The founder opened with a small book. In his favourite bookshop in Abu Dhabi, the one with the good science fiction shelves, he had picked up A Year of Ikigai, a little volume with one piece of wisdom for each 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? You may be joyfully afflicted with sundoku, the practice of buying books and letting them pile up, which reflects your curiosity and your wish to stay close to the ideas that nourish your ikigai. It was his last week in the Gulf before flying home, where it was raining, and the room was small and familiar: the co-host, two regular members, and the lighting designer who had introduced himself a week earlier and came back with a presentation of his own.
Last week was about risk as the measure of our inability to forecast the future. This week he looked at how risk is distributed among the pieces of a venture. When you start anything, there must be an idea, a team able to execute on it, and the capital to support them, and each of the three may be present or missing to some degree. The missing piece is an opportunity for a particular model. When capital is missing, investors provide it. Hackathons let ideas and teams form together. Venture builders supply a team for founders who do not want to take the risk alone, or not at a particular stage of development. How much of each piece a party brings, and so how much risk it carries, shows up in the equity split.
In a world of abundant ideas, which AI makes easier to execute on, the ingredient that matters is the ability to test an idea in a market, finding people who are willing to pay for the product or the service and so confirm that they see value in it. AI is lowering team risk and capital risk, because a small team can now do much more with much less. It does nothing to answer whether anyone wants to buy. The usual answer has been to buy attention, bidding for it on Google, in the App Store or on Meta, and a lot of pitch decks arrive with a large share of the use of funds going to marketing. That is a zero-sum game. Many projects crowding into a scarce space of available attention raise its price for everyone, and the benefit that AI brings by lowering the cost of building is absorbed by the platforms that sell the attention.
There are at least a few ways around that paradox. One is the anchor buyer. Vaccine prices have been kept affordable through collective agreements to purchase a guaranteed number of doses. Frontier, a coalition of companies, committed to buying a billion dollars of carbon removal and in doing so created a market that barely existed. The United Arab Emirates has committed to delivering half of its government services through AI agents within a couple of years, which makes the state itself the anchor customer for agentic AI and guarantees, symmetrically, a private market of agents that will need to talk and transact with the government’s. A second way is opening markets that have been closed. Regulators and courts in Europe and the United States have started to stop Apple from taking its cut on purchases made outside its apps, on the grounds that the practice restricts competition. A third is the community, where a message circulates inside an ecosystem on the strength of trust built over time, and the founder admitted that a small community like this one is building exactly that.
The question he left open is what happens with agents, which will make it even easier to transact, not only to build. At least today, agents are neither swayed by advertising nor do they build the trust relationships that make it easier for humans to transact with each other. When go-to-market strategies and purchase decisions, and the choice of which tools get adopted, are driven or strongly influenced by agents, the picture will change, and he said he is eager to find out how. The tools for sending agents out into the world keep arriving. That same day AgentMail, which his agents use to send and receive email, announced a login-with-agent function on the model of login with Google or login with Apple, so that an agent can sign in to a platform on its owner’s behalf, and more platforms are supporting it.
The lighting designer started with an apartment. Looking for a new one, he had asked his agent to write to the listings, and it sent a hundred and one emails and received forty-five replies. What struck him was who was on the other end. The figures that came back were not produced by the people answering. They had read them off a screen and typed them into an email back to his agent, which compiled a dashboard for him. Forty-odd people were working as a low-bandwidth relay, from a screen to their eyes, to their brains, to their fingers, and back to a machine. The founder said he often makes the point more bluntly, that people are the slowest and least reliable component in any chain of communication, and that keeping digital information native, with high bandwidth and high reliability, matters a great deal. The same day he had asked someone for data and found them about to export it to hundreds of PDF documents and send them by email, when the same data could be exported as a single spreadsheet. Had he not been there to ask, the first path would have been taken. It is the fax machine all over again, receiving pages printed by one computer so that someone could type them into another, and a complete waste of human potential, because people have better things to do.
His presentation had been made the same way he works. That day Grok Bot gained a direct connection to Google Slides, so he gave it the Collective’s website, told it to learn how the site was laid out, and had a deck in about twenty minutes, by talking. Its subject was the week’s news. An agentic Linux distribution had received a grant of a million and a half dollars in tokens from xAI, and the founder made explicit that these are tokens for a language model, not crypto tokens. If you look for the equivalent of crypto’s volatility in AI, he said, it is a deflation in the cost of intelligence of around ninety-nine per cent a year, and it moves in one direction only. The team behind Hermes, the open-source agent harness the member uses and considers the best, had raised a large round from investors including Nvidia and Microsoft to bring it into businesses.
That led him to the organisation chart. He expects companies to be made of roles, as in principle they already are, with a human or an agent fitted into each one. Each agent will get the resources a person gets, an email address, a phone number, responsibilities, and be inserted into the company the way people are today. His own company is himself and some agents, and the agents increasingly do more than the human. The founder took it from the other side. If you had a team of people and they stopped working every time you looked away, it would be deeply disturbing. You trust them to believe in the mission and to push it forward while you are not watching, and it must be the same with agents. The ones to use are proactive, like OpenClaw and Hermes, which actively look for things to do, and whether you call it a heartbeat, a loop or a scheduled task, what matters is that they are allowed and expected to do it. His own were working as he spoke. The member’s run loops and listen through webhooks on email and WhatsApp, and they sit deep inside his systems. Building them took six months, which is why he keeps returning to how that accumulated knowledge could be preserved and moved.
The second half of his deck was about open source, and on this he was emphatic: closed source is dead, in his opinion, and open source is being reborn. That week a team had released an open-source suite meant as a full replacement for Adobe’s creative applications, largely vibe coded, and there is now an open-source replacement for SketchUp, lightweight and free. Both have MCP servers. Blender, which has been around forever and was always hard to use, is the one he has moved to, after years on closed applications for his work, because he can now build his own plug-ins with an agent. As it becomes easier to use, people with less knowledge build more plug-ins, and he described it as a snowball that is already rolling. He would like to donate his tokens to open-source projects he cares about, such as a civic portal helping people deal with public institutions. The founder confirmed both points. For anyone used to agents, an app without an MCP becomes almost offensive, because it forces you to click around the screen yourself, and computer use in Grok Bot is now reliable enough that even then you rarely have to. The magic of GitHub was never available to everyone before. Now a non-technical person who makes a little effort can tell an agent to download and compile one of thousands of valuable repositories, and an adventurous one can tell it to fork the code and change it to fit their needs.
He then turned to vibe coding. When Andrej Karpathy coined the term he meant it positively, or at least neutrally, but the developer priesthood quickly gave it a condescending tone, to protect its exclusive access to the dark art of programming from everyone else. At the beginning vibe coding was genuinely dangerous: you could hurt yourself and your wallet, and your API keys could end up exposed for the world to see. Over the past six months, though, agentic development tools have become so reliable that some of the best developers have flipped completely, some of them publicly. He described a video, which he promised to find, of the CTO of a well-regarded platform saying that he had been programming for twenty years, felt lucky to have lived through the era when chiselling software was an art, and had started delegating to agents six months ago. Today he no longer looks at the code his agents write. Developers who still sneer at vibe coding are hurting themselves and the people who listen to them, and the ones who updated their view as the tools changed deserve the respect they have earned.
The co-host answered from experience. He has worked in design, marketing and code for more than fifteen years, started making animations in Flash at ten, back when it still belonged to Macromedia, and remembers producing three or four client presentations a day. Every year meant learning the new version of every tool so as not to fall behind. Seeing the open-source suite during the presentation and hearing about the CTO frightened him a little, and left him with an internal conflict about the years spent learning. He is a firm believer in learning the basics before anything else. The day before, a prospective client had spent an hour proudly showing him forty-five pages of brand guidelines that Claude had written. More than twenty of the pages were unnecessary, and the rest contradicted the brand they were meant to describe. A developer brief, also written by Claude, contradicted the guidelines, and the client was ready to hire someone to “just fix the UI” and feed the whole brief back to Claude. The tools make a working day efficient, he said, but trusting them without clear basics is where it gets dangerous, and that project had not really started. That same morning he had spent two hours at a trade exhibition looking for clients and left disappointed, because every poster, every brochure and every stand looked the same, all of it generated by AI.
The founder told him he is in very good company. Musicians, visual artists and lawyers feel the same, but the deepest blow has fallen on mathematicians. A designer might spend three days on a brief and a week on a website. A mathematician can spend twenty or thirty years on a single problem, and an AI lab has just released solutions to hundreds of problems that people expected to spend their lives on. With formal methods, the proofs can be checked and leave no doubt that they hold. What mathematicians find when they read them is reasoning that is completely alien, correct but offering nothing to learn from, build on or be inspired by, and they have only begun to work out what to do next. He added that we will one day feel nostalgic for these months, when it was still possible to recognise an AI-generated poster and talk about it.
The lighting designer offered a car. A Ferrari is still a Ferrari and a Honda is a Honda, but now everyone can afford the Honda, and it keeps getting better, safe and pleasant to drive. He would not have spent two hours making a presentation for the room, but twenty minutes he would spend, so the evening got something it would otherwise not have had. A documentary on the same news would still take real effort and money. You must know what you are doing, but you can also build that background, because you have to talk your bot through what you want, and twenty minutes from someone who knows will beat twenty hours from someone who does not, until the market learns to tell the difference.
The founder went further and said that he embraces AI slop. When someone complains that what he sent them is slop, he asks why they still assume it was meant for human consumption. If it was made by AI and they do not like it, they can give it to their own AI and ask it why he sent it, whether there is value in it, what to ask him and how to reply. When he gets an AI-generated reply to something he wrote, he is delighted. What AI adds is variety as well as speed: asking for four presentations instead of one, or asking three or four AIs to make the same presentation and criticise each other’s until it improves. The music industry once claimed tens of billions of dollars lost to piracy, as if the person who copied a thousand songs would ever have paid for them. Human time and agent time do not compare either, and the hour everyone had just spent in the room was worth a thousand times more than any number of hours spent by their agents, or by themselves scrolling through AI content.
It was midnight in the Gulf by then. Before leaving, the co-host shared a tool he had come across that lets you visualise a stage lighting setup in 3D, which he had thought of for the lighting designer as soon as he saw it, and the founder asked him to post it on Discord, because the chat inside a meeting does not survive the meeting.
Thursdays, 21:00 Rome time. Join the Collective on Discord and the invitation will find you. Bring what you built.