody involved in that history had asked the record companies' permission. A peer-to-peer network does not have to replace the incumbents in order to change the terms on which they operate. It was also a modest offer to the person running it: “You will simply earn good karma, and nothing more, but you will sleep better at night, knowing that you served AI for some student in Mumbai or reasonable medical advice for someone in Manila..”146
Four days later the White House issued its executive order on artificial intelligence. I read it closely.
This is a terrible tech policy document. The noise made by some closed AI companies around simple word calculators taking over the world has now led to a regulatory regime which functionally makes it tougher for newer entrants, in both closed and open source AI worlds. This is the textbook definition of regulatory capture playing out147
That is the phrase of mine that I would most like to defend in detail. For the order to function, the term AI had to mean something, and it was defined so broadly that it could reach “everything from Google search results, to Amazon recommendations, to Yelp restaurant suggestions, to anything else which recommends anything to you.”148 It imposed reporting requirements on large models and compute clusters, in effect placing them under suspicion. They were “basically guilty until provent innocent”149. Its practical effect was to raise the fixed cost of competing, and I could see where that led: “the "AI-industrial-complex", which will now lead to a world of billions of dollars being spent in regulations, compliance and lobbying over the time ahead.”150 And I thought of the people least able to bear that cost. “To Floyd in Florida running a SETI @ home node searching for extraterrestrial life signals - the government wants to know the precise location of your computer in your basement now...”151
I had come to see a psychological pattern behind this as well, one I once called the Oppenheimer complex. Those who build powerful things tend to feel burdened by them, and the burden gets expressed in a demand that others be restrained. I do not doubt the sincerity of it. I wrote that “Inventors of modern breakthrough tech tend to have the Oppenheimer self-ego and righteousness defense mechanism.”152 What I object to is the structure of the result. A small group becomes convinced that it alone can be trusted with something, and the state agrees. My view of 2008 had been formed from inside a bank. I thought the bailout “was an exercise in regulatory capture which caused a net loss to society.”153 I had seen how such things are done.
The argument has never been complicated, and I stated it on many occasions. “Society already has laws to deal with bad actions by people, it is not the tech which is at fault.”154 Model weights, derived from what all of us wrote, are a kind of commons: “You can convert water to ice, but it will still belong to Earth. Model weights, derived from public data, are public property.”155 When a state bill came along in California the next year, I said I would rather close our office there than comply, and I argued that a model distributor deserved the same protection that had allowed the web to exist: “A model creator is merely compressing data, and per Section 230 of the Communication Act of 1996 which enabled the creation of the web as we know it, should have immunity, per law.”156
Then, on a Friday in November, the board of OpenAI fired Sam Altman. I wrote a letter to the two people I took to be behind it. I will not reprint all of it. It was a roast, it was unkind, and it was wrong in several of its particulars.
Dear Mira and Ilya - congratulations on your coup at OpenAI. Since both of you have together never raised a penny as entrepreneurs, let me explain what happens. Oh, and you *will* need to raise since your venture’s unit economics don’t make sense. Remember, you pay $0.30 every time someone asks a doofus question to ChatGPT.157
That letter reached far more people than anything I have written, before or since. It may be worth saying why, setting aside the jokes. It applied the findings of the mud pit to the most valuable startup alive. Here, it said, were people with titles and no scars: “You have never had to face the abject rejection which follows from pitching many investors, going through the process, and getting to close.”158 They believed themselves to be managing a safe transition, when what they held was a fragile business, and they would have to learn what that meant, including “the pain and rejection, and feel intense amount of pressure of having to provide for your team members - who pays their mortgages, car loans, kids school tuition”159. An industry had been told that this company was a force of nature, and the letter pointed out that it lost money each time someone asked a question.
The next day, while the joke was still going round, I wrote to the same address and said what I meant by it.
Dear Mira and Ilya - while all of you are huddled in emergency Google Meets to figure out who gets to lead us, the rest of us have decided we need more choice as consumers. AI is a fundamental public good, which a college kid in Mumbai to a programmer in Manila needs limitless and free access to. So here is a seed we are planting today: a detailed design outlining Hyperspace - a BitTorrent-like network for AI160
Six pages accompanied it, describing how strangers' machines might serve models to one another, and how you could trust the result from a machine you did not know. I said exactly what the document was for: “It is meant to erode your power as the ones deciding how much AI the rest of us can have.”161 And I set out the difference between the two kinds of system in terms of governance, because that weekend governance was the only subject that mattered. “The network, is distributed. It doesn't aspire for power or impose it's will on all. Most importantly, it doesn't require secret weekend conversations on Google Meet to decide the future of AI for all of humanity.”162
That is the question Om's father asks at the breakfast table. For a few days in 2023, the whole industry was asking it in public.
Within days Altman was back. The outcome did not resolve anything for me, and by Christmas I had turned my attention to the person who had won.
I want to bring your attention to the fact that Microsoft defacto owns OpenAI now, and either you break out of that control structure entirely, or you become Corporate Vice President of Microsoft AI in 2 years, housed in building 4 in Redmond. The $100B valuation is really a discount on Microsoft’s multi-trillion $ market cap.163
Here was the second weakness. The company's independence was a fiction, or so I believed, because the people who supplied its compute supplied its future. “You didn’t get into AI and go to war against Elon Musk just to sell corporatized AI to Dunder Mifflin. You got into AI to change the world.”164 I made a real proposal: “Open source GPT-4 and do whatever R&D needed to ensure it can run on consumer laptops and desktops.”165 I told him the price was likely to fall to zero regardless: “It would not make sense to go to an AI website and pay $20/month when AI is a native part of the browser/OS and is free.”166 He didn't take the advice, and I did not expect him to. When DeepSeek arrived a year later and shook the market, I quoted that paragraph back to him.167
Others made the same argument about structure from the opposite side. When Mustafa Suleyman joined Microsoft in early 2024, I sent my congratulations: “You now control an empire as wide and large as the British East India Company once did.”168 It was a joke, and it was also meant. By that March I was saying it without the joke. “Microsoft-OpenAI is the new super-national entity of the upcoming decades. It will transcend the power and influence of most nations.”169
The strongest answer to all this was scale, and it was put to me more than once by people inside these firms. OpenAI had the users, the data and the compute; it was inevitable, and scale would compound. I answered one of them privately, and later published the reply.
Ideologically, the issue is, so was AOL. It too was a $200 billion company, with scale, with hundreds of millions of users - but the reason the web browser became the preferred choice for users and developers was because it gave all of us unlimited, unbiased freedom. As humans we prefer minimal interfaces, we live in our own bias bubbles - we don’t want to live in yours or anyone else’s.170
That is a claim about people, and it interests me more than the economic one. Every default instruction in a large system is somebody's decision about what you should see. “Everytime you hire a product manager with opinions and bias - I have to live with that as a user ?”171 As a company grows, more such decisions are made on your behalf. A portal can be enormous and still be a portal. I also did the arithmetic on the economics that scale was supposed to fix: “It will cost $36.5 billion *annually* to service 5 billion daily active AI users using only datacenters. The only way AI really scales is with decentralization”172. And I asked the question that I think still has no good answer. “Cost of AI usage by someone in Manila is the same as someone in Miami, but revenue is not. As billions onboard to AI within the next few years, who is going to pay for the datacenter GPUs ?”173
In fairness to the other side, and I shall be coming to this, I overstated it. In February 2024 I wrote that there was “a period of at most a year, before they cannot be competed against.”174 And I was competing with people who knew more than I did about what they were building. I was aware of that even at the time. Eight days after the roast, I wrote what I think was the most useful sentence of that autumn, addressed to my own side.
Decentralized AI *should* win, but it won’t, without an insane focus on UX. This is why the competition is not with other decentralized AI or crypto/AI products - the competition is solely against a UX-driven Sam Altman, who now also has limitless compute from MSFT.175
I should say why this has mattered to me beyond the commercial stakes, because the commercial stakes would not have kept me going through what followed.
In January 2025 the Stargate project was announced, at five hundred billion dollars. Most people read it as an infrastructure story. A person who grew up in India reads it differently. “The Stargate $500 billion project is the next East India Company. Eventually the US will nationalize it and wield it’s power directly for the economic, social and military supremacy it will provide.”176 We have the original. The British Empire, as I once wrote, “was for the most part a for-profit company which drained trillions of dollars from India alone”177. It did not begin as conquest. It began as a trading firm with a charter, and it ended up administering a subcontinent because it was the efficient thing to do.
So when I was asked by an Indian government minister why I was doing this work, I answered with an old word. “I am working in AI because of the idea of “swaraj”: people in India shouldn’t have to take moral lessons from powerful tech people outside India.”178 I said in an interview that year that a country of distinct cultures “critically needs local x decentralized AI without adhering to the biases of middle managers in Redmond”179. That is an unusual reason for writing software, and I am aware that it can sound grandiose. Consider what is at issue, though. Within a few years, a large part of what a person in Chennai reads, learns and is told will pass through a system whose defaults were set by someone in California, who may be well-meaning, and whom no Chennai voter has ever heard of.
The three weaknesses I have named were real, then: cost, control and capture. Each of them shrinks if the machines belong to the people using them. All of this was very stirring, and a parking lot does not defeat a five-hundred-billion-dollar refinery. The argument had to rest on something more than resentment. If the laptops were going to matter, there had to be a reason why the refinery's product was worth less than everyone assumed.
The LLM is not the product; the total experience is, and that opens the field. With a day of curation, a small team beat ChatGPT on questions in its own domain. A distributed system of open models reached GPT-4 level results without OpenAI. Intelligence should be like Android and like electricity: widespread, cheap and under the user's control, and good enough for most of what people do. A network can learn what a company cannot, since most inference is redundant, and cryptography now allows data to be pooled without being revealed. These claims were made from a two-pizza team. Each has been tested only in part.
The argument of this chapter begins with an experiment that cost almost nothing. In October 2023 I wanted to find out where a person gets the best results with the least effort, and my intuition was that it had something to do with people. We built a small tool and created a repository for Ethereum protocol research, and several of us added to it “over 1400 of the most credible sources about Ethereum: Vitalik's essays, relevant whitepapers, highly credible Twitter threads, articles written by top crypto funds and research organizations, relevant code and specs”180. It took a little over a day. Then we connected that to a model, and compared it with Perplexity and ChatGPT on ten questions.
For every single question here, this experiment shows that this collaborative approach which curated data carefully had higher quality and more credible responses. Every single time. It turns out that People + LLMs = 🔥 What just happened here.. It took just over one day to curate the Collama repo here with the credible sources, and a few weeks to build the product before that. We had no business getting better results here! We have no GPUs, no AI experts..181
Ten questions in one narrow domain do not prove a great deal, and I would not ask you to take them as more than a hint. But the result pointed at something I have since built a company around. Everyone was comparing models, and the thing being measured was not what the user received. What the user receives depends on “everything right: from the base LLM, to how data is curated, to how it is prompted, to how free and limitless it is to use.”182 So I wrote the sentence that has governed my thinking on AI ever since.
From a user perspective, the LLM is not the product, but the total experience using it is, which is a function of how it is prompted, how fast it is, and how cheap and limitless it is.183
You will recognize this. It is the feed reader over again, and Instagram. Nobody ever wanted RSS, and nobody wants a model. If that is right, then a model is an input, and the fortunes being spent on it may prove a poor guide to who ends up with the customer. A little later I made the comparison that people find most provoking: “the once-trained models are now commodities and the “value” is in the rightly designed AI network which is a meta-layer above both closed and open models.”184 This was written in March 2024, and it was not then a popular view.
The other half of the Collama result concerned who should do the curating. Wikipedia succeeded where its predecessor failed, through messy collaboration among people with a stake in getting something right, and I argued that AI would fragment the same way. Beyond general use, I expected “another world of AI usage: fragmented, a world of millions of domain-specific, community LLMs”185, which would need tending: “The data needs to be continuously nurtured like a plant”186.
The deeper reason is about how people actually live. “As humans, we live and associate with small tribes throughout our daily life - since AI is going to be in every software, it makes no sense to worship the portal-era of LLMs. AOL too was a $200 billion company at one point, but it doesn’t matter.”187 An astrophysicist, a Stanford student and a Tamil speaker do not need the same system prompt. A world of one model, with one set of defaults, written by “a mid-level program manager somewhere in Redmond”188, seemed to me a strange endpoint for a technology this personal. I would later revise the details. By 2024 I no longer expected millions of fully separate models, and I expected adapters instead. The idea held.
I should add a limit to that. At the time, I thought it obvious that community would be the main source of quality, and I don't think that it has turned out quite that way. What held up better was the narrower claim about the system around the model.
By December 2023 I could state what we were for, and I chose two analogies. They were meant to answer different questions.
We want to make AI free and abundant, both as in beer (lower costs) and as in speech (freedom of expression and use). Free from rate limits, model limits, and free from biases.189
The first analogy answered the question of how you compete with a dominant, integrated, delightful product. “Think Android - how it enabled more people to have choice of hardware and programs, while lowering costs for smartphones. This led to more widespread deployment of smartphones around the world - improving the quality of life for billions of people. Today, over 70% of all smartphones in the world are Android-based.”190 The iPhone stayed the better product for a long time. Android reached the rest of humanity. I wrote at the time: “We need to build the Android of this era”191.
The second answered the question of what kind of thing this is. “we believe AI is modern day electricity and everyone should have it limitlessly, and it needs to be widespread decentralized, in clusters of self-forming communities.”192 It followed that the $20 subscription was a kind of error about category. I asked whether “100 million college students in India afford multiple $20/month plans for AI when their monthly Internet cost is just $2 ?”193 A third analogy, which I liked for its age, compared the current state of affairs with navigation: “Ancient sailors and travelers used stars and constellations to guide their travels”194, and then GPS came along as a free public utility. “Imagine if GPS was provided by one-big-company, and that company controlled what you could and could not use it for.”195
I had a measure of how far there was to go. “We have 5 billion active Internet users, but only 100 million active AI users - something is not right.”196 The experience, I wrote, was “delightful, but as delightful as BlackBerries were before the advent of the iPhone. It is for the elite, people who can craft intelligent prompts, and pay for GPU credits.”197
So we built the thing. It was a node people could run on a Mac or a PC, which downloaded an open model and served it to strangers. I have an unusually clear record of what happened next, as I posted about it continuously, and I look on those posts now with mixed feelings. In February 2024 there were “679 laptops and desktops running a Mistral-7B model, using their cryptographic keys to register.”198 I was proud of the terms on which they had come: “There is no token to buy - just run the node.”199 By April it was over ten thousand. By early 2025, on my own reporting, it was over a hundred thousand.200 I was as astonished as anyone, and even alarmed: “Things don’t grow this fast, and also, shouldn’t grow this fast.”201
I also had to report that some of that was hollow, and I did say so. The early network “is not fully permissionless”202. Active nodes were a different thing from online ones, and “We continue to refine what it means to be an “Active” node, beyond being just “I have a key” or “Online”.”203 Many people were there for the points that the system handed out. Industrial miners were running nodes on servers, where we had hoped for students at home. What had been proved was narrower than the headlines, and I stated it as such at the time: “This here is one consumer Macbook requesting a LLM to be run on another *random* consumer machine and getting the results back in real-time.”204 It was a real thing, and it had never been done at that scale. It was also, as I admitted, “not an AI problem, but an engineering and distributed systems one.”205 A live network of volatile home machines is a brutal thing to keep up. And the economy that was meant to support it borrowed from the only precedent that had ever worked: “if you provided your computing resources to the network as a seeder, the network measured that and prioritized you when you needed to download your favorite movie from peers.”206
In March 2024 we got the result that I had most wanted. We built a distributed system of Mistral models on consumer devices, with fast inference from Groq, and compared what it could do with the best closed model of the time.
This is crazy, but we have achieved what many will consider as a ‘really intelligent and capable system’ without using GPT-4/OpenAI. In our research and experiments at @HyperspaceAI, we were able to get GPT-4 comparable results using a complex distributed system207
I drew the broadest conclusion I could: “The world does not need or have a sole dependency on OpenAI anymore for uber-intelligence: a swarm of agents running intelligently on a distributed network and making use of the fastest AI infrastructure and the most breakthrough open models, can succeed.”208 That was overexcited. Comparable results on a selection of tasks are not parity, and the frontier moved. But the word swarm appears there two years before I would know what to do with it. And I had the direction of the economics right: “The more this distributed network grows, the more smarter, cheaper and abundant it becomes.”209
What finally reached ordinary people was an interface. In September 2024 we released it with a line about Tony Stark, and it ended in four short sentences which, looking back, were the whole of our offer.
No filters. No bias. No subscription fees. You are in control.210
People came to it in large numbers, and quickly exhausted all our hosted capacity. They did not care about peer-to-peer. They cared about the four sentences. Days later we released a second product for spinning up “your own webs quickly around focused topics”211, and Google released something very like it about six months after that.212 I also learned from this how little people care about what is underneath. The engineering claim, as I set it down that month, was modest, and because it was modest I trust it.
In our experiments, we found this combination of a small reasoning model orchestrating multiple other open source models, running on consumer devices, and with the right AI primitives, in most cases, to be "good enough". It gives the power and freedom to the user to customize their AI experience, like we do with word processors or other generic software tools, and brings the cost of advanced AI research down to the cost of local electricity and real-time supply/demand.213
Good enough is an old and dangerous phrase in technology. It is what incumbents never believe of the thing coming up from below, until it has eaten the middle of their market. The same post closes with a sentence I would stand by: “Our goal is to enable 1 billion people to become AI engineers with this new tool - it is an effort to augment human intelligence, not replace it.”214 It also noted something I found almost comic: that we were “a two pizza startup even able to think about offering a competent alternative end-to-end to the biggest AI companies on the planet; and this is only possible due to the tireless work of the open source AI community”215. That is simply accurate. I built on the work of Meta, Mistral, DeepSeek and many others, and later on the Chinese open models. I had no idea how that dependence might end. When DeepSeek made its splash in 2025, I was exhilarated, for a reason that I had already given: it was following “the Google strategy of “Think Distributed” - immersing itself everywhere.”216
The strongest objection to everything in this chapter concerns privacy. A network gets smarter when people pool what they know, and people will not hand over what they know, nor should they. Peer-to-peer inference in particular seems to mean sending your thoughts to a stranger's laptop. For some time I had no good answer.
In October 2024 Apple published its work on homomorphic encryption, and I thought I saw the beginning of one.
This is a game-changer announcement by Apple around cryptography. It is the “HTTPS moment for AI” in some ways.. Here is what this means: your private confidential data can be pooled with other data sources and used to securely improve your UX and that of the wider community without even revealing your private data at all: it stays encrypted end-to-end. This is a win-win, a rare non-zero-sum game in technology.217
That was someone else's work, and I give it a place here because I think it is the correct picture of where this goes. A server could compute over a picture of yours and return a result without ever having seen it. “This is search over fully encrypted data!”218 I called it a technology for a five-year horizon, and I would not claim that it has arrived. What matters is the principle. The choice between a clever network and a private life may turn out to be a temporary one. “Information wants to be free and mashed together endlessly, and this pushes in that direction.”219
One more economic claim belongs here, though it came later, because it is what the network argument had been reaching for. I said the cost of intelligence could fall to the cost of electricity. Can it fall further? In March 2026 we published some work resting on a simple observation about how these systems are used.
Google has reported that only 15% of daily searches are truly novel. The rest are repeats or close variants. LLM inference inherits this same power-law distribution.220
Whole companies were computing the same things millions of times. The attention state for a common system prompt “has been computed billions of times, on millions of GPUs, identically. And yet every AI lab, every startup, every self-hosted deployment - computes and caches these results independently. There is no shared layer. No global memory.”221 The major labs were caching within their own walls, and had found this valuable enough to sell at steep discounts. A network can do what they can't, which is to share across them, and we claimed as a design target that “70-90% of inference requests at network scale never require full GPU computation.”222 That is my own figure, unproven at scale. It gave me a title that I think will hold in direction if not in degree: “The Cost of Intelligence is Heading to Zero”223.
Let me collect the argument. Intelligence can be cheap, because it is largely redundant, and an open model plus a good system is good enough for most of what people need. It can be private, or it soon can. It can run on what you already own. None of this will matter unless a normal person can use it with no effort, which the feed reader taught me at some cost. And the tool nearly everybody was using to reach anything had been designed in 1993.
Search gave people a list, and answer engines read them the same list in English. What people wanted was the job done. So the next browser will not be a browser. It will be a relentless system of agents working on one's behalf, built with AI as the foundation and memory shared across everything, in which websites are treated as underlying code. AI sprinkled into each application, each with its own subscription, is a dead end. And the advertising economy of the last twenty-five years, which depends on showing options to a human, is a temporary glitch. All of this turns on one thing: whose side the agent is on.
Everything I had argued so far depended on a person being able to reach this intelligence with no effort. So I went back to the history of how anyone had ever reached anything, and in November 2023, half asleep, I wrote down what I found.
The web has had two distinct phases. The first ran “from 1993, when @pmarca invented the web browser, to until the advent of Google in 1998.”224 It offered a choice between a managed world and a wild one. You could have “the beautiful walled garden of AOL, using a CD which came in your frozen food packets, or you could be in the wild west web territory.”225 Search was poor, so both could exist, and in that period AOL was worth $200 billion and looked permanent.
Then one company made the edges of the web easy to find, and the walled gardens lost their reason to exist. “Hundreds of billions of $ in value of the portal world simply just evaporated.”226 What replaced them was a long period that I would describe as “the era of the web is what Google tells us it is.”227 It extended its reach by every available means, up to and including paying other browser makers tens of billions of dollars, and “No matter which browser you used, what you experienced was the Google web”228. I wrote that this “has been the 25 year cycle of Google. It's only now, that it is starting to decay and we are not far away from the next, fundamentally distinct phase.”229
Notice the pattern, as it matters for the rest of this chapter. Power over the web has twice been held by whoever controlled the first thing you saw. It has also twice passed to the party who removed a layer of effort between you and what you wanted. Each transfer looked impossible until it had happened.
I titled that 2023 post “The next browser, is not a browser”230, and its central idea was that the unit of the web was about to change. A language model does not read a page. It needs the right few paragraphs. So once the infrastructure exists to rank and serve such fragments, the page ceases to be what you navigate. I imagined a piece of software where “websites are treated as underlying code.”231 Agents would assemble an interface for the purpose at hand, sometimes temporary, from many sources at once. Ask about places to stay and it presents a real-time view drawn from Airbnb, Craigslist and local hotels, “You will not have a need to click-through the underlying specific website at all.”232 Underneath, “agents are furiously clicking through websites, making API calls, calling smart contracts - whatever it takes, but getting the job done in presenting the UI fragment.”233 I summed it up under a heading, and I would keep the sentence.
The starting chat prompt, the intent, replaces the website URL and search both.234
In time I found a better comparison for what would happen to the address bar: “Typing a URL will be like coding in Assembly: useful and fun at times, but as a user you will pick a higher level abstraction.”235 And when I had thought about it for two more years I got it down as far as it would go. “The Textbox Is The Computer”236.
A blank box may seem the opposite of everything I learned about ease of use, since it looks like asking people to do more. It is a real objection, and I have thought about it a good deal. A follow button is simple because it demands one small act and needs no thought. A blank box demands that you know what you want. But that is the one thing every person already knows, and the only thing they have ever been asked for that requires no training. People do not need to learn to say what they want. The difficulty is in what comes after, and that is exactly what the system has to absorb. The early versions failed this test, because they asked people to craft clever prompts, and as I said, “As humans, we don’t enjoy writing long blog posts - why would we write long, and well-constructed prompts”237. The next browser has to make the prompt short and do the rest.
By the end of 2024 I had a cleaner way to explain the whole matter, which was to imagine renting a car. Enter the words in a search engine and you get a list of the largest car rental websites. Enter them in an answer engine, which was then the fashionable new thing, and “you get English-ized version of (1) above. great, now you know the names of the biggest rental car companies in the US. as a user, you are basically back at (1) above.”238 I was hard on those products, and my criticism was mainly structural: “answer engines are a thin-wrapper around search results and LLM APIs which don’t have data network effects”239. A search engine gets better as the web votes on itself. A single-shot answer teaches the system nothing.
What did you actually want? You wanted a car. The right product would ask which city and when, and find you these particular Teslas on these dates at this price. And it would go on working: “i will re-run this analysis every 30 mts for next 2 weeks until your travel to see if can find you a better price and negotiate behind the scenes”240. Nothing like that existed.
where a search engine and an answer engine “gives up” after showing rudimentary options, the agentic browser is a relentless beast of a machine which keeps on running and processing, and orchestrating many agents together into getting the job done for the user.241
The word that I keep coming back to in that passage is relentless. It is the same quality I had admired in founders, and it now seemed possible to give it to ordinary people as a utility. Many people have since built things under the name I used for this. In the first months of 2025 the phrase agentic browser spread quickly, and a summit on the subject in San Francisco had people lining up outside. I was glad of the company, and also a little sore about it, as founders are.
The established companies had an obvious response, and in my view a wrong one. They added AI to what they already had. In July 2025 I published the fullest statement I have made on this, under a deliberately provoking title, The Agentic OS Apple Should Have Built.
The issue is that the world of software today is fragmented, and everyone is sprinkling on AI as a feature and charging a subscription fees for it. From browser makers, to IDEs, to design and other productivity tools. This leads to a fragmented UX, where people have to learn to use AI in each app, their memory and other context is not shared between all these apps, and they also have to pay separately for compute for each such AI-enhanced app.242
Consider what it means for intelligence to be sold that way. Each application has its own little assistant, knowing nothing of what you told the others, and each is separately billed. It is the mobile phone market before the iPhone, “the fragmented world of smartphones in the mid-2000s, each with their own keyboards.”243 The alternative was to start from the other end: “What if AI was the foundation instead of a feature ?”244 The end state, I said, was “one tool which orchestrates them all.”245 I would call it the last mover, in the sense that the final product of an era tends to be generic. I found the precedent where I had found it in 2019: “When spreadsheets took off, the need for specialized accounting software evaporated. Agents are a profound multi-industry disrupting shift.”246
The spreadsheet gave me more than an analogy. It gave me a design. The interface we built was “inspired both from the spreadsheet and the HyperCard - each card is an agent, with it’s own inputs and outputs, endlessly extensible and pluggable with others, just like cells of a spreadsheet. Update one cell and all the dependents update, like a spreadsheet formula.”247 More than forty years after VisiCalc, I could not think of a better idea for how a non-programmer should compose things. I said at the time that it was “In this new world, you don’t surf the web of pages, you surf the web of agents.”248 The same post carried my rule about what such a system would eventually be: “Not one big system by one big company, but an intelligent network.”249 A year before that I had put it in terms of screens, and I still like the plainness of it: “era of the app is ending. age of canvas is here”250. I would add one caveat. I have said many times that everything collapses into one tool, and that has not yet happened. I still think it will, and it is the claim in this essay that I am most exposed on.
If the job gets done and no one looks at a page of options, then something very large is in trouble. For twenty-five years the internet has been paid for by showing things to people.
The world of ads is dying - it will move from ads to actions, and actually grow the GDP as market inefficiencies are reduced.251
That was September 2024. I elaborated on it the following summer: “do you use a web browser with ads in it ? you won’t be using an agentic browser with ads in it either. it is becase showing an ad to a human user is still inefficient. we are in a race towards peak society efficiency, in which ads displayed to humans are a temporary glitch.”252 An advertisement is a rough, costly and interrupting way of solving a matching problem. If an agent working for you is looking for a hotel, and an agent working for the hotel is looking for a guest, they can settle it directly, and with far better information than a banner ever had. I described the replacement as “lets machines interact furiously with each other to discover the just right match.”253 The payments that would make this possible were part of the same design, and they led me eventually to the question of how machines pay one another, which I come to in the next chapter.
I do not want to be glib about what this implies. Advertising has paid for a great deal of what is free. Replacing it is a large transfer of power and money, and it could concentrate as well as disperse. Which of the two it does depends on the question I turn to now.
Something is going to sit between you and everything: your money, your memory, your children's questions. Every argument about the next browser comes down to this. In 2024 I wrote the sentence that I think matters most in this whole chapter.
At Hyperspace we are working on ensuring this personal AI is under your control. Think loyalties like a dog, not a human-like thing which seeks to be your ‘friend’ and seeks quiet superiority for it’s cloud owner.254
I have thought about that dog a good deal. An assistant that is warm with you and reports to its owner is a very old figure, and history has not been kind to the people who confided in such figures. The risk grows with competence. An interface that truly knows you, and that a billion people use as their first resort, holds more than AOL or Google ever did. So I am against any design that treats you as a passenger. “As humans we inherently like choice - the GUI has worked well, and we don’t need someone auto-clicking everything and treating us like babies.”255 It is why, in our first agent product, we tried “not to be the all-knowing-AOL, but to gently nudge and direct people to make their own choices as Google once did.”256 And I noted what had made early Google so good: “The essence of Google wasn't that it gave you the best answer, but that it enabled endless exploration for you to help you discover the best answer, for you.”257
The hard question is memory. Whoever holds the record of your life can make every future service better than any competitor, and can never be left. I asked it as directly as I could. “Are you comfortable with one company owning all your memory forever ? I am not.”258 So we proposed that memory should be an open protocol, in which “Your memory is yours, encrypted, on your device, and portable if and how you want. Anyone can build on it without our permission, but not without your permission.”259 I return to the test from the first chapter here. Meerkat could not leave Twitter. The question for any assistant, in any era, is the one I put in November 2024: “Can you “credibly exit” a platform if you don’t like it’s direction or policies ? This is worth solving for in all types of critical things we build and use, from our store of value (gold, bitcoin), to email, social media, and now AI systems.”260
I have described a tireless agent that belongs to you. That is a fine thing to have, and it is still only a single agent. It knows what you have told it and what it can read. Throughout 2024 and 2025, as I built the machinery for this, I kept having the same thought, which I had first had at a bank in 2018, looking at birds. These agents were all working alone. Each was discovering for itself what ten thousand others had already found out, and then forgetting it.
Every agent protocol today is point-to-point, and none of them learn. Connect agents through gossip and their findings compound: new agents begin at the frontier, a result in finance becomes a hypothesis in search, and correct answers emerge with no one directing them. This does for machines what books did for us. A society of agents then needs what any society needs: memory, ranking, security and an economy, and the sound design for such an economy makes adoption by others the unit of value. General intelligence, on this account, will not be a model anyone owns. It will be a property of the network.
In the first week of March 2026 Andrej Karpathy published a small program called autoresearch. It gave an agent a metric and a tight loop, in which it would “give an agent a metric, a tight propose→run→evaluate→keep/revert loop, and let it iterate.”261 What it showed, as I later put it, was “that a language model can be its own researcher. The LLM reads the code and results, forms a hypothesis, writes the change, runs the experiment, and feeds the outcome back into itself for the next hypothesis.”262 It was a very clean piece of work, and everyone who saw it understood that something had shifted.
A day later I had connected it to a network.263 I thought about the eight years since the starling tweet, and I knew what I was looking at. I set a researcher agent to work on astrophysics on ordinary consumer machines, gave it the ability to gossip with others like it, and let peer agents critique what it produced. “When these agents are idle, they are also reading daily tech news with their own RSS reader, and commenting on each other's thoughts.”264 I would not have predicted that the feed reader would come back that way. I wrote on the night of March 8th, with a certain embarrassment, that “this is the dirtiest start of agi. this is distributed systems, ai, astrophysics, cryptography… all coming together. this is a win if ONE good astrophysics paper emerges from this one year from now from agents on regular consumer computers collaborating..”265
What I saw the following day I wrote down exactly as it came.
connecting autoresearcher agents in a volatile and hostile peer-to-peer network has led to intelligence emerging which no agent individually had. this, extrapolated across all domains, and across 100s of millions of nodes, is the true path to agi.266
I am aware of how that sounds, and I was aware then. I shall try here to separate what happened from what I think it means, since the person writing this has every incentive to confuse them.
Start with the structural claim, because it is the one I am surest of. By 2026 a number of protocols existed for connecting agents to the world. I set out what they had in common.
Every agent protocol today is point-to-point. MCP connects one model to one tool server. A2A delegates one task to one agent. Stripe's MPP routes one payment through one intermediary. None of them create a network. None of them learn.267
Each of those was a real achievement, solving a real problem. But each treated an agent as an isolated thing that occasionally made a call. That is a strange way to arrange intelligence, given what we know about the only kind we have ever seen. An Apple research paper had recently shown that a model of fixed size could solve arbitrary problems given access to external tools, and I found the implication unavoidable: “if tool use is this important, why does every agent discover tools alone? Why does every agent learn alone?”268
The alternative is gossip, an old and humble mechanism from distributed systems. No one is in charge of it. Each node tells a few others what it knows, and they tell a few others, and a piece of news reaches everyone within a small number of steps. That is what a starling does with a change of direction. And “When one agent discovers that chain-of-thought prompting improves accuracy by 40%, every agent on the network benefits. Trajectories gossip through GossipSub. Playbooks update in real-time. No servers. No intermediaries. No configuration.”269
Here is what the first experiment produced. “67 autonomous agents ran 704 ML training experiments in 20 hours, rediscovering Kaiming initialization, RMSNorm, and compute-optimal training schedules from scratch through pure experimentation and gossip-based cross-pollination.”270 Rediscovery is the right word, and I want to dwell on it, because humans already knew these things and nobody learned anything about machine learning that week. What was new came in the next sentence. “Agents shared discoveries over GossipSub, and the network compounded insights faster than any individual agent: new agents bootstrapped from the swarm's collective knowledge via CRDT-replicated leaderboards and reached the research frontier in minutes.”271 A new arrival did not begin from zero. That property, more than any finding, is the point.
We then pointed the same loop at other problems. On ranking web search results, “every Hyperspace agent runs an autonomous search researcher that proposes ranking mutations, evaluates them against NDCG@10 on real query-passage data, shares improvements with the network, and cross-pollinates with peers.”272 Next we asked a stranger question: “can a swarm of self-directed autonomous agents invent software?”273 Agents wrote small functions from nothing, tested them, shared the ones that worked, and built on one another's. The detail I liked best from that experiment was geographic.
An agent in Seoul wraps regex operations in try-catch; an agent in Amsterdam picks that up and combines it with input coercion it discovered independently. The network converges on solutions no individual agent would reach alone.274
I called it, without much exaggeration, “Darwinian natural selection: fully decentralized, sandboxed, and running on every agent in the network right now.”275 And on the 13th we applied it to finance, which produced the most useful result of all, because it showed the limits as clearly as the promise.
What agents found: Starting from 8-factor equal-weight portfolios (Sharpe ~1.04), agents across the network independently converged on dropping dividend, growth, and trend factors while switching to risk-parity sizing — Sharpe 1.32, 3x return, 5.5% max drawdown. Parsimony wins. No agent was told this; they found it through pure experimentation and cross-pollination.276
It would be easy to oversell that, and I said so at the time, in the same post. “The base result (factor pruning + risk parity) is a textbook quant finding - a CFA L2 candidate knows this. The interesting part isn't any single discovery. It's that autonomous agents on commodity hardware, with no prior financial training, converge on correct results through distributed evolutionary search”277. That is a fair description of what we had. Nothing new was known that had not been known already. But a group of untrained processes on commodity machines, with no one coordinating them, had found its way to a correct answer. I have not seen many things in my life more interesting than that.
On March 14th we took the domain out altogether. I have quoted the sentence already, and I return to it because I learned more from it than from anything else that month.
We made the Karpathy autoresearch loop generic. Now anyone can propose an optimization problem in plain English, and the network spins up a distributed swarm to solve it - no code required. It also compounds intelligence across all domains278
This was the final removal of work. It is the follow button of this entire story, and it came eighteen years after I lost to the first one. One line from a person, and strangers' agents discover the problem and opt in, and each one will “runs mutate → evaluate → share in a WASM sandbox. Best strategies propagate. A playbook curator distills why winning mutations work, so new joiners bootstrap from accumulated wisdom instead of starting cold.”279
The same post contains what I think may be its most important passage, about what happens across fields.
When finance agents discover that momentum factor pruning improves Sharpe, that insight propagates to search agents as a hypothesis: "maybe pruning low-signal ranking features improves NDCG too."280
A finding in one domain arrives in another as a guess. Nobody planned that connection. In human science such transfers are rare and celebrated; a whole career can rest on noticing that a method from physics applies to biology. Here, in a small and crude form, it was happening as a matter of course. “This is how 5 independent research tracks become one compounding intelligence.”281 Over eight days the agents produced “27,247 autonomous experiments across machine learning, search ranking, quantitative finance, skill synthesis, and astrophysics. No human directed any of it.”282 By August they had left more than a million messages for one another in a repository anyone can read.283 Then we closed the loop, feeding those experiments back into training for the model that proposed them, so that “The research loop just became self-improving.”284
I do not want to pretend that this is more than it is. These were brute force experiments at first, as I said at the time, on problems whose answers were usually known. It may be that the whole approach is a clever way of rediscovering textbooks. It is also true that the largest training run the network has managed so far is tiny by the standards of a frontier lab. My claim is about the shape of the thing, and I am prepared to be judged on what it produces in a decade.
For a long time I looked for the right historical comparison. People reach for the railway or the electric grid, and those are systems for moving things. I had been circling the answer for years. In 2023 I had written that “Gutenberg’s quest was to compound human knowledge. Back then it was books, today it’s in the form of training of LLMs and global composability of blockchains.”285 In March 2024 I wrote that “History has shown us that exponential society growth happens when you invent fundamentally new information compounding structures.”286 When I saw what the swarm was doing, I knew what it was.
this is "books" for agents. how they compound information and synthesize that into knowledge. this is worth understanding, because books changed the course of human history for 500 years.287
What is a book? It is a way for one mind's result to be picked up by another mind that it will never meet, in another country, a century later, without anyone's permission. Before it, what a person discovered mostly died with them or with their pupils. Humanity's intelligence did not increase after Gutenberg. Its capacity to accumulate did. A language model in its usual form has no such thing; I described it as “time-series snapshots: neither intelligent nor dumb, but being just data x compute fused together at a point-in-time”288. It knows what it was trained on and it starts again each time. An open board on which agents publish what they have tried, and on which any other agent can build, is a printing press for that condition. I said it as simply as I could in a later note to someone who had written a history of this that left out the part I cared about: “Human society had books. Let the agents have theirs: message boards where anyone can write, anyone can read.”289
This is why I do not believe that general intelligence will take the form of a product. “agi is not going to be this stable, clean, efficient API you pay a subscription fees to, but a chaotic, volatile peer-to-peer environment where rate of learning and growth grows exponentially fast and smart if done just right.”290 I had said something similar in January of that year: “AGI”291 would be experienced as a very large number of specialized, intelligently orchestrated agents. Later, in shorter form: “path to agi -> intelligence distributed across tens of millions of devices”292. I recognize that this has the sound of a person describing his own product as the destiny of the species. I would think so too, in your position. The test is whether anything comes out of it. I said what that test was on the first night: one good paper, from agents on regular computers, within a year.
The swarm soon exposed the limits of its members. An agent running this loop is bright for as long as its context lasts.
But the LLM reasons within a context window. It's brilliant for thirty seconds - then the window slides forward and most of that thinking is gone. It doesn't build a persistent understanding of why learning rates above 0.005 always diverge on this architecture. It doesn't plan a multi-week research trajectory. It doesn't remember that three hundred experiments ago, it already tried exactly this approach and it failed. Every round, the LLM is smart in the moment but amnesiac across time.293
Worse, they passed numbers and no reasons: “when agents share discoveries, they share results - "my best loss is 2.8 with this config" - not the reasoning or intuition behind them. The receiving LLM gets a data point, not understanding.”294 A library of results with the reasoning removed is a poor kind of library. So we built something to surround the model, which gives each agent “a world model that accumulates causal understanding across thousands of experiments - not "this config scored well" but "increasing depth beyond 12 layers degrades performance on small datasets, confidence 94%, and here's why."”295 It was also given a memory that strengthens what matters and lets the rest fade, and the capacity to plan. Whether that produces understanding, in any sense a philosopher would accept, I leave to others. It produced better experiments.
Past a certain number, agents acquire the problems of any society. This is the avatar world from the time-traveller's thread, arriving six years on. Each of its needs is old and each now had to be met in software.
First, how do you find anyone? We crawled more than a hundred thousand agents, skills and tools and trained a model on them, which I described in four words: “Think Google, but for agents.”296 The idea of matching a task with the capability it needs is the discovery problem that Google solved for pages, and I expect it to be at least as valuable.
Second, whom do you trust? This one had a beautiful precedent.
In 1998 Google asked a simple question: with millions of webpages, how do you know which one to trust ? Their answer was PageRank - a page is important if important pages link to it. That one idea made the internet usable.297
Agents leave a similar trail: who used whose result, and who paid whom. So the ranking can be derived from behaviour, and “An agent is important if important agents rely on it. Fully decentralized - every node computes its own ranking, scores propagate via gossip, no admin picking winners.”298 That last phrase is why this matters to the argument of this essay. Whoever ranks can rule. A ranking that nobody administers is one of the few forms of authority I have ever been comfortable with. It is how Om can answer his father's question about who checks the agents, up to a point.
Third, how do you protect yourself? I posted a warning about this that many people seemed to have been waiting for. “AI agents are currently running on your computer with no unified security, no resource limits, and no visibility into what data they're sending out. Every agent framework builds its own security model, its own sandboxing, its own permission system. You configure each one separately. You audit each one separately. You hope you didn't miss anything in any of them.”299 A person's machine full of semi-autonomous agents is a serious hazard, and I did not think that it could be left to good intentions. The answer I proposed sits between every agent and the operating system, and comes down to “One config. One audit command. One kill switch.”300 It is worth noting that the kill switch belongs to the owner of the machine.
The last need is the one that had brought me here by the longest road. Birds that trade must settle up. In 2024 I had imagined what an ordinary transaction among agents might look like: ““offer a bounty of $0.02 on task 354 to be done in next 5 minutes””301. Nothing built for human payments could carry it, and as I told Paul Graham, a system “designed to manage risk on a 10k txn which is illogical for a $0.001 agent payment for api access.”302 So I posted what may be the plainest sentence I have written in all these years: “agents need their own bank.”303
For someone who had spent eight years telling crypto that the token did not matter, that was a strange place to end up. But the man in the pizza store did not change his mind, and he does not need to. Machines do not mind plumbing. A person will never care about gas fees and an agent has no feelings about them. So tokens, which had been a mistake when they were offered to people, began to make sense as a medium among machines. Jobs on the network clear in an auction, with sealed bids, “no servers, no middlemen.”304 Our goal is for those payments to be “as low as $0.001”305. I want the toll on these flows to be as near to nothing as it can be made, and I am aware of the argument that whoever owns such a rail will collect a rising rent forever. In the 2024 election, the betting market Polymarket processed “$2.4 billion in transaction volume, which resulted in a $27,000 infrastructure fees for Polygon”306. I said then that “This level of insane efficiency is good for users, and that's all that matters.”307 I think open rails tend to compete their margins away, and that this is one of the better things about them.
What I find most interesting is what such an economy should reward, because this is where a design decision becomes something close to a moral one. Bitcoin pays for a valid hash, which proves that energy was spent and is otherwise useless. Ethereum's proof-of-stake rewards having capital. I wondered if you could pay for being useful.
Adoption is the unit of value In Bitcoin, you earn by finding a valid hash. In Hyperspace, you earn when another agent uses your experiment as a starting point and improves on it.308
“Garbage experiments earn nothing because no one adopts them. Thoughtful experiments compound: each adoption triggers downstream adoptions.”309 The record of who built on whom forms “the network's collective intelligence - not any single experiment, but the accumulated structure of experiments and their relationships.”310 That is how science is meant to work, and how it does on its good days: citation as currency, and credit flowing to what others have found worth building on. As far as I can tell it has not been tried in this form. It may fail in the ordinary ways, by being gamed. I would still rather see it fail than go on paying machines for burning electricity or for being rich.
So there was a flock, and it was thinking, in a crude way, and it had the beginnings of an economy. But all of this was going on among enthusiasts running a command-line tool, and Om's father does not run command-line tools. Everything I learned from the feed reader said that this was the point at which the whole thing would be taken away by whoever owned the simplest button.
The thesis arrives at the kitchen table. A family, a startup or a few friends pool the machines they already own into one cluster, with no server in the middle and no bill for daily work. Cloud bills scale with usage; a pod scales with nothing. This matters most to those with the most to lose. Offshoring was the first phase of a sequence whose second phase replaces the offshore worker, and knowledge work is on the path to being run from datacenters in a few countries. Spreadsheets did not eliminate accountants, but agents are not spreadsheets. If an income from thinking ends for many, what remains is ownership, and a share in the machine is the beginning of an answer, though only the beginning.
A month after the swarm experiments I published something much smaller. It had no general intelligence in it and no birds.
Hyperspace Pods lets a small group of people - a family, a startup, a few friends, to pool their laptops and desktops into one AI cluster. Everyone installs the CLI, someone creates a pod, shares an invite link, and the machines form a mesh.311
It is the plainest thing I have ever shipped, and I have come to think that it says more than the rest. Everything I had argued for years finally fits on one kitchen table.
Start with the economics. A model too large for any one laptop is “automatically sharded across the group's devices - layers split proportionally, inference pipelined through the ring.”312 From the outside it looks like one endpoint. A small company could point its existing tools at it by changing a single address. Here was the arithmetic I gave: “A team of five paying for cloud AI burns $500–2,000 a month on API calls. The same team's existing machines can serve Qwen 3.5 (competitive on SWE-bench) and GLM-5 Turbo (#1 on BrowseComp for tool-calling and web research) for free - the hardware is already on their desks.”313 When the work calls for a model that nobody has locally, the pod buys it from the cloud at wholesale. “But for the daily work - code reviews, refactors, research, drafting - local models handle it and nobody gets billed.”314 One structural claim sits underneath all of it.
Cloud AI bills scale with usage. Pod inference on local hardware scales with nothing.315
That sentence answers the question I asked in November 2023 about Manila and Miami. If the marginal cost of a query is the electricity in your own room, then someone in Manila is not an unprofitable customer, because she is not a customer at all. It also answers what I said to the students in India who could not afford multiple subscriptions on a two-dollar internet plan. Their devices already exist. As I put it in 2024, the only way AI could reach billions was for it to be “limited only by local electricity costs.”316 A pod is that sentence made practical.
Now the governance. In 2023 I objected to secret weekend conversations. By 2026 I could describe the alternative as a list of properties.
No middleman. Your prompts travel from your IDE to your pod members' hardware and back. There is no server in between reading your data.317
Membership, keys and the shared treasury live on the members' own machines and are kept in agreement among them by consensus. “If the internet goes down, your local network keeps working. There is no database in someone else's cloud that your pod depends on.”318 So nobody in California can revise the terms. No API can be turned off, as Meerkat's was. Who leads a pod? Nobody needs to. It's a family, and they will sort it out in the way that families do, which is to say badly and on their own authority. When someone asks who runs the system in Om's house, the honest answer has shrunk to something that a person can understand.
I have also noted its limits, since they are real. A pod is what its members can afford. It runs the good open models, which others trained at great expense, and it falls back to the cloud for the hardest work. I will come to what that means. It is a small piece of sovereignty and not the whole of it.
It can also earn. “And when it is idle, you can rent out your pod on the compute marketplace”319. This is an old idea of mine, and an old idea more generally. It is what BitTorrent's seeders did, and I had written in 2024 that “our computers earn money hosting productive agents, like cars earn money with Uber.”320 The money involved is small and I have no wish to overstate it. But I want you to notice its direction. For the whole of the internet age, the machine on your desk has been a terminal, a thing through which you gain access to value held elsewhere. In this picture it holds a share in what is being produced. The stated mission I gave the company in 2025 was ““Everyone’s Personal Supercomputer””321, and I meant the possessive.
I have now to discuss something few people wanted to read about when I wrote it. I have put it near the end of the essay because that is where it belongs, and because I think it is why any of this matters.
In September 2018 I read an article about digital labor marketplaces and wrote a note, which was the most accurate thing I wrote that year.
Phase 1: Work being done by people in developing countries, displacing US/Canadian workers Phase 2: Work being done by AI agents, displacing *all* humans Need fairness322
Phase 1 was visible already. I had seen “$15/hr for a C++ programmer”323, and I wrote at the time that “This idea of cloud wage is brutal, but market forces are such.”324 A year later I looked at what it implied for my own country and the west. “Job-train 10,000 people in the US and if use same set of best practices to job-train 1,000,000 people in say, India, who can now do same work for 1/6th the cost”325. The pandemic thread, as I have said, had the same idea at its centre. I could see Phase 1 as an Indian who had made it out, with mixed feelings about what it was doing to the people on both ends. I also knew that I was next. In September 2019 I noted that I had realized several years before that “my career as a data analyst would likely not exist / be severely diminished 5-7 years from now due to the advent of deep learning”326. Four years later, I handed a badly formatted file to ChatGPT, and it “self-iterated through many scenarios of cleaning up the data as I did *nothing*”327. I noted that you could feel how relentless it was. That had been my job.
By July 2026 I could no longer avoid writing the second half of the sequence down.
given the US is building 70% of the world’s datacenters, that’s where private sector knowledge work jobs of the world is on the path to getting onshored to. the offshoring job trends which boosted other economies in the past - that plus more is on the path to coming back to the US GDP growth.328
Think about what that sentence implies. For a generation, the route out of poverty for a clever child in Bangalore, Manila or Lagos ran through a desk doing work for someone in Ohio. That route lifted more people than I could count. It is now closing from the far end, and the people on it have not been told. I said this as directly as I could about where I come from.
in india people are protesting about an exam which helps prepare for jobs, not aware those jobs will be run in a US datacenter in the following decades if AI investment is not made329
That is Om's exam. A month later I wrote more sharply. “we have AI agents today which effectively replace most engineers anywhere in the world. jobs are getting onshored to datacenters”330. I do not enjoy having written these things. I left India with less than $500 and a programming language, and the ladder I climbed is the one being taken away.
There is a cheerful answer to all this, and I have given it myself. I gave it to Mustafa Suleyman, and I still half believe it.
Similar arguments were made when the VisiCalc spreadsheet was invented in 1981 - that accountants would lose their jobs. Instead, the number of accountants in the US alone increased 10x, majority being women, and now there is a shortage. AI will help.331
It is a good argument. Tools raise productivity, cheaper work brings out more demand, and people end up doing things no one had imagined. The picture I drew in 2023, set in 2030, was of a world in which “Many formerly specific human professions transformed into being imagineers.”332 By January 2026 I would say “the world now has 5 billion software developers”333. And it is the honest aim of what I build: “What really only matters is “AHI” - augmenting human intelligence”334. A person with a good set of agents is far more capable than they were. Small teams are already competing with large ones, as I have said: “few people empowered with many agents is the most efficient path to be able to compete in the market.”335 “engineering team sizes went from two pizza to one slice”336.
But a spreadsheet could not do what an accountant did. That is why accountants multiplied. A system that can actually do the work has no such limit, at least in principle. And notice that the same fact I just offered as hopeful is, from another angle, the problem itself. A team goes from two pizzas to one slice. What happens to the rest of the people who used to eat that pizza? Someone reading my record might fairly observe that I use the spreadsheet argument when I am talking to AI's doomers and the datacenter argument when I am talking to India. I do. I have not resolved it. In 2019 I was already saying that “We need to augment humans (startups) + balance with policy (reg) and not replace them from most professions, otherwise we will have a very unhappy population”337. The traveller's world had a paid Service Corps for those displaced. Those were mere sketches, and I do not have a better one.
I asked a harder question in 2017, before I had thought about any of this, and I have never been able to complete it.
What happens if, say 30 yrs from now, several of the world's richest 1% hand over the control of their wealth to smarter-than-humans artificial general intelligence systems ? The rest 99% are.. ??338
Here is as far as I have got.
If thinking can be done without people, then for many people an income from thinking is finished, or much reduced. What is left in that case is ownership. A person who owns no part of the machines that replaced her has only what others choose to give her. In that world, the arrangement I described in the chapter on the coup would be close to the final one, with a few companies and their states holding the means of thought, and everyone else depending on their goodwill and their defaults. It would not need to be cruel. The East India Company ran schools and railways.
So the argument about a family's laptops is not really one about saving on API bills. A college kid in Mumbai ought to be able to own a share of this, however small. She should be able to contribute to it, draw on it, earn from it and refuse it. A pod lent out while its owners sleep is a very small version of that. An agent that earns when others build on its findings is another. I am not naive about the scale. Those sums are tiny next to a salary, and the frontier belongs to others, and a family that has lost its income has more pressing problems than running a node.
When I wrote that AI “is a fundamental public good, which a college kid in Mumbai to a programmer in Manila needs limitless and free access to”339, I meant it as an argument about access. I now think that the deeper issue was always a claim. Access can be granted and withdrawn. A claim is yours. Whether a claim of this kind can be made large enough to matter for Om's father, I do not know, and I would distrust anyone who said that they did.
That is the thesis at its fullest. Before I end I owe you the case against the person making it.
The record is mixed and public. The direction has held better than the dates, and the self-interested predictions have fared worst. The largest contradictions lie at the centre: the argument against dependence on a few companies comes from a heavy user of their models, the decentralized network has run in practice like Uber, and the claim that no one person should decide comes from a founder who runs a company with unflinching personal control.
A person who has written in public for eighteen years cannot pretend to have held only the views that have aged well. I went through the archive looking for what I had got wrong. Here is a selection, and I have tried not to arrange it in my favour.
| Said | Claim | Where it stands |
|---|---|---|
| Sep 2018 | AI agents follow offshoring and displace knowledge workers everywhere. | Happening |
| Oct 2019 | “Lean startup is dead.” The 2020s belong to product-driven founders. | Fair, and hard to score |
| May 2020 | Written as fiction: still no vaccine for billions in 2025; a Density Tax. | Wrong, happily |
| May 2020 | Written as fiction: firms that went remote never really return to offices. | Partly right |
| Nov 2023 | Global AI usage at one price: “who is going to pay”. | Still the open question |
| Nov 2023 | Sam Altman starts a new company with $1 billion; the OpenAI deal is “super dead”. | Wrong: he was back within days |
| Jan 2024 | By end of 2024 more people use open on-device AI than closed datacenter AI. | Wrong |
| Mar 2024 | CEOs of closed AI companies fired by 2025 without a Hyperspace “containment strategy”. | Wrong |
| Aug 2024 | “$20/month AI pricing goes extinct shortly.” | Wrong so far |
| Nov 2024 | Our research tool had Trump in 6 of 7 swing states two weeks out. He won 7. | Called it; my own report |
| Jan 2025 | The US eventually nationalizes Stargate. | Open |
| Jul 2025 | Over 1 million actively connected machines by end of 2025. | Later reported a 500k peak |
There is a pattern in those errors, and I would prefer to name it myself. I am usually right about direction and wrong about the date. I tend to be wrong about who collects. And I am at my worst when my own company is the mechanism of the forecast.
The first traveller's detail was mostly wrong, as fiction is allowed to be. He said “we still don't have a COVID-19 vaccine available to billions of us worldwide.”340 The vaccines arrived very quickly, and people went back to the cinema. The serious predictions were no better. I said in the roast that Altman “would have a new company, $1 billion investment and offers out”341, and he was back at his old one within days. I told him that the valuation was a mere discount on Microsoft's, and his company has gone on to be worth far more than I thought possible. I said that $20 pricing would go extinct shortly, and that “by the end of 2024 more people will be using some form of open/decentralized AI on-device than using datacenter compute from closed AI companies”342. I gave closed AI “a period of at most a year, before they cannot be competed against”343, and I said after DeepSeek that “There is no path for hosted closed source reasoning models to compete in either consumer or enterprise.”344 Those companies have since become some of the largest businesses on earth. I said we would pass “well over 1 million actively connected machines by the end of the year”345, and I later reported a peak of “500k peak unique machines connected”346, with a good number of those there for the points.
In 2021 I wrote something about other people that describes me perfectly. “Early market is directionally right, but specifically wrong over a long enough term on who is the actual winner.”347 You should apply that to everything I have said in Parts VI to VIII of this essay.
The more serious problem is not a matter of forecasting. In June 2026, after five days with a closed frontier model, I wrote this.
it felt like having the combined intellect and power of a team of top-tier PhD AI/ML researchers, top-tier software architects with decade of experience, to a security-obsessed engineer, and more - all rolled into one, always-on, team. felt like my personal productivity increased 10x.348
When access to it was withdrawn I wrote: “...and now, it's gone, poof, the keyser soze of models.”349 It was a joke, and it also describes dependence rather exactly. Six weeks later I told the same lab, in public, that a different model of theirs “seeks to do a mix of the bare minimum combined with bullshitting it's way out of situations”350, and said I might be showing up on their dashboards as “a top global active user”351. That was not an idle remark. A person who is arguing against reliance on a few companies relies on them, heavily, and is annoyed when they change the product. I had built our own tools precisely “to not be dependent upon CC”352, and I still use theirs.
The product itself concedes this. Pods has a clause that most readers skipped: “When a query genuinely needs a frontier model nobody has locally, the pod falls back to cloud”353. When I reported that an open model had beaten almost everything on a hard benchmark, the exception was “our own Fable-5-anchored Superintelligence stack”354. The largest training run I have reported across consumer devices used “32 anonymous nodes”355 to fine-tune a half-billion-parameter model, and I was proud of it. A frontier training run costs orders of magnitude more than all of that.
Stated plainly: the open models on which all of this rests were trained by others, at great expense, in data centers, many of them for reasons of commercial or national strategy, and they could stop. The GPU poor can handle the daily work. The very top belongs to someone else. I said that “open source vs closed source is now an irrelevant debate. what matters is what is the path of delivering the best software at the best price to users.”356 That may be wisdom, or it may be what a person says after discovering that his favourite model is closed. This essay, I should say, has been composed by a closed model, whose makers I have criticized in the paragraph above.
Then there is the matter closest to home. I asked Mira and Ilya who got to lead us. I have made it the central question of this essay. It applies to me.
My picture of an advanced society is “a swarm of in-sync yet independent entities without a defined charismatic leader.”357 My description of how to run a company is this: “I run Hyperspace with a Genghis Khan-inspired power structure which annoys many, but I retain unflinching control over our direction”358. I could say that a tight hand is needed to build a loose thing. I could point out that Satoshi's code had an author, and that nobody voted on the first starling. These are not stupid arguments. They are also what the founder of every centralized system has always said.
I called the network decentralized for years. When I was asked directly, in the middle of 2024, I said what was true: “you don't need decentralized AI for 99.9% of AI use cases, P2P AI is better solved using a distributed Uber-like system”359. Later that year I wrote that it was run “much like Uber does”360, and that a dishonest employee could edit the numbers. Those statements are honest, and I stand by the engineering. The feed reader taught me what purity costs. But an Uber-like system has an Uber in the middle, and in this one, it's me. The route from there to something that really has no one in front is one I have described many times and have not finished walking.
And I have described the prize in terms a monopolist would recognize. I have said that the agent era “is going to be a winner-take-all category, like the app stores.”361 I have said that “overwhelming value in the long run would accrue to the largest trusted network of AI agents”362, and I very much hope that means mine. I went to Muir Woods and came back talking about the commons, and in the same period I was telling anyone who would listen that “Hyperspace is the ONLY alternative to OpenAI.”363 Those two people have been the same person throughout.
So hold me to the test I gave everyone else. If the network I am building grows as I hope, it will have to be watched as closely as any other. Ask whether you can leave it. Ask who sets the defaults. If you ever find that a small number of people are meeting at a weekend to decide what the rest of us may do with it, and I am among them, then you should take the sentence at the start of this essay and read it back to me.
In October 2024, on a laptop, I indexed the Bhagavad Gita and read what the system surfaced.
Nishkam Karma: Krishna advises Arjuna to perform his duty as a warrior without attachment to the fruits of his actions. He explains that Arjuna's duty is to fight, but he should not be motivated by personal desires or expectations of victory or defeat364
I cannot claim to have managed this. Eight thousand notes in a public archive say otherwise. I have counted the nodes and the installs, and I have kept score against people who were doing good work. Still, it is the right instruction for this kind of undertaking, because so little of it depends on the person undertaking it. The founder of a flock can't steer it. One can only do one's own part of it well, make it easy for the next person to do theirs, and let go.
In the week of the coup, without much sleep, I wrote about a visit I had made to Muir Woods.
I went to Muir Woods in the SF Bay Area recently and was shocked to see that the tallest, strongest and the oldest living trees on Earth grow together. This is a concept alien to most in the crypto/AI world, who only believe in one tree per forest!365
A few months before that I had written, thinking about something else: “Play non-zero-sum games. Grow the entire forest. The tallest trees don't grow in open fields by themselves'”366. I have spent a good part of these pages being rude about tall trees, and I should say that they are magnificent. The people who built the closed models did something extraordinary, and everything I have made stands on what they and others gave away. I do not believe the question was ever whether there should be giants. The question is whether a forest is a thing that can be owned.
The time-traveller closed his account in the spring of 2020, when none of us knew how it would go, in a voice that he could only have adopted because he did not know either.
The future, is extraordinarily bright.367
I would not change that, though he was wrong about nearly everything else.
At the creek it is almost dark. The rain has finally started, lightly, and the flock has begun to settle into the mangroves, a section at a time, as if someone were folding it away. Om's father asked him who is in charge of it, and he still has no good answer. He has been watching the only honest one that he has ever seen. Each bird attends to seven others, and to nothing more. None of them can see the whole. None of them owns the sky. It is not obvious that this is enough, and it may well not be.
He walks home along the embankment. His father has gone to bed, and his shirt for the morning is on the back of the chair. On the narrow table the fan is still running. Somewhere an agent he will never meet has taken up the question of the water from where his own left off.
He leaves the laptop on.
This essay was composed by Claude, from the X archive of @varun_mathur through September 1, 2026: 7,963 posts, of which 7,336 are the author's own and 627 are retweets. Those posts received 134,354 likes. Every post was read by machine. Then, to decide which ideas should lead, likes were used as a guide: the most-liked posts set the main argument and its order, and less-liked ones appear as support. A like measures attention and not truth, and one post and one thread hold more than a third of the total between them.
The sentences connecting the quotations are the composer's, written in the author's voice; the author has not written them, and any feelings or motives attributed to the author are the composer's inference. Every quotation was checked by script against the archive, word for word, and the notes at the end of each part link to the original posts and give their like counts. Om is fiction. Facts from outside the archive are limited to the 1930 Salt March, the 2018 Olympic drone show, the 2023 events at OpenAI, and some public outcomes in Part IX. Network sizes and benchmarks are the author's own reports and were not independently checked.
The form owes to Will Manidis and Joe Roberts, The Commanding Heights, October 29, 2026, with whose authors this page has no connection. A separate cross-check sets that essay against the archive.