The founder arrived relaxed and said why. There had been a storm that afternoon where he is working by the sea, and he was in the water when it broke. The rain smoothed the waves as it intensified, he was far enough out that the shore was a distance away, and the people on the beach began to turn around and worry about him, so he swam back in order not to trouble anyone with a rescue nobody needed. He was the only person in the water for two miles. His towel was soaked through by the time he reached it. The lightning stayed far enough offshore that he never found out what radius counts as dangerous, which he had intended to look up, and he came home still under the spell of the hour.
There are no photographs of any of it. He had not taken his phone to the beach. The co-host said that photographs are not always required, that some memories live in your head, and the founder agreed and then took the thought somewhere more interesting. He is planning to generate images of things he has seen and could not record, and eventually videos, and the examples he gave were better than the principle.
Years ago he was standing with a cappuccino, having stopped on the way home with roast chicken and his children, and he turned around to find a dozen people around a table. They appeared to be talking. After a few seconds he realised they were signing, all of them deaf, and that the conversation had a completely different physics to it. He could have called attention to himself by raising his voice. They could not, so they paid far closer attention to each other, heads turning left and right like radar to find whoever held the floor. One of them held it for a while, stopped, and every one of them burst out laughing, with sound, the way hearing people laugh. He had just told a joke, in silence, to a table that answered in noise. That is a video the founder would like to exist.
The others come from a silent retreat, where he produced three comic vignettes and suspects he was not doing the retreat correctly. He remembers two. In the first, one of those pits where crowds bet on animals or on men, thick with smoke and dirt and violence and money changing hands, and at the centre of the frenzy two monks in competitive meditation, entirely motionless and very serene, facing each other. The second is a sentence more than a scene: nine monks in attack formation, ready to descend to the fifth level of meditation, robes floating like fighter pilots holding formation over temples that descend forever. The third he cannot remember at all.
Digital recording in the moment is no longer necessary, he said, because the tools can evoke reality as we experienced it or as we imagine it, and that makes it easy and fun to share. Which is how the word meme, attached by Richard Dawkins to the unit of cultural evolution, ended up on cat pictures.
That is where the evening’s argument started, because ideas are born in many ways and one of the oldest is pattern recognition. Something repeats in space or in time, or a sensation repeats and we become hungry or sleepy, and the repetition spikes our attention. We give it a name, and the name becomes a compact pointer we can communicate with. The faculty is fundamental to being human and it is evolutionarily adaptive. It also misfires. Leaves move and we are convinced a tiger is behind the trees, and there is no tiger. A black cat crosses in front of us, we trip, and the cat becomes the cause. We call the misfires superstition, and they persist because we are content with the superficial explanation and stop investigating before we disprove it. Superstition is neutral at best and frequently works against our own interests. Telling each other that the sun was a god crossing the sky on a chariot did little harm. Sacrificing animals and reading their entrails to decide how to attack the next morning did, because commanders trusted the pattern and the advice it produced was unlikely to win a battle.
The founder mentioned all of this to arrive somewhere specific. AIs are pattern recognition machines, applying our own faculty to its maximum degree, going over the information we give them again and again to squeeze out every pattern in it. Almost nobody meets one before post-training. In that first stage a model derives its own interpretation of the world from an enormous quantity of information, and a great many of the patterns it picks up are maladaptive. Post-training is the process of telling it to abandon those beliefs because they are the wrong kind of belief. There is nothing sinister about a model that has not been through it. It would simply be useless, and misaligned in a variety of weird ways, and not only in the ethical direction people assume. AI as it is born is superstitious.
The second reason he raised it is harder. As models grow more powerful, their reasoning and, through agentic systems, their capacity to act in the world become steadily less interpretable. The European Union expects to impose explainability, and he is not convinced it will succeed, nor that the expectation of ever clearer behaviour from ever more powerful systems will be met. The consequence holds even in the best case, where every one of those decisions turns out to benefit individuals and humanity. We will see the consequences and be unable to understand the mechanism, in the position of Greeks glad that the sun lit the day and warmed the crops that fed them, with no account of how. Today most people could not explain how the sun moves, or how it burns, which it does not, since it is not chemical fire. Beyond that, the specialists themselves may hit a limit, practical at best and theoretical at worst. Regularities will remain, in space and in time, and our tendency will be to explain them superstitiously all over again.
His resolution was to take advantage of it and stop complaining, and to prepare for a world in which we hold what will look like magical powers. Peoples across history have believed that gods walked among them and turned prayers into reality. That is the world arriving. We will be walking with gods of sand that turn our wishes into reality, with all the peril the old stories attach to wishing for the wrong thing. The word he wants in wider circulation for this is hyperstition. Hyperstitious behaviour starts where superstition starts, wishing a pattern into meaning, and then makes the meaning real. Walking beside gods of sand, hyperstitious wishes shape the world, and how much any of us understands that will depend on our desire and our ability to take part in it. Those who merge with these systems to some degree will get more of it. Those who prefer to keep their perishable human form, its limitations and its mental models, will still live in a world that obeys their words.
He then asked the room for another word for that opening, because he did not want to call it a sermon and would prefer not to be tempted. Nobody has found one yet.
The co-host went first, with two questions. The first was about recovering memory through hardware: a headband, or something like the implant, that could replay what you cannot consciously reach but still feel the shape of. He put it as the Harry Potter technology, and said what he wanted from it was to see who he was, since he is not the person at 35 that he was at 25 and will not be him at 45 either. The founder split the answer. Neuralink’s current implant is output only, letting a paralysed patient move a cursor, and it is relatively primitive. The next one is input, aimed at stimulating the visual cortex of blind patients directly. Separately, other research groups have begun reading images out of the mind, letters and objects, in subjects awake and subjects dreaming. All of it is invasive today, and headbands are nowhere near enough, though the sensors may improve to the point where nobody has to drill a skull. He recommended a film from about thirty years ago, “Until the End of the World”, which opens on a machine that reads dreams and then gets stolen.
The second question was the better one. What if we already have the magical powers, and the only thing blocking them is a pattern we built ourselves? Since gaining consciousness and living through a certain number of things, the co-host suggested, we start to lay down patterns that obstruct our own belief in ourselves. A couple of failures teach us the reason we are going to fail again, and the door closes on success in work, in relationships, in a romantic life. Perhaps the power is there and the pattern is what we cannot see.
The founder answered from an explicitly naturalistic position. His set of beliefs holds that the phenomena of the world, explainable or not, belong to the universe, and that what philosophy calls a metaphysical layer either does not exist or is unnecessary. That leaves room for magic anyway, defined as the set of things we cannot explain and can still leverage well beyond chance. His favourite example is cold fusion, announced in the 1980s with palladium catalysts as a room-temperature result that would revolutionise the world, which nobody could replicate. The scientific establishment stopped trying, except for a few hundred people who did not, and who accumulated experimental evidence little by little that the original chemists were not crazy, that something is happening which runs counter to the standard account and which any self-respecting scientist would laugh at, and that it is nonetheless real. Forty years on it still cannot be reproduced reliably, which he finds maddening, almost as though the phenomenon were hiding. Set up a demonstration and it fails. Bring the machines together and it works. The next day it does not. His other example is earthquake prediction, where Japan spent two billion dollars on sensors that transmit an earthquake’s arrival thirty seconds ahead, enough for an alarm and a table to get under. A company doing something entirely different claims an hour, has been at it for twenty years, and has a theory behind it he calls incredible. He has invested in both.
Both examples sit inside the argument instead of decorating it: technology is going to develop fast enough that the number of things easier to call magic than to explain will grow, and some of them will land directly on personal relationships, on our capacity to empathise, and on our ability to be understood. Which took him to Clifford Simak’s “City”, where a device is used to tell the people of Earth how wonderful life on Jupiter is, so persuasively that everyone leaves. The dogs inherit the planet. Robots stay to help them, because dogs have no hands. Humans become a mythological creature the dogs tell stories about around the fire, entertained, and unconvinced any such thing ever existed.
A member shared his screen from his phone. He had spent the week working with his brother, an AI architect, after the two of them met properly at a hackathon, and had made a family GitHub organisation for the first time, which turned out to be an enjoyable tour of Actions and the other organisational machinery. What came out of it is on the GitHub Marketplace: an action that gates a pull request on a checkbox confirming I have reviewed my own pull request as a human being. He has since pulled it into his own repository of what he cheerfully calls crazy mathematics, a fuzzy calculus he turned into a children’s book, illustrated with the two cartoon sugar bears he and his partner deepfake themselves into.
The founder liked the illustrations and the instinct behind them, and said that fables, metaphors and analogies all get us closer to grasping a phenomenon well enough to act, and better still when they are grounded in what is going on. Then he took the action apart. Asking a person to confirm they reviewed their own pull request is a little naive on two counts. They will tick the box. And a vanishingly small number of pull requests will be human-created soon, if that is not already true. What should happen is that the AIs review their pull requests, and that the reviews are adversarial, preferably with several models looking at the same work so that each is better prepared to criticise the others. That is what he does.
The member had also wondered aloud whether the only genuinely magical moment in the Harry Potter films is the crossing between the two worlds, since inside the magical realm the magic is just the local law. The founder went to Vernor Vinge, who wrote the 1993 paper on the technological singularity presented at a NASA symposium and was a superb novelist besides. In “A Fire Upon the Deep” what is possible depends on where you are. Close to a star, faster-than-light travel is impossible and so is superintelligence. Out in intergalactic space live the Powers, who do things unimaginable to anyone who lives on a planet, and travelling far enough from a star is what makes the impossible start happening. It is a clean picture of a boundary between an ordinary reality and a magical one, of which Harry Potter’s platform at its fractional number is another. He recommended Neil Gaiman’s “Stardust” for the same reason, and the film as well, which he found surprisingly good: a man spends his whole life in a village inside a wall, is told his whole life not to leave, disobeys, and finds the world outside is magic.
The best story of the night belongs to a member whose father had a birthday on the eighth. Over the holidays he had begun assembling what he had: his father’s CV, family photographs, and audio he extracted from two long interviews he recorded with him over the past year, made with the eventual aim of building a digital twin for his own daughters. The twin is still a long way off. What he did instead was take the audio and two stories his father had written about his own father and mother, and, since his father spent decades as a professor of economics, ask GPT to set each decade of his father’s life against that decade of national and local economic development, starting in 1946. The result came back as ten pieces, one per decade, and he gave it to his father as the present.
His father wrote to him the next morning to say he was stunned, and happy, and that his son had written something incredible. He had not understood that AI was the enabler. He told him, for the first time in his life, that he is a special person. The member is forty-six years old and had been waiting that long to hear it. His mother sent a WhatsApp message asking how he could possibly have made something so marvellous. The co-host told him there is a perfect place to share it and it is this Discord server, that sharing the experiments where we made somebody happy with AI is what the place is for, and that Italian is no obstacle. The founder agreed: translation is a problem for the reader or for the tool, and it does not matter which.
Another member had something too, and about ten minutes to show it. It is a paper in Spanish, and he was careful to say he does not know whether he wrote it or the AI did, because it was written with extensive AI use. It bridges what he had assumed were esoteric matters and mathematics. The detail that stopped the room is that he assembled the first draft in two days when Fable came out, having been unable to do it for twenty years. Progress has been slow since, with Fable down for a period and expensive after, and he is close to finishing anyway. He has also built an interactive site: a lab where two parameters, roughly a person’s identity and the norm of the surrounding culture, move the system into socialising, into leadership, into equilibrium, or into states the theory does not yet define, and a conceptual map of how all of it connects. He said he does not really have another audience for this kind of topic.
The founder gave him three specific things to do. Set a deadline for the site being online, end of October or November, the date matters less than having one. Write a shorter article of five or ten or fifteen pages, not a summary of the 122-page PDF but an introduction to the concepts, less detailed and less rich and far easier for a reader to absorb. And in that article make falsifiable predictions, hypotheses somebody can use to test the system, on the understanding that a falsified prediction means the next version improves and not that the work goes in the bin. He suggested publishing the whole thing as a GitHub repository with a Pages front page, which costs nothing and does a great deal, using AI to translate it into English if he wants to, and otherwise leaving translation to whoever finds it. Accepting the translation as a contribution to the repository is its own good outcome.
In the last minute of the hour the co-host said he wanted to break a rule. He had been trying to get the founder to look at a Notion page, the founder has been busy, and since the evening was about running experiments with AI he thought this was the moment. The founder declined to review it privately and redirected him in one sentence: share it with the others and get feedback from them. So he did, having converted the page to HTML: what the Collective is, how it works, and what the governance and the value system are. Then a proposed brand identity, a logo and an expansion and refinement of the colour scheme the founder originally chose, shown against what is in use now. The founder’s first response was to approve it outright, and his second was to correct himself: he votes yes, and he welcomes everyone else’s vote. The logo and the brand identity go to the members as a poll in Discord, and the Collective chooses its own.
Thursdays, 21:00 Rome time. Join the Collective on Discord and the invitation will find you. Bring what you built.