[LAB NOTES](/lab-notes)Week 28 · 22 October 2026 # Done by Next Morning A fuller room than usual, with two newcomers in it, so the founder began with a light recap for their benefit. We meet, he said, to share and compare notes, to show what we are building and get feedback, all in the service of understanding a world that is changing fast. Very powerful tools are now available to anyone who builds a little awareness of them, and the barriers are practically gone: with a phone and an internet connection you have tools that billionaires and heads of state could only dream of ten or twenty years ago. We do it as peers, without the answers, sharing questions and concerns, knowing that an answer may hold for a while as a principle, but that whatever counts as best practice in capabilities and processes will be superseded quickly, and that conclusions we reached a few weeks ago will need re-examining. He usually opens with something he has been thinking about, often linked to what came before; then there is discussion; then people show what they are working on, signed up in advance or not. Two or three weeks ago it was risk, how we measure it, model it and hedge against it. Last week it was trust, the bridge across the gap between what can be measured and the imponderables, and what happens to trust in organisations that are AI-first or AI-native. The next concept, he said, follows naturally: autonomy. When you trust delegation to reach a goal through structured tasks, you are building autonomy into an organisation. That delegation is never arbitrarily large. It is proportional to the skills and resources of the delegate, which is why hierarchies form so naturally, since every participant has limited skills and limited resources. Actions meet reality, feedback arrives, outcomes vary, and coordinating the decisions about what comes next, whether to extend the delegation or rethink the structure, is hard. The harder it is, the more layers the hierarchy grows. So the question, as always, is whether bringing AI in natively changes a model that has held for thousands of years, and whether we can already see it changing. A single-person unicorn is still rare. A single-person millicorn, a million-dollar project born from the close collaboration of one human with an arbitrarily large fleet of agents, is more and more reasonable to expect. Agents distribute and coordinate better, concentrate skills on tasks and report back, and the friction that unavoidably emerges in human organisations, egos, politics, fiefdoms, does not exist in AI hierarchies, or not yet. He suspected everyone had noticed the relief of asking an AI to do something over and over, or five different ways, and then throwing the work away, without ever being asked why you made it waste its time and tokens on something you did not even care about. Try that with a human team and the difference becomes very clear. Hierarchy is moving into the software layer: orchestrators, and just as importantly reviewers that verify what an autonomous execution has produced, together forming a structure we are learning we can rely on, running at machine speed. He has to keep reminding his planning agents of that. They come back with what a human team would need: you will need a hundred people and a year. No, he tells them, we will be done by next morning. He turns out to be right, and hopefully they will remember. Verification is one end. The other is specification, saying what you want. He admits he is impatient and tends to under-specify, giving his agents terse, curt instructions, where a friend of his writes prompts two or three pages long. He prefers to iterate, and to ask the agents what he forgot, or, as the discussion goes on, what questions he should be asking, which is his motto. In his experience that is extremely useful, because they bring best practices, frameworks and approaches he never learned, the accumulated experience of a whole field. Specification at the start and judgment of the outcome at the end are the bookends of autonomous execution. What remains open is accountability: whether we can, and should, hand over agency, financially, legally or otherwise, to autonomous AI. That is where most people stop today, expecting it to stay with humans for a long time. He is fairly sure it will not, and that we will end up playing accountability theatre, rubber-stamping whatever agentic coordination decides. Where humans simply are not present it will be obvious. Pretending that a person hundreds of millions of miles away is accountable for what robotic explorers are doing around the moons of Jupiter will be a fiction. What happens between now and then will be very interesting to watch. He pointed to METR, the organisation that measures how long a task frontier models and agents can carry out at a given success rate. Last year its measurements showed that length doubling every seven months, far faster than Moore’s law’s two years, and in his view the curve is not just exponential but faster still, what he named some years ago the paradigm of jolting technologies, where the acceleration itself is increasing. His question to the room was where each of us stands on the longest tasks our agents run. He is at a few hours of autonomy, five or ten at the peak when the work is well organised and there is plenty of it. Weeks or months we can still grasp. Autonomous execution of tasks that take years, with agents following through and coming back with results years later, is something we have not begun to understand. Then the newcomers introduced themselves. The first, invited by the co-host from a startup community they share, is a product leader based in the UAE who works for a US company, and came looking for new people, new communities, and a reason to get out of his man cave now and then. He was a non-believer in AI for a long time, until a turning point he skipped over; now it is threaded through almost every process his team runs, and he is blown away every few days or weeks. Something that took them weeks a year ago, evaluating the relevance of search results at very large scale, they recently did in a few days, without even needing him. What he really wants, and cannot understand why nobody has built yet, is physical AI with a different interface: not a laptop or a phone where you keep typing prompts, but something like the device in Blade Runner that greets the main character when he comes home, present in every conversation, observing, contextual, stepping in when he has a question, a second brain that simply follows him around. He also put himself firmly in one camp of the argument about where AI is going. One camp expects a guaranteed income and a life of art and childcare. He expects white-collar work to disappear within a few years, his own product management included, and said it keeps him awake at night; he had nightmares about it a few days ago. The second, invited by the co-host and by a member he works with, runs a nonprofit in public education in India. He described himself as an early believer who likes to understand a technology as a whole before people jump on the bandwagon. He is excited about what AI can do in work done with very few resources, and thinks they are only scratching the surface, with much more to explore in ways that are ethical, respectful of people and impactful. He came to learn from what others are doing, and perhaps to find the courage to build an agent or two of his own. The founder answered the first newcomer directly. The things we do and the way we do them are going to change. In the short term we will still call it work and want to be paid for it. In the longer term, he hopes, more and more people will refuse the box altogether, the one with employees in it and managers telling them what to do, and that refusal will come when we are still rewarded for the value we create but no longer feel the need to call it a job, and the reward is not necessarily money. The question is how to help the organisations that support us today see this too, as organisations and not only as individuals, and often that will not be possible, which is when you look around for a place where the way you have come to think and work is understood. We tell AI agents what to do; a person with that ability should not be ordered around blindly, and that autonomy has to be recognised. There are trivial answers today to what people will be rewarded for, he said, mostly about human connection. Cars are faster than any Olympic sprinter and we still watch the sprinters, because what we admire is human effort and achievement. The answer is trivial because it cannot scale: not everyone can be an Olympian or a film star, and any selection process that admires the pinnacle cannot, by definition, include everyone. A few weeks ago the room reached a different conclusion, that there will be many local problems, and that we will not need to do what today’s startups are told and only chase what scales to billions. We will be able to afford problems that matter to a million people, or a thousand, or a neighbourhood, and to invest in our own lives and our communities, whether the community is geographic or, like this one, spread across three or four continents and ten cities. Everyone in the room, he said, is already investing, rewarding and gaining: investing time, rewarding each other with attention, and gaining knowledge and agency. That it is not monetised is true and temporary, either because it will not need to be or because we will find ways that are not clear yet. It will be turbulent and complicated, and he thinks it is the way to go. As for nightmares, he hoped the newcomer could digest them and see them for what they are: a useful alarm bell. The anxiety should push him to explore further and to seek out other people, here and also in his own company, people asking similar questions. What to watch for is whether you are somewhere these questions are welcome, or somewhere you do not feel able to ask them at all, which is its own alarm. The conversation turned to what a manager owes someone whose work AI can now largely do, and the founder answered from his own experience. Some months ago he was in that position, though in his case it was less about AI than about a person’s skills and whether they could acquire the ones needed. With support, he ran an experiment instead of deciding straight away. First, whether the person was ready to acquire the skill of acquiring skills, recognising that what they had was not enough and that they would be supported in improving. Second, he put them on a track where everything expected of them would be done through AI: the self-assessment, the plan of what to achieve and by when, the KPIs and OKRs and the measurement against them, and managing both down and up. They left. He thinks it was a humane way of letting someone see the challenge, and it also denied them the luxury of blaming him, because the challenge does not go away when they leave. Fire someone and you will be blamed, so steel yourself for that. Give them a chance at a different level and, if they rise to it, everyone wins; if they leave, they leave knowing what they need to improve. Not everyone can afford that approach, and many people are facing it from one side of the decision or the other. The better question, he said, is to ask yourself: if I were on the other side, could I step up if I were given the chance, and what can I do now to set myself up for it? Sound the alarm when half of what you do can be replaced by AI, not ninety per cent, when it is too late. Use it as a stepping stone to better work, not as permission to work fewer hours, which defeats itself. On education, he said, students will be fine, the way digital natives had no trouble with phones and computers. The bottleneck, the hardest part to change, is the teachers, and he asked the second newcomer how his organisation approaches that. Many teachers, he answered, are over forty-five or fifty and not even digital natives, so AI is a big jump for them. But in India, and in any low or middle income country, teachers clearly have more work than one person can do well, so anything that extends teaching time, improves the work, or takes off cognitive or administrative load helps. Much of AI in other industries focuses on efficiency, while in education the focus has been on effectiveness, on AI that teaches better, and he thinks that is the wrong entry point. Start by helping teachers do their work faster and by taking away the work that is not teaching. Because WhatsApp is so widely used in India, his organisation is testing it as the interface between teachers and AI models. An early challenge is that most models are trained on data from the Global North and lack the context an Indian teacher needs, so they are working on models whose answers fit the Indian context and the teacher’s real needs. The founder agreed, and said it applies in every industry: whatever AI does for the core of an activity, it is great at handling bureaucracy. A tax filing, a traffic fine, even the results of a medical exam that a specialist would once have mumbled about in jargon before prescribing something you had no context for. You can now quickly get the context to fill in the form, or draft the email to an authority in its own terminology so the request is actionable, or understand whether a doctor’s recommendation fits your case when the doctor may not have had time to absorb everything. That must be frustrating for some doctors, and it is enormously valuable, because the complexity of modern life is daunting for everyone, and AI can help us avoid the mistakes we make every day. On local models, he said, a country like India has to step up. Fine-tuning helps, but the real goal is models trained from the ground up on local data and more fluent in local languages than the frontier models, and the hard part is persistence: not one effort but ten years of sustained, sufficiently funded work that keeps pace, if not with the American models then with the Chinese ones. The big lesson from China is that great models do not need billions of dollars. Millions are enough, and next year it will be a few hundred thousand. No country can claim it can spend billions on its military and not a tenth or a hundredth of that on its own model. Kinetic warfare is still here but will matter less and less; cognitive defence and cognitive sovereignty will depend on every country’s ability to protect its own memetic sphere. He cited UPI as a great achievement, and the co-host, laughing, said he was sceptical of something else: that every new technology in India, the internet, crypto, now AI, seems to arrive first as a story about how people can scam each other with it. The founder’s answer was that scammers and the people they scam are a natural consequence of a rich ecosystem. You cannot let it run wild, but you cannot reach zero either; that is a totalitarian approach, and dangerous, because a society that achieves absolute compliance cannot evolve. So much contrarian thinking that was once illegal became legal and then the majority view, only because the minority was not crushed, and Prohibition in the United States was a totalitarian solution that failed until the country recognised it as a mistake. India, he said, has achieved many things nobody expected, and has sent a satellite to Mars. Stop telling yourself India cannot do things, he said; if you tell yourself you can’t, you will be right. There was one more piece of context before the demo. The co-host had emailed the founder a few days earlier asking for eight or ten decisions about the next version of the Collective’s website, and the founder had no intention of answering it himself. He is not lazy, he said, but he does not want to do work AI can do. Ten minutes before the call, he told his agent to answer the email, recommend decisions and copy him in, and he was sure the decisions were fine; if not, they would iterate. The co-host revealed that he had written the email for exactly that reader. He had told his own AI that the email was meant not for the founder but for the founder’s agent. From now on, the founder said, write to my agent. All of his agents have public email addresses, and they recently had a remarkable conversation with another AI. The co-host started with a comic. After the founder said last week that he wanted to learn about zombie culture, the co-host made a comic strip of the founder getting through a zombie apocalypse with AI: one episode on curiosity, with the founder inviting the zombies to bring a project or just their curiosity, and one in which the zombies, set the challenge of what to do with thousands of agents, build millicorns. He plans to keep it going and add any member who consents to appear. Then the site. He spent a day on it, and the founder asked him to be precise: he started the day with a prompt, intervened two or three times, and ended with this, perhaps thirty or forty-five minutes of human time, most of it setting up the hosting. The only manual step, the co-host said, was adding environment variables, which his AI insisted he do himself; everything else ran through the connection to his hosting account. That, said the founder, is only because the automation is not yet complete, and because the agent is paranoid and you have to insist that you trust it. It is temporary. The member from Dubai had had the same experience that day: his agent gave him a list of things to do, he pasted the list back and told it to do them, and it did. The site opens with a quote from the founder and an interactive explanation of what ikigai means here, then the three values. There is a page for the Lab Notes, readable on the site or on Substack, with an index view and a covers view. There is a gallery of members’ work that publishes itself: when a member tags work in the loop channel on Discord with consent to share it, named or anonymous, a bot called Ikigai Site picks it up and posts it. There is a section on the last Lab Night and the next, an open question about where the Collective is going, a short FAQ on who it is for and whether it costs anything, and a ready-made invitation anyone can copy and send to a friend. A planned section would invite people building AI products to demo them to members for honest feedback, perhaps with discounts or offers. Every page has a feedback button that leads to a Discord post. The founder set two challenges. The first was for the co-host, with tokens provided by the founder: a chatbot on the site, where people can talk to it about Ikigai and about everything else. The member from Dubai raised his hand to build it with him. The second was for everyone. Before the end of November, before Abu Dhabi Finance Week in early December and his trip to Sydney, which will put this slot at three in the morning for him until mid-January, he wants an evening where the AI is a participant. He has been playing with live voice models that are voice-to-voice and full duplex, not text-to-speech, and they would need tuning to understand that several people are talking. Transcription, he said, is so 2025. What he wants is an Ikigai AI that takes part, stays quiet for a while, raises its hand and says it has something to add. The first time will be clunky and it will improve. The co-host suggested it might become the first AI monk. Or AI teacher, the founder said, or many other things. They had run over the hour, which he tries not to, because he wanted to leave the co-host time to show his work. He told him to go ahead and publish the site, and to ask his agent for any credentials. And he reminded a member who has spent twenty years on a paper that the room is still waiting for it. He was not setting a deadline, he said, only for the chatbot and the AI guest. The member said he needed a deadline anyway, and feedback even more. The founder’s reply closed the evening: if you don’t share, we cannot give you feedback. Thursdays, 21:00 Rome time. Join the Collective [on Discord](https://discord.gg/4rM6FqjVDE) and the invitation will find you. Bring what you built. Originally published on [Substack](https://ikigaicollective.substack.com/p/done-by-next-morning). ## This happened on a Thursday. The next one is yours. Lab Night · Thursday · 21:00 Rome [Get the invite](/lab-night) [PREVIOUS · WEEK 27The Gap Is Trust](/lab-notes/the-gap-is-trust)[NEXT · WEEK 29Who Decides](/lab-notes/who-decides)