Eight of us, the largest room the weekly hour has had, and for the first time a woman in it. The founder joined from the island of Elba, where Napoleon was held under house arrest in two beautiful villas, and where he is spending two weeks working, eating fish and swimming in the sea. Other people would call that a holiday. He does not, and the reason turned out to be the first argument of the evening instead of a pleasantry. He has called himself jobless for twenty or thirty years, because the common definition of a job is the thing you do in order to earn the money that keeps food on the table, and he does not accept that definition. Whatever he does still supports him, his family and whoever else depends on him. He refuses the word that would file it under that description, and the consequence of refusing it is that he is never on holiday either.
The same refusal, an hour later and at a much larger scale, produced the idea he had brought to the room. Everyone in startups knows what a unicorn is, a private company valued above a billion dollars, and everyone knows the name came from the rarity of the sighting. They are less rare now, for three reasons that have nothing to do with the companies getting better. The dollar inflated away its own value. Some fields need enormous capital before there is a viable business at all. And the public markets raised the bar and the cost of floating, so fewer companies list, they list later, and they need more private money to get there. Above the unicorn came the decacorn at ten billion. Above that, nothing has quite arrived: SpaceX AI touched 1.7 trillion at its IPO but was under a trillion at its last private round, and Anthropic and OpenAI are both approaching a trillion while still private.
He reached for the word for a thousand unicorns and caught himself mid-sentence. A thousand of them is a kilocorn. A megacorn would be a quadrillion dollars, which nothing on this planet is. The prefix carries the whole argument, so he fixed it before going on.
Since open claw shipped at the beginning of this year there has been steady talk about agentic AI letting a single founder build a unicorn alone. He thinks that is misguided, and he was blunt about why. It replicates and multiplies the Silicon Valley mythology that celebrates extreme sacrifice and self-immolation, a mythology whose accuracy is demonstrated by how rare the outcome is. New tools mean more people can attempt it, with one in a million succeeding and likely fewer. That makes a Hollywood film or a novel that needs to hold your attention. It cannot be a recipe for the balanced and scalable life this group exists to work out.
What the tools genuinely do is create value, and if the admiration is misplaced the capability is not. So he inverted the prefix. There is a millicorn in everyone’s future: not a thousand times a unicorn but one thousandth of one, a private project which, if you valued it, would be worth a million dollars. He was careful that project is the right word and business the restrictive one. A stable multiple for a small concern is one times revenue, so a million dollars of revenue a year at a ten or twenty per cent margin is a millicorn. In much of the world that is a great deal of money, and in some places it has unfortunately stopped being so.
The question is whether that is within reach of practically anyone with a computer or a phone, surrounded by an army of experts. Analysing the opportunity, designing the business model, branding it, building the go-to-market, designing and testing the product, working with suppliers, standing up whatever commerce plumbing delivers the thing, supporting the customers who buy it. Every one of those functions is now either done by agentic AI or strongly supported by it. Multiply a millicorn by the eight billion people alive and set the result beside the size of the global economy, and the scale of the shift is visible without assuming anything exotic. No radical revision of how we think about the future is required. It is existing AI applied to existing processes and existing businesses, and it already delivers a large increase in well-being wherever it lands.
The first pushback came immediately and it was the right one. If agentic workflows are what make a private project reach a million dollars, one member asked, are those same workflows not already flooding the world with repetitive content and repetitive projects? Look at the websites AI builds and they are the same site in different colour combinations. Look at the content and it is the same content. When people started using these tools it felt different, because someone who could not write was suddenly able to put their own thinking into readable prose from a few bullet points. Now it has made people lazier, to the point where they publish without reading their own output. So among the millions of projects that already exist, how do you find the one that has value?
His first answer was uncomfortable. Anyone who has used and produced AI content has primed themselves to recognise it in the work of others, and can respond to that recognition with sympathy and compassion or with condescension. The only reason you can tell is that you have done it too. The second answer was lighter. The ages of AI are accumulating in ever-shortening cycles, and each one has a visible taste. There was a period when AI was in love with neon, when everything was neon and futuristic. Now everything is beige with a little orange, and he is curious to watch the waves of machine taste keep arriving. He is not a designer, takes what is available, tries to keep it consistent, and is not worried that the result is legible to others as the product of a given moment’s capability.
The third answer is where the argument sits, and it is not about AI at all. Our sense of what a startup should be doing has been skewed for decades by a perverse incentive. Investors need startups to pursue impossible goals, because only a sufficiently large multiplier at exit hedges the losses on everything that failed. That requires founders to live in a constant state of unbalance and fear. Take money and hold it in the bank against a downturn, and the question comes back: why did I give it to you if you are not spending it? Go out, capture the market share, and when you run out you will raise the next round, and if you cannot raise it, too bad, because other startups will. A millicorn is not built to give anyone those returns. It probably never takes outside investment at all, perhaps a thousand-dollar loan from an uncle or a grant from a foundation, and then it grows organically on the strength of its agents. A business that grows that way and does not need investors already has a name in investor circles, and the name is a slur dressed as a category. It is called a lifestyle business, as though it were unworthy of the passion the people building it put into it, purely because it does not need the money. He read that as sour grapes: it cannot be a good business, because it does not need my money.
One consequence follows directly. Far more of these projects will be local rather than global. The standard accusation from an investor is that the problem a founder has identified is real but merely local, with no path to scaling worldwide. Under the millicorn that stops being an accusation. Solving a local problem well, and inspiring another million people to solve theirs, is the outcome. An entire ecosystem will grow around it, with incubators and accelerators and consultants and advisors, and all of them will be AI. What nobody yet knows is which features make a problem a good candidate for this, and he put that question to the room unanswered.
Then the founder turned the argument against itself. The group assumes people should move towards an entrepreneurial disposition and believes everyone would benefit from it. But a great many people value security above the eventual rewards of a risk they do not want to take. Are we deluding ourselves? Or do these tools lower the risk enough that someone who never wanted to carry it will now try?
He put the question to the newest person in the room, who had introduced herself an hour earlier as the opposite of jobless: she has a job, she likes having one, and it gives her a sense of security. She leads people, learning and culture at a nonprofit that builds the capacity of government school teachers across three Indian states, work she came to through psychology, fashion and two years of teaching. Her answer separated two things that usually get collapsed. She does not take the kind of risk under discussion and prefers security, and inside her role she takes calculated risks constantly, which is where AI has come in for her. Appetite also moves with life stage. The risk she would have taken five or ten years ago is not the one she would take now, and anyone trying to build the entrepreneurial mindset in others has to account for that rather than treating reluctance as a defect. She then took a risk with AI in her own work, with no background in technology or code, and built something for the organisation.
Her illustration reached back to 2020, when the schools closed and nobody had any idea how to teach online. Teachers who had never taken a video call were pushed into it in weeks. Her organisation has since built a chatbot where a teacher can say what is going wrong in their classroom and receive resources drawn from the government bodies they work with and from recognised pedagogies, delivered then and there. Getting from a teacher uncomfortable using a phone to a teacher using a chatbot took five or six years of accumulated, unglamorous risk-taking by a lot of people. Most of us try not to remember what those years were like, and she had just counted what they built.
Another member heard something in that answer and went looking for the connection. He is building a project to help people identify what they were born for and activate it, and over the summer he decided to start from the bottom, meaning NGOs and people who genuinely need the help, before going anywhere near corporations that would use it to add margin to a balance sheet. He asked about her foundation because he wanted to know whether the two projects touch, and offered to present his in the coming weeks. It is the second time he has found a serendipitous link in this room.
The demos began with a member who had put his website live and asked, before anything else, for critical feedback. He had also built a German version locally, in which every string is translated except one paragraph he cannot yet explain, with Italian and Spanish planned. Arabic defeated him, on orientation, alignment and a flow he found to be a different game entirely. Asked why not Hindi, he gave the answer of someone who has thought about it: Hindi commits you to Tamil, Punjabi, Gujarati and Marathi, because every state has its own language, and there is no obvious place to stop. The founder’s counter was Paytm, which serves India in thirty-two.
His second share was the more useful one. He had come across a repository called GTM skills and it changed how he thinks about outreach. The copy AI wrote for him had no value system in it, only a statement of what he does and why he is reaching out. Run through the skills, the model began criticising its own messaging, and he concluded that all of it should be scrapped and rewritten. The lesson he drew is the one the tools keep teaching and people keep missing: the model will always give you an output, so the quality has to come from a skill you hold yourself, or from finding the right set of skills to feed the agent. He has since found a book-to-skills repository that turns any book into a skill an AI can think with, and offered to feed the founder’s own book into it to see whether it starts sounding like him.
A third member picked up the language thread from an angle nobody had raised. He has been joining community builds in agentic work, and has got his mother and his brother onto a hackathon build team. What interests him is that if languages parse differently, the architecture itself may be introducing bias, and other architectures already exist that developers use in other contexts and could be borrowed. He cited a recent paper out of Google and is hoping to develop something less linear with a family member.
The founder closed with the small computer that came out of its box two weeks ago, and a question he had set himself: beyond running local models, what else is it for. He showed a Mandelbrot zoom. The set is self-similar, so the little black shapes keep recurring however far in you go, and the equation that generates it is trivial and merely has to be iterated. He had prompted the machine to find interesting areas without ever defining interesting, and left it to decide. Across the top of the frame runs the size of the original image’s border as the zoom descends. At the point he showed us it was as wide as the solar system, and a minute later as wide as the Milky Way at a hundred and fifty thousand light years. The next step is running now. It needs twenty days of every GPU in the box to produce a picture that will look exactly like the one we were watching. When it stops, the border of the original image will be larger than the universe.
Three things about that struck him as strange. That we discover such objects at all, and calculate pi to trillions of digits, is already odd. That the objects exist in a state reality does not reflect is odder: the coordinates he was showing us can be recalculated by anyone and will produce the same geometry, at a precision deeper than the span of orders of magnitude between the largest and smallest things that physically exist. And then the third, which is that we build everything at three or four digits of precision. The bridges, the skyscrapers, everything we are proud of. Our best theories match experiment at thirteen or fourteen digits and there is no practical use for any of it. He called the whole thing a beautiful waste of detail. The exploration is the equivalent of picking a leaf from a tree in one random place on Earth, then picking another leaf in another random place, and finding them alike, which they are, and which they also are not.
He noted in passing that the computation is not the AI. The AI wrote the code, and the code parallelises across every GPU rather than making them wait on each other. His working pattern is Fable to plan, Sonnet to implement and Opus to test, with instructions to loop until the tests pass. What he finds funny is how much the agents care. Tell them anything and they report that it is the most important thing in the world, and he is content to say that he kind of cares about the result. For the next twenty days one of them will be busy calculating a few more minutes of a film. A member asked whether it might collapse into a black hole, which sent the founder to the Strugatsky brothers and a novel in which information density has an upper limit, and reaching it produces a small black hole that instantly evaporates into a literal explosion, so the researchers must approach the limit without crossing it. His computer has not become a black hole yet.
He left the room with homework, and it is the most concrete thing he has asked of anyone. He currently runs several dozen agents at once. Some of the communication protocols between them he designed and others the agents wrote themselves, they run on different platforms so the protocols have to bridge, and some channels are near real time while others are deliberately slowed so he can supervise what is happening. Reaching hundreds looks comfortable from where he stands. The question he wants answered next week is what kinds of problems will require tens of thousands of agents at the same time, and how we would go about analysing and addressing them.
Thursdays, 21:00 Rome time. Join the Collective on Discord and the invitation will find you. Bring what you built.