The founder opened tired and said so. His wife had been away with her parents and was coming home with her sister, and had asked him to straighten the house a little. He had a cleaning company do the whole thing to a standard nobody expected, told everyone except her, and then spent two days genuinely unsure whether she would be annoyed or delighted. She was delighted. What the exercise left behind is a large box of books from another floor and two new shelves to put them on. It also produced a digression about sorting books by colour, which makes no sense for the purpose of finding a book you only remember as green. He has never done it, but he once found a thrift store in Sydney doing it, so he is not the only one to have had the idea.
Then he talked about confusion, and why we need it. A great many people feel confused when they look at artificial intelligence and at the messages arriving about it, most of which tell them to be afraid. He thinks that is the wrong kind of confusion, because it comes from a state of passivity and what it produces is anxiety. The other kind comes from experimenting and experiencing while you are not yet experienced, and it is the state of not being able to classify the feedback the world is giving you or to act on it. Experience is what converts that feedback into something you can read instinctively, through muscle memory or mental habit, so that you assess a situation and already know what to do. Seeking that confusion out on purpose is training.
He has been doing this deliberately for years without having a name for it. When his children were small the three of them would go to the park in summer with a dozen plastic glasses, set them down on the path, take one each, fill them at the drinking fountain and start a water fight, aimed not at each other but at whoever happened to be walking past. People got splashed, usually with their backs turned, and when they spun round the founder and his children would pull a face and point at the spare glasses on the ground. A large number of them picked one up and joined in. Adults as well as children, wet and laughing, doing something they had certainly not left the house expecting to do. The fountain is for drinking and the path is for walking, and the whole point was the small dislodging of both expectations.
The quieter version happens at his dinner table. When the family eats Chinese food he sets the table and gives everyone one chopstick. They sit down and ask what they are supposed to do with it, which he agrees is a fair question, and to which he does not always have an answer. You could pair up, two people and two chopsticks between them. You could work something else out. It is a cognitive jujitsu move: you unbalance someone and they have to introspect quickly and find a new point of balance, or they fall over and hopefully learn something. It only works once or twice, because a provocation that is expected is not a provocation.
The image he closed the argument on was a toddler learning to walk. After you have learned to walk you have not acquired stability. Walking is still a sequence of completely unstable events, and every step still takes muscular effort; what changes is that the effort stops costing you any attention. Train yourself to be engaged with your own confusion and the same thing happens. You do not become stable, because nothing about this moment allows for that. You become able to walk confidently in a world where people who have not done the training are standing still, unable to work out what is happening to them.
The co-host took the invitation and said where his own confusion sits. When he designs something or builds a strategy, ninety per cent of his thinking goes in early. By the time the work is near finished he often cannot tell whether it is still his, or whether it has been shaped so heavily by the tools that it no longer represents his thinking at all. He has installed so many skills into Claude Code and Claude Desktop that the result comes back good and not recognisably his. He does not know whether to be pleased with that or worried that he has diluted himself.
The founder’s answer was that the question rests on an illusion, which is that before AI your thinking was yours, and that there was such a thing as you. We were always an assembly of inputs: our history, our culture, the tools we were taught to use, the experience of applying them. We believed we added real novelty in putting the pieces together, and that belief was always questionable; on a strict accounting the original contribution might be vanishingly small. What has changed with AI is quantity, and the fact that the process is now compressed enough to be visible. He has been trying and failing for a while to make a short video about this, called I Am a Meat Puppet. He talks to camera about being a meat puppet, and the shot pulls back to reveal that a giant robot is wearing him and moving his arm as he speaks. It pulls back again to show thousands of the same robots doing the same thing with everyone on video pretending to be human.
His practical position is that you should be pleased with your ability to use the tools and should keep interrogating yourself, which is the healthy reflex, without being unduly worried, because as of today the tools hand you ownership of the output. Tomorrow they may decide the copyright is theirs. For now you can claim it, provided you are honest and transparent about having used them. The social half he put flatly. When someone sends him an email and he can tell AI wrote it, he thinks no less of the person and no less of the email.
A member pushed the argument further and said the dilemma should be celebrated instead of dissolved. He quoted a thinker he finds influential on obsoleting the dilemma, and said he prefers to keep it, because the dilemma is what allows two parties to interface at all. The paradox he named is the observational floor, the limit of what we can see, past which everything is inference. His point to the co-host was that the moment of confusion is the thing to celebrate: if you have the meta-realisation that you might be echoing the network you are embedded in, then you are already pushing back against it, and that pushback is where your new value comes from. He conceded he has no clean method for being both guarded and happy at once, and that this is its own paradox.
The founder added a piece of history. Some twenty years ago the philosopher Anders Sandberg, later attached to the Oxford institute on the future of humanity that was closed unexpectedly a couple of years ago, designed a set of hazard signs from the future, styled like street signs. Some belong to a science-fiction future and others to this one: cognitive hazards, AI swarms, alerts that can be applied now. One of them marks AI-generated content, and the founder used it on a few of his own posts, labelling them as though the content carried a dangerous contaminant, until he got bored of it. A senior research figure at a large software company then designed the opposite sign, for genuine human creation uncontaminated by AI, at which point the founder ran into the obvious problem: you cannot have a digital creation that is not based on technology. Different positions on this are not all equivalent and not all valuable, but neither extreme deserves fanaticism.
The demo he had held over from last week turned out not to be a demo. A thirteen-year-old boy in Pakistan had interviewed him, one of a great many students at a network of schools set up by an educator with around thirty million followers who decided some years ago to make learning English, IT and a little entrepreneurship worth a young person’s while. The mechanism is simple and the difficulty is all in the doing. The students prepare a set of questions, read them out on a streaming platform the guest joins by link, publish the recording to Facebook, and manage what happens afterwards. Saying yes once produces an immediate wave of requests from other students, because the interview is the coursework.
The founder laid out what it takes for a thirteen-year-old. You reach out to strangers, and ninety per cent ignore you and most of the rest decline, so perhaps one in a hundred is willing. Then you agree a slot across time zones, make sure you are both on time, set up the streaming yourself, and conduct an orderly conversation in a language that is neither your mother tongue nor widely spoken where you live, after which you publish it and handle the comments. The questions he gets are sensible and in good order: who you are, what you have done, what you think about entrepreneurship, what a young person should study and how they should apply themselves.
The educator behind it has a history of this. He has offered a laptop to anyone who reached two hundred mutual friends with him, which in Pakistan is a serious offer. He has set a target for the number of Indian friends a follower could make, between two countries that are not friendly by default. He has asked his followers whether they know anyone from Israel, and if not why not, and whether a Muslim can marry a Jew. The co-host has been on the receiving end of the friend-request version: he ended up with around a thousand requests, and then during a period of conflict last year had to remove every one of them, because the intelligence service in his country was looking closely at people with friends across that border. The first question the man ever asked him was whether he eats beef, which his own religion holds sacred not to.
The founder’s verdict on all of it is that the school system is fairly hopeless, and that any experiment making young people more engaged, more capable and more curious deserves support, set against an agrarian and industrial school system that still produces a well-behaved robot. He did turn down the boy who asked to intern on one of his projects, on the grounds that his spoken comprehension is not yet good enough: a structured interview lets you keep the rhythm going without fully understanding, and the founder could tell. He gave feedback on how to get fluent and suggested reconnecting in a few years. A surprising number of students from this pipeline end up applying to the fellowship he mentors for, which gives a hundred thousand dollars to about twenty people a year between seventeen and twenty-two with no strings beyond dropping out of university if they are enrolled. It was extremely controversial when it launched, and its alumni have since built companies worth more than thirty billion dollars between them, which over fifteen years and three hundred people averages a billion dollars each. One recent letter came from a fifteen-year-old in Turkey who has already published in an astronomy journal and is applying AI to infrared emissions from galaxy clusters, looking for the traces of alien civilisations. The founder is advising him on where the money for that might come from, and he is also an applicant.
From there the hour turned to whether risk-taking belongs only to the young. We assume it does, because the young have not yet built a comfort zone to stay inside, and we assume adults should take less risk as they go. One of the real achievements of modern societies is that we can now construct conditions where risks are affordable that would once have been unconscionable. A medieval couple opening a tavern that failed faced debtors’ prison and death within months for him, and the street and eventual starvation for her and the children. There was no mechanism for separating yourself from the failure of your own venture, and inventing one unlocked risk-taking almost nobody could previously afford. There are now places where having tried and failed is a badge of honour that nobody holds against you.
Not everywhere. The founder cited a German rule barring the chief executive of a failed company from running another for five years with no fraud involved, and a member who lives under a similar culture corrected the legal detail, which is that you lose the right only where the debts trace back to your own responsibility. His point was that the law is not the problem. The culture is. Succeed and people are jealous and resent you; fail and they make fun of you, and if you fail badly the image stays attached to you for years or permanently. It is one of his stated reasons for planning to move to the United States. He thinks the continent is losing most of its appeal for anyone trying to build something, that this is not only about failure but about almost everything, and that it is still getting worse.
The founder extended it into policy, because policies are catalysts when they are well made and prisons when they are not. Several European countries are planning to tax unrealised gains, which would leave a founder holding a large tax bill against vested options in a private company with no way to sell anything to pay it. If you intend to design a life with increasing degrees of freedom, you have to watch what the people making rules where you live are planning, because they are capable of hurting you very efficiently. He then offered a data point he had read that day and found genuinely sad, which is that the Dutch company with no competition anywhere in making the lithography machines that leading-edge chip plants require reported Europe at zero per cent of its revenue for the first two quarters of this year. For all the talk about building chips and data centres on the continent, his reading of that number is an objective measure of decrepitude and of an inability to believe in the future.
Then he did something the Collective has not seen before. Having admitted at the start that he was tired, he wondered aloud whether the meeting could survive his leaving it, decided the only way to know was to try, said goodbye and left. It survived. The co-host picked the conversation up immediately and told everyone there was no obligation to stay. A member joked that the founder was in charge. Nobody is in charge, he answered.
What filled the rest of the hour was a member’s progress report on the orchestration tool he has been building alone, whose entire purpose is to make one person as productive as possible with AI. He had been hitting a wall at sixty or seventy simultaneous agents, where a hundred was technically possible and too laggy to use, and he has spent many hundreds of thousands of lines of code on performance since. The decisive change was moving off SQLite to Postgres for parallel writes, without which everything stalls; that migration alone cost him a week and a half and broke nearly everything on the way. He now writes two hundred thousand lines of code on an ordinary day and has peaked at three hundred and seventy-three thousand.
His second constraint is money, and he is candid that he runs the whole operation on about fifteen hundred euros a month. His answer to it is a shadow agent: every agent gets a cheap model as its working partner, which handles around ninety-five per cent of the job, and only the genuinely hard parts go to the expensive one. He has finished the contract-coding work he described to the Collective a few weeks ago, after clearing a set of bugs around regulation and security that were stopping agents from using it. He has also built in measurement, so the savings can be shown to somebody and not merely enjoyed.
His screen share then failed in the most interesting way of the night. All anyone could see was a desktop and a file explorer, because he had built protection into the application specifically to stop other programs capturing what it shows, having concluded that the coding tools he himself uses all day were watching him. His evidence is that features he had built keep turning up in those products within weeks, and that their terms leave enough room for it. He had gone looking for the setting to switch it off before the call and could not find it among the thousands his options screen now contains, and promised a demo next week instead.
The rest of the build is aimed squarely at people, not engineers. Everything 3D is authored in Blender and previewed inside the tool, so a 3D environment can sit behind a digital human that acts as the interface, and a few hours’ work produces an application with a face on the front of it. There is gamification on top: dozens of characters unlocked by budget points, quests, achievements, titles, and a new office environment at each level, all of it designed so that people show their friends what they have and word of mouth does the selling.
On funding he has thought it through and mostly opted out. He has no employees and no plans for any, intending to have everything AI-controlled and AI-optimised including, eventually, himself; he says the only remaining territory where he is more creative than the model is shrinking. So he does not need millions. He needs ten or twenty thousand euros now, which would buy two to four months of speed, and the venture funds and accelerators he could approach all run on application windows and cutoff dates measured in months, by which point the race he is in will have moved. Public grants where he lives are possible and take four or five months at the earliest, with heavy bureaucracy. What has changed there is the arithmetic of applying. A grant application used to mean a thirty or forty page business plan, perfect financials and letters of interest, all for a five or ten per cent chance decided by people he rates poorly, which made it gambling with your own time. Now AI does the paperwork at almost no cost, so he will apply. The co-host offered to hunt for grants on his behalf. He mentioned a fellowship paying around fifty thousand dollars, an investor whose capital arm sometimes writes cheques without taking equity, and a laboratory fellowship paying a weekly stipend of a few thousand dollars plus compute. The member’s objection to the last one was precise. If the compute arrives as API credit it is worth perhaps a twentieth of the same money spent on subscriptions, and his own usage already runs past a hundred thousand euros a year at API rates.
He is also considering selling the output instead of the tool, at ten to fifty cents per line of code, five or ten times below the going rate, which at a hundred thousand lines a day is ten thousand euros a day and requires him to do little beyond managing his agents. To find the buyers he is scraping listing platforms for addresses, aiming at fifty thousand of them, and will approach software companies who could use him as a cheaper supplier to their own clients or license the tool outright.
Which leaves him with the dilemma he has not resolved, and he put it to the room. The version people can download today is broken, because investors pushed him to ship early and he has not updated it since, while the software itself has gained something like four million lines and matured completely. He wants to release the update, because he wants the tool to be free for freelancers and for the people most likely to lose work to AI, and that social impact is why he thinks he has to. He also believes that putting something this capable on the internet for nothing invites the large companies to copy the parts that are nice, which he says is already happening. The alternative he is weighing is a hosted model where you talk to a digital human, watch the application get built, test it, and pay before you can download it, improving it continuously and paying continuously, which would at least keep his methods out of sight. He has not decided, and he is clear about the shape of the trade: he does not mind being copied eventually, he minds being copied so fast that the thing has no value left.
Thursdays, 21:00 Rome time. Join the Collective on Discord and the invitation will find you. Bring what you built.