Five of us. The founder opened from Abu Dhabi, where he had spent part of the day in his favourite bookstore, a branch of a Japanese chain in a mall on one of the city’s islands. It has a strong science fiction section and he never leaves without a few books. This time he picked up a small one called A Year of Ikigai, a piece of wisdom for every day of the year, and let it fall open where it wanted. It opened on books. Do you have piles of them on your desk, it asked, in your home office? Then you may be joyfully afflicted with tsundoku, the Japanese word for buying books and letting them pile up, which reflects your curiosity and your desire to stay close to the ideas that nourish your ikigai. He took the serendipity as it came and enjoyed it.
Then he went back to risk, which had been last week’s subject, and asked a narrower question: what kinds of risk support entrepreneurship, and how do they distribute across the parts of a project? Anything that starts needs three things. An idea, a team able to execute on it, and the capital to support them. Each can be present or missing to some degree, and every missing piece is an opportunity for a particular model. Where capital is missing, investors supply it. Hackathons let ideas and teams find each other. Venture builders supply the team for founders who do not want to carry the risk alone, or not at a certain stage. How much of each ingredient is present, and who brought it, shows up in the equity split.
AI changes the triad unevenly. It lowers team risk and capital risk, because far more can be executed with far less. What it does not lower is the test that matters most in a world of abundant ideas that are now easy to build: whether anyone will pay. You still have to find people who see value in what you made and confirm it with their money. For a long time you could buy your way towards that answer by buying attention, bidding on search ads, on the app stores, on social platforms, and plenty of pitch decks still arrive with a large share of the use of funds set aside for marketing and go-to-market. The founder called it a zero-sum game. Many projects crowding a scarce space of advertising raise the price for everyone, so the savings AI brings to building a product are absorbed by the platforms that sell the attention. Build cheaper, and the advertising giants and the app stores collect the difference.
He offered at least three ways around that paradox. The first is an anchor buyer. Vaccines have been kept affordable through collective commitments to purchase a certain number in advance, and carbon removal got a market when a group of companies under an agreement called Frontier committed to buying a billion dollars of credits. The UAE’s commitment to deliver half of its government services through AI agents within a couple of years does the same job: the state becomes the anchor customer for agentic AI, and symmetrically guarantees a private market of companies building agents to deal with it. The second is open markets. In Europe, regulators found that Apple’s rules unfairly restricted competition, and the commissions it takes on purchases made outside its apps have been cut sharply as a result. The third is communities, where messages circulate inside an ecosystem of trust. He said, admittedly of a small group, that this one is building that kind of trust.
The open question is what happens when agents start buying. They will make it even easier to transact, not just to build, and today they are neither swayed by advertising nor do they take part in the trust relationships that make human transactions easier. He does not know what go-to-market will look like once purchase decisions are led or strongly supported by agents, and he is eager to find out. The tooling is arriving fast. That very day the service he uses to let his agents send and receive email announced a login with agent button, the same idea as logging in with Google or Apple, so an agent can now sign in to platforms on its own, and more platforms are accepting it.
The member from Dubai who had joined for the first time last week was back, on his phone, with a presentation he had sent the founder in advance. Before sharing it he told a short story. He has been looking for a new apartment, so he asked his agent to write to properties, and it sent 101 emails and got 45 answers back. What stayed with him was the other end of the exchange. Forty-odd people had sat reading numbers off one screen and typing them into another, sending back data none of them had created, so that his agent could compile it into a tidy dashboard for him. Human eyes, a human brain and human fingers, used as a very low bitrate bridge between two machines. He found it hard to see what those people were for.
The founder said he often thinks the same thing and says it more bluntly: people are the slowest and least reliable component in a chain of communication, and keeping information native, digital and high bandwidth matters a great deal. The same thing had happened to him that day. He had asked someone for some data, and the person was about to export it as hundreds of PDF documents and email them over, when the data could have been exported in one go as a spreadsheet. Had he not been there to explain, the PDFs would have arrived. It is the fax machine all over again, printing computer documents so that someone at the other end can type them back into a computer. A complete waste of human potential, he said, and everyone should be told they have better things to do.
The presentation itself was a demonstration. One of the week’s announcements was that the founder’s favourite agent tool could now work directly in Google Slides, so the member gave it the Collective’s website, asked it to learn how the site was laid out, and had it build the deck in the same style. He made it by talking, in about twenty minutes. The first item was a token grant: SpaceXAI had given a million and a half dollars in Grok tokens to Omarchy, an agentic Linux distribution, to fund its development. The founder made sure everyone heard the word correctly. These are not crypto tokens, they are model tokens, and if crypto is known for volatility, AI tokens have a volatility of their own that only ever goes one way, deflating by ninety to ninety-nine per cent a year. He finds it interesting to watch what carries over as people move from one kind of token to the other.
The second item was the lab behind Hermes, the open agent harness the member uses and rates the best available, which has raised more than ninety million dollars from NVIDIA, Microsoft and others to bring it into businesses. That, he said, connects to the founder’s opening. Once agents are inside companies, the org chart becomes a chart of roles, and any role can be filled by a human or an agent. His own company is him and his agents, and more and more he finds the agents doing more than the humans could. He expects agents to be given what employees are given today: an email address, a phone number, responsibilities, a place in the company.
The founder took that further. If you had a team of people and they stopped working every time you went for a bathroom break or looked away to concentrate on something, it would be deeply disturbing. You trust a human team to believe in the mission and to push it forward when you are not watching. It has to be the same with agents, and the problem is not that we need to learn to trust them, or that they stop when unobserved. It is that we forget how effective they can be. The ones worth using are proactive, agents that go looking for something to do, and he named the two best-known open harnesses as the examples. His own were working as he spoke, and so, he was sure, were the member’s. Call it a scheduled task or a heartbeat, what matters is that agents are allowed and expected to find their own work.
The member confirmed it. His agents run on heartbeats, are fully free to do whatever they want, and sit inside a loop of webhooks: every incoming email sets them off, and now every incoming WhatsApp message too. Which is why he keeps returning to last week’s worry about preserving what they know. It took him six months to bring them to this point, explaining things to them and connecting them, and he has connected them deeper than the newer tools do. They administer his workspace, one lives on the computer in front of him, the main ones live on a server with full permissions. Building that takes time, and once it is built it does remarkable things, because the more an agent knows about you the faster it understands what you mean. Six months in, he said, he is still amazed. Tokens, in his view, are the new gold, and because you can build so much with them he sees the rebirth of open source. Closed source, as far as he is concerned, is dead. That week alone brought a full open-source replacement for Adobe’s creative suite, vibe-coded, arriving into a user base tired of subscriptions, and a free, open replacement for SketchUp that anyone can contribute to. He would like to be able to donate some of his own tokens to projects he admires, and named a civic portal he had heard about that helps Italians communicate with their institutions. And Blender, the 3D platform that has existed forever, has come back to life now that agents can drive it. For years his work kept him on closed-source tools. Now he can write his own Blender plugins and finally use open software.
The founder confirmed both points from his own habits. Using agents becomes so ingrained that he now feels offended when an app has no MCP and he is forced to click around the screen like a monkey. When there is no MCP, his agent’s computer use has become reliable enough to stand in. After every Lab Night he restructures his remarks and agents do the posting, through APIs where platforms offer them, and on LinkedIn, which he said does not believe in computers, by having an agent copy and paste so he does not waste his life doing it. The second point was about GitHub. The magic of thousands of open repositories used to belong to programmers. Now anyone willing to make a small effort can tell an agent to download and compile one, and the adventurous can tell it to fork the code and change it to suit them.
The member closed with a slide made for the co-host. He had tried to work out which language to use for him and ended up putting in several of India’s, which led to a short exchange about how many there are: hundreds, and thirty-five spoken by more than twenty million people each.
The founder thanked him and encouraged everyone to do the same with whatever catches their attention, then made one more remark, about a shift in the last few months. When Andrej Karpathy coined vibe coding, he meant it positively, or at least neutrally. The developer priesthood quickly gave it a condescending tone, because it wanted to keep its exclusive access to the dark art of programming away from the unwashed masses. At first there was something to the warning: you could hurt yourself, your wallet, and your secrets, with API keys exposed for the world to see. But over the last six months agentic development tools have become so reliable that the best and most honest developers have flipped, some publicly. He described a video of the CTO of a well-regarded platform saying he had been programming for twenty years, felt lucky to have lived through the years when chiselling software was an art he was proud of, started delegating to agents six months ago, and has reached the point where he no longer looks at the code they write. The lesson the founder drew is about updating your priors when conditions change this fast. Developers who still sneer at vibe coding hurt themselves and the people who listen to them. Those who saw that what was unreliable six months ago can now be trusted deserve the respect they earned.
The co-host answered from the other side of that shift. He has worked in design, marketing and code for more than fifteen years. Early in his career, at a large consulting firm, he designed three or four client presentations a day, alongside reports and vector graphics, deep in Photoshop, InDesign and Illustrator. He started animating in Flash 5 at ten years old, back when it belonged to Macromedia, and watched Adobe absorb it. Every year meant learning the new version of every tool so as not to fall behind. While the member was presenting, he had been looking at the website of the open-source creative suite, and it frightened him a little. Together with the CTO who no longer reads his code, it left him with an internal conflict: years spent learning these things, and now watching people make things without them.
He is, he said, a great believer in basics, and he had a recent example. Someone had shown him a set of documents an AI had produced for their business: a brand guideline and a brief meant to guide the next piece of work. Much of the guideline was padding, what remained did not hold together as a brand, and the brief contradicted the guideline it was supposed to follow. Nobody had noticed, and the plan was to hand the whole thing to a tool and expect the right result to come out the other end. Using these tools to make your working day efficient is wonderful, the co-host said. Trusting them without the clarity of the basics is where it gets dangerous. If you have not started properly, how do you expect the end to be beautiful?
The founder told him he was in very good company. Musicians feel the same contradiction, and visual artists, and lawyers, the sense of having invested in a field and then watching AI persuade people they can do without that expertise. The most unexpected and most devastating blow has landed on mathematicians. A designer might spend three days on a brief and a week on the website. A mathematician can spend ten, twenty or thirty years on a single problem, and OpenAI has just released 722 manuscripts on problems people expected to spend their lives on. Modern mathematics can formally verify a derivation so that it leaves no room for doubt, and in the founder’s account what mathematicians find, where a result checks out, is reasoning so alien and confusing that there is nothing in it for a human to learn from or be inspired by. Correct, and leaving them with no idea what to do next. Not everything held up: within a day of the release OpenAI withdrew three of the manuscripts after an error was found, and only part of the set came with formal proofs. Navier-Stokes was only a month ago, and there are already seven hundred results. There are also rumours that important results are being held back. In spycraft, he said, information and its absence matter equally, and the shape of what was not released sits around cryptography. People are starting to conclude that breakthroughs there exist and are being kept quiet. He did not want to speculate further, but found the silence extremely significant.
The co-host had one more piece of evidence, from that morning. He had gone to an exhibition at eight o’clock looking for clients who might need his services, and found that every poster on every stall had been made by AI. The resemblance between stalls was obvious; every brochure and every poster looked the same. He left after two hours, disappointed. The founder’s reply was that we will laugh about this and even feel nostalgic for these months, when it was still simple, when you could recognise AI art and AI posters and talk about them.
The member from Dubai had a quick analogy. A Ferrari is still a Ferrari and a Honda is still a Honda. More people can afford the Honda now, it keeps getting better, it is safe and pleasant to drive, and if you want a Ferrari you can still have one. You still have to know what you are doing. His own deck took fifteen or twenty minutes, and by hand it would have taken two hours, which, with affection, he would not have spent on the room, so the evening had something in it that otherwise would not exist. A documentary about the same news would still take much more time and money. Background matters, and it can be built, because an agent system takes time to build and you cannot just ask it for a presentation without knowing what you want. Those who build without understanding will make something less than a Honda, but they will make something, and eventually the market will see the difference between twenty minutes from someone who knows the work and twenty hours from someone who does not.
The founder said he loves embracing AI slop. When someone complains that something he sent them is slop, he asks why they are still assuming it was meant for human consumption. If AI made it and you do not like it, give it to your AI and let it spend its time on it: why did he send me this, is there any value in it, what should I ask him, what should I reply. When someone answers one of his emails with an AI-generated reply, he is delighted, because it means they are exploring the directions that will matter more and more. AI also gives you variety as well as speed: ask for four presentations, not one, or ask three or four models for the same presentation and have them criticise each other until it improves, which nobody would have done by hand. That, he said, is what the studios and record labels never understood when they claimed fifty billion dollars lost to piracy. The people who downloaded a thousand songs were never going to pay for a thousand songs, because they did not have the money. Time works the same way. What we spend as humans and what our agents spend on our behalf are completely different currencies, and the hour everyone had just spent together was a thousand times more valuable than any number of hours from their agents, or an hour scrolling AI content.
It was midnight in Dubai and Abu Dhabi and half past one for the co-host, so the founder began to close. Before he did, the co-host shared a tool he had seen on social media and thought of the member for at once, a 3D visualiser that shows how a lighting setup will look on stage. The founder said he would be astonished if a lighting designer did not already know it, and asked where it had been shared. In the call chat, the co-host said. Post it on Discord, the founder told him, because these chats do not survive the call.
Thursdays, 21:00 Rome time. Join the Collective on Discord and the invitation will find you. Bring what you built.
Originally published on Substack.