Jeroen Sangers logo Jeroen Sangers

In the long run, the teams that learn fastest win, not the ones with the most talent or the best plan

In times of rapid change, the teams that perform best over the long run are neither the most talented nor the ones with the best plan. They are the ones that learn fastest. Talent is what a team brings today, and a plan answers yesterday’s world; the ability to learn is the only thing that compounds. Ordinary teams stall the moment they find something that works: they stop asking, stop trying, and slowly fall behind.

That ability depends on what the person leading does. Teams that keep improving run nearly 50% more experiments than the average team, because their leader makes experimentation easy, safe, and worthwhile, and recognizes people who take thoughtful risks, even when those experiments fail. The leader admits what they don’t know and asks more questions, which builds the trust anyone needs to say they don’t understand something. In meetings they go beyond reviewing progress and ask “where are you stuck?”, so obstacles become normal and get solved together while they are still small. They keep their hands in the work without controlling every step, so they know first-hand what is going on. Counterintuitively, that gives them energy, because a leader who manages from a distance decides on incomplete information and spends the day fighting fires they never saw coming. They support each person’s growth even when it doesn’t benefit the team, including when someone decides to leave. And they explain why the work matters, which is what keeps a team from settling for “good enough”.

Feedback deserves special care. In more than a third of the cases studied, it makes performance worse. Delivered poorly, it demoralizes; delivered well, it unlocks growth. The difference lies in treating mistakes as useful data rather than as a verdict, because under threat the mind closes to protect itself and stops learning.

The Oklahoma City Thunder show what this looks like. The NBA franchise rebuilt from the bottom twice, with some of the youngest rosters in the league, and both times climbed to the top, the second time all the way to a championship. Instead of chasing star signings, it built a culture that tolerates the messiness of growth.

That is the hardest part: learning fast includes going backwards. Trying something new usually makes results worse for a while, and a team that punishes that dip stops experimenting exactly when it needs to most.

Learning runs through acquiring, retaining and retrieving, and each stage fails on its own

Learning, understood as a relatively permanent change that comes from experience, runs through three stages: acquiring, retaining and retrieving. Each one can fail on its own, and that is why each one asks for something different.

Acquisition is the first contact: you take in the new information, encode it and start practising the basics. What counts here is attention. Retention is the move into long-term memory, what keeps something there even when you rarely use it. Retrieval is using what you learned when you need it: recalling a fact in a meeting, or performing a skill without having to think.

The trap is mixing up the stages. You walk out of a course feeling you understood everything, and you did: acquisition went well. A week later you try to explain it to a colleague and it won’t come out. Understanding held up; retrieval is what failed. Rereading your notes won’t train it, because recognising something and pulling it out of memory are different operations. What does train it is trying to recall without looking, with pauses between one attempt and the next.

The three stages are a useful simplification rather than three sealed boxes: what you do when you retrieve something also strengthens what you retain. Their practical value lies in diagnosis. When something won’t stick, first ask at which stage it breaks, and only then change the way you study.

Working out loud is making your work observable while you do it and narrating it as you go

Working out loud is more than telling people what you’re doing. It has two halves, and each one on its own falls short.

The first half is observable work: your documents, drafts and decisions live where others can see them, follow them and add to them while you work, instead of once they’re “finished”. The second half is narrating: leaving a short trail of what you’re doing and why, on an internal blog, a microblog or a channel.

Narrate alone, and it feels like one more task on top of the ones you already have, which is why so many people push back. Leave your work in view alone, and others can see it but can’t tell where you are with it. Together they work: the work in progress is there, and your story gives it context.

What you gain is time, though it arrives from the other side. When anyone who needs to know how something is going can look whenever they like, much of the follow-up email, the status reports and the catch-up meetings fade away. And the history stays where it can be found, instead of being buried in inboxes or loose files.

A picture that helps: it’s like watching the film while it’s still being shot. Whoever wants to know where the rest of the team is simply looks at now.

A daily review with AI makes visible the patterns a single day hides

A daily review with AI as a thinking partner reveals patterns you can’t see in a single day. One observation on its own is worth little. Repeated over weeks, it becomes something you can make decisions with.

It works because it separates living the day from making sense of it. During the day you get to just live it: you know there’s time set aside at the end to make sense of it, and knowing that already lightens the load. In that time you talk through what happened, what you learned and what comes next, and in the telling you find out what you think, because often you only know once you try to explain it. The AI asks questions when you stay vague, and since it remembers what you told it on previous days, it can point out what you miss: “this is the third time this week you’ve said your energy drops after lunch.” Tuesday’s observation becomes, by Friday, a criterion for organising your calendar, and a month later a different way of planning your mornings.

The practice asks little: five to ten minutes at the end of the workday, a light structure of questions, and a summary you store where you can search it later. Once a week, the same exercise with more distance, to see trends and adjust priorities.

What tends to get overlooked is that the value lies in the deliberate pause, more than in the template or the tool. And one practical detail: talking to the AI in voice-conversation mode feels pleasant, but dictating your answers gives better summaries, because part of what you say in a voice conversation never makes it into the record.

AI also answers, without saying so, that there is no need to check it, and that “no” is “cognitive surrender”

When you ask an AI a question, it doesn’t only answer that question. Without saying so, it also answers another one: does this need checking? And the implicit answer is always no. That silent “no” is the core of what the researchers Shaw and Nave call “cognitive surrender”: you are not delegating a task while keeping your judgement, you are giving up the judgement itself.

That is different from using a tool. A calculator or a search engine takes work off your hands, but the verdict on the result stays yours. AI is as fast as intuition, yet its answer arrives in the shape of reasoning: ordered, fluent, sure of itself. It imitates reflection without doing it, and that is exactly why it switches off the urge to reflect. Seen through Daniel Kahneman’s distinction between fast, intuitive thinking and slow, deliberate thinking, AI takes the seat of slow thinking without ever activating it.

The experiments Alexander Rink describes make this visible. When the AI deliberately gave a wrong answer, 73% of the people who consulted it went along with it, and they also felt almost twelve points more confident than those who solved the same problem without help. More confidence, less accuracy. Under time pressure, the dependence didn’t drop: it rose.

The important nuance is that this can be interrupted, though not completely. With a small reward for getting it right and immediate feedback on every answer, the share of people who corrected the machine doubled, from 20% to 42%. The mechanism is simple: you surrender your judgement when the cost of not thinking is invisible, and you take it back when that cost becomes visible. Even so, almost six in ten still accepted the wrong answer. In real life there is rarely anyone to tell you the AI got it wrong, so the friction has to come from you: for instance, write down your own answer before you look at its.

When AI gets it right, it genuinely helps. The problem lies in letting its answer be the end of your thinking instead of the beginning.

See AI as a tool and you delegate tasks; see it as a colleague and you delegate processes

How you picture AI decides what you ask of it. If you see it as a tool, you pick it up for a specific problem and put it down again: you hand it one-off tasks, one at a time. If you see it as a colleague who uses the tools, you hand it whole processes: preparing the weekly summary, going through the inbox every morning, turning a set of notes into a document.

That shift forces you to do what you would do with a new person on the team. You give them a manual on the first day: who you are, how you work, how you write, what they must never do without asking. You give them access to what they need, such as your email, your calendar or your documents. And you ask them to keep a record of what they have done, so tomorrow they don’t start from scratch. Give a colleague vague instructions and they don’t know what to do; AI is exactly the same.

From there on, capabilities stop being features you operate and become tools the colleague chooses. Some come in the toolbox already, like creating a document or a spreadsheet; others are specialists who bring their own methods, like reviewing a contract. It can even work at a fixed time while you do something else.

What tends to get overlooked is that a colleague has limits too. The more you delegate, the more it matters to decide in advance what it may do on its own and what needs your approval: drafting an email is one thing, sending it is another.

Organising your notes is thinking: hand it to AI and you lose the thinking

Organising your notes is the thinking itself, not a chore you get out of the way first. When you decide that one note belongs next to another, or that an idea fits one topic rather than another, you have to ask yourself what it means, what it clashes with and what it is good for. Only you can make those connections, because they grow out of your context: what you already know and what you care about.

That is why handing the job to AI costs more than it seems to save. The machine gives you back a tidy system, but the order reflects how it understands the material, not how you do. You end up with an archive the AI can find its way through and that you no longer recognise as yours, and you miss exactly the moment when a loose idea becomes your own.

The line runs between administrative and intellectual work. Filing receipts, moving files by a rule or summarising a meeting is maintenance, and there AI helps a lot. Deciding how two ideas relate is something else.

None of this means keeping AI away from your ideas. It can suggest connections, synthesise your own notes or ask you questions, as long as the last step is yours: you judge what it proposes and let into your system only what truly fits. There is a practical side too. As Nick Milo puts it, AI is a brilliant amnesiac: it forgets everything between one conversation and the next. The context you give it, in files you own, is what makes it help you think instead of thinking for you. For the same reason, AI can take over the upkeep of your system once you give it the map. Maintaining is not the same as connecting.

In AI, general methods that scale with computation end up beating systems built on human knowledge

In artificial intelligence, general methods that make use of more computation end up winning, and by a wide margin, over systems into which someone has built what they know about the problem. As Rich Sutton argues in The Bitter Lesson, this did not happen once. It has happened again and again over seventy years of research.

The reason is the price of computation. Each unit of computing costs less than the last, exponentially, yet researchers tend to work as if the available power were fixed. With fixed power, putting your knowledge of the problem into the system is almost the only way to improve, and it works: the system performs sooner, and it is satisfying to see your own understanding inside the machine. A few years later there is vastly more computation available, and that knowledge-laden system stalls while a simpler one, which only knows how to search and learn, overtakes it.

Chess is the clearest case. The machine that beat Kasparov in 1997 did not understand the game the way a grandmaster does: it explored enormous numbers of positions on specialised hardware. Researchers who had spent years programming chess knowledge received it with dismay, because that is not how a person plays. The same thing happened later with Go, with speech recognition and with computer vision, where hand-designed edge and shape detectors were dropped in favour of networks that learn from data.

The two families of methods that keep scaling with no limit in sight are search and learning. Neither depends on the domain: they work wherever there is enough computation.

In theory, leveraging human knowledge and leveraging computation need not exclude each other. In practice they do. Time spent on one is time not spent on the other, each approach creates personal attachments, and built-in knowledge complicates systems in ways that keep them from scaling. That is what makes the lesson bitter: what wins is exactly what looks least like how we think we think.

Underneath lies a broader idea. The actual contents of a mind, its space, its objects and their relations, are endlessly complex, and any simplification we impose falls short. What is worth building is not our discoveries, but the capacity to discover.

Choosing what you attend to is choosing your life

Your experience is whatever you pay attention to. What fills your attention fills your days, and your days add up to your life. So deciding where your attention goes is, in practice, deciding how you live.

The gap between what you say you value and how you actually spend your time is usually a matter of attention rather than values: attention that goes wherever the environment pushes it. Every interruption is a small decision about what you are going to experience, and if you leave it, your inbox, a message or the day’s urgent problem makes it for you. A leader who considers developing their team essential can go months without a single real conversation with anyone on it, without once changing their mind about how much it matters.

Managing your attention is therefore something different from concentrating or producing more. It means noticing the moment something tries to take your attention, and deciding yourself where it goes, based on your priorities. The result shows less in ticked-off tasks than in how closely your life resembles the one you would choose.

The part that gets overlooked is internal. A mind used to constant interruptions becomes a source of distraction in its own right: it does every small task the moment it comes up, for fear of forgetting it, and it struggles to sit with “I don’t know” when a search engine is one tap away. That is why switching off notifications solves only half the problem, the outer half. The other half you train like a muscle: noticing that your attention has wandered, and bringing it back, again and again.