The person I keep meeting

There is a person I keep meeting, in one form or another. A beginner who wants to learn AI engineering. Bright, motivated, genuinely curious. And I watch them lose two weeks to the difference between two vector databases and the mechanics of asynchronous jobs — real topics, badly timed — while the thing those topics were supposed to serve never gets built. They do not fail from laziness or lack of talent. They fail because they were handed a roadmap, and a roadmap optimizes for coverage, and coverage is not capability.

If you only look at social media, you would not think this is common. The feed shows you a sliver: the people chasing the next dopamine hit of a saved “roadmap” they will never open, and the small handful who have already figured it out and are performing their mastery. Between those two extremes sits almost everyone in real life — students, experienced professionals, whole organizations — and their journey toward knowledge, skill, and capability is almost never at full momentum. It moves in fits and starts. People push themselves so hard that they never develop any love for the thing. Or it bores them. And so we reach for the easy verdict: the schools are bad, the courses are bad.

Maybe. But that is not the real question. The real question is: how do we help a person — or a team, or a company — learn and build capability in a way that keeps their agency intact, instead of grinding it down?

This is a named problem, not a private complaint

I want to be clear that this is not just my frustration talking. The Anthropic Institute frames a version of it directly: interfaces shape what people become. A television turns you into a passive viewer; a computer can turn you into a generative creator. Their open question is what interfaces can be built so that AI systems improve and promote human agency rather than quietly removing it. That is exactly the question underneath everything here.

And the evidence says the bottleneck is not capability. We now have extraordinarily capable models available to nearly everyone, and yet the numbers on how people actually cut through and learn have barely moved. The gap is not intelligence. It is that the interface and the mindset are not aligned with each other. We keep bolting more power onto the same passive shape.

The dominant “fix” — personalization — often makes this worse, not better. I have called it, elsewhere, personalization as an autoimmune disease: a system builds a fixed portrait of you, then filters all of reality through that portrait, defending the model of you against the actual, changing you. It does not evolve with your intent or your method. It freezes a snapshot and calls it service. The result is scatter, not clarity. So if we want an interface that raises agency, it cannot be built on a frozen portrait. It has to be built as something that learns your method and keeps moving with you.

Two ways of seeing

Before the reframe, the fork has to be named, because everything turns on it — and it is not really a fact about systems, it is a mindset. There are two ways to look at almost anything. The reductionist way breaks a thing into parts, understands each part, and assumes the whole is their sum — the world as a machine you can lay out on the bench. The holistic way looks at the whole first: the relationships, the feedback, the way the thing evolves and organizes itself, the properties that exist only because the parts are together. Same object, two lenses — and the lens decides what you are even able to see.

This is not a new quarrel. Fritjof Capra made it the center of The Turning Point, arguing that the crises of our age come from applying the old Cartesian-Newtonian, mechanistic, reductionist worldview to a reality that is actually a web of living, dynamic, self-organizing relationships — and that the deepest shift already underway in modern physics is precisely the move from the machine to the system. Philip Ball, in How Life Works, shows the same correction landing inside biology: the comfortable story — the genome as a blueprint, genes as instructions, the cell as a machine — turns out to be misleading. Life is a system: multi-tiered, modular, evolving, driven by interactions inside and outside the cell, and shot through with agency, where the cell is the orchestrator and the genes are just one of its tools. Not a blueprint read out from the top, but flexible rules and resources from which form emerges. That systemic reading of life — evolving, modular, self-organizing — is the same lens everything below is written through.

The holistic view does not throw reductionism away; you still have to burrow into the parts to know how a protein folds. It re-situates the parts inside the living whole. And it marks a line the reductionist eye keeps crossing: a complicated thing — an engine, a tax code — can be decomposed, and its behavior predicted from its design; a complex thing — a market, a city, a person learning something hard — has feedback, emergence, and surprise, and cannot. The reductionist mistake is to treat a complex thing as if it were merely complicated.

Which is exactly what a curriculum-as-checklist does. Fifteen weeks, one topic per week, on the assumption that the whole is the sum of the weeks and that understanding will assemble itself at the end. It never does — and it manufactures a second harm, the guilt of the unfinished plan, when you fall behind a schedule you invented and mistake a modeling error for a personal failing. Adopt the holistic lens instead, and learning becomes what it always was: a living, evolving, self-organizing system. That is the reframe.

The reframe: a curriculum that is alive

So treat the curriculum as a system — the kind biology actually runs on. Not a line to walk, but a landscape that grows, reorganizes, and stabilizes as you move through it, with you inside it as a participant rather than a recipient. A few properties come with that, and each one changes how learning is supposed to feel.

Self-organization. Order is not imposed from a plan; it emerges from interaction. Clarity does not arrive because a syllabus scheduled it for week nine — it arrives because you kept pressing on live questions until an understanding fell into place on its own, often after friction. Chaos and conflict are not failures of the system. In systems held far from equilibrium, they are exactly where new order comes from.

Canalization. This is the stabilizer, and without it self-organization is just drift. Repeated passage carves grooves; certain understandings become deep valleys that later learning flows into and returns to. Exploration gives you divergence, canalization gives you convergence, and a living curriculum needs both or it is either rigid or formless. Watch it in your own notes: when the same few ideas keep reappearing across unrelated entries, that recurrence is a canal forming.

Modularity and redundancy. Understanding accretes in reusable units, and nothing starts from scratch — new learning attaches to what is already carved. Redundancy, which a checklist treats as waste, is mostly a feature: meeting an idea through several different routes is what makes it robust rather than brittle. The reflex to eliminate all overlap is the reductionist reflex, and it is wrong.

Multilevel and multidirectional. Understanding moves up and down levels, not just forward. A concept met at the surface sends roots into the mathematics beneath it; a piece of deep theory suddenly explains a surface behavior you noticed months ago.

The engine that drives all of this is a loop I have leaned on for years and written about as Learn–Experiment–Interact. Interaction is what makes you learn something new; experiment is what tests and validates it and shows you the gaps; and the learning reorganizes itself through the doing. It is not a rigid sequence — sometimes you interact first, sometimes you learn first — but the three feed each other, and understanding self-organizes out of the loop rather than being poured in.

The learner is inside the system

The most important move is to stop imagining the learner as someone a curriculum is delivered to. In a living system the learner is a component, and learning is what emerges from the interaction between the learner and the material — not a substance transferred into a passive vessel.

This is where agency comes from, and where the interface question becomes real. A roadmap makes you obedient to a sequence. A living system makes you a participant whose interactions actually shape what the system becomes. That is the difference the Anthropic Institute is pointing at: an interface that promotes agency has to put the person inside the loop as an actor, not outside it as a consumer of recommendations. Automation done wrong takes the loop away from you. The goal is the opposite — to hand you a sharper loop.

Just-in-time theory is the metabolism

You do not front-load all the theory and then hunt for problems to use it on. You hold a live question, hit the exact wall where you lack a tool, and go get that tool — with real motivation behind it, which is the only reason it sticks.

Metabolism is the right word, and for me it is not a decoration. I manage a chronic condition that means I cannot take a heavy, fatty meal all at once; so when I eat one, I bring an enzyme to that meal, and it eases the digestion precisely when the load arrives — not stockpiled uselessly in advance. Learning theory works the same way. You do not pre-digest all of mathematics. You bring the specific lemma the proof in front of you actually needs, exactly when the difficulty demands it.

And this is not a quirk of biology or of learning. It is how well-built adaptive systems behave in general. A ride-hailing price is not read off a fixed table; it is computed on demand from the live state of supply and demand, and its depth scales with how volatile the moment is. Good software evaluates lazily — it does not compute a value until something asks for it. The immune system does not keep a finished defense for every pathogen in stock; it mounts a specific response on demand, then keeps the memory. On-demand depth, summoned by the situation rather than scheduled in advance, is the signature of systems that stay alive in changing environments. A curriculum should have that signature too.

Just-in-time learning has exactly one failure mode, and the structure fixes it: you cannot reach for a tool you do not know exists. That is what the map of levels is for — it tells you which layer you are chronically under-resourced at, so you can sense your own blind banks and go carve them. Emergence with no scaffold is drift. Emergence with a scaffold is self-organization. The scaffold is the banks; the water still finds its own level.

The same shape — and it is not only about individuals

None of this is a new belief arriving from outside. It is the same shape recurring across everything I work on, at three different scales.

At enterprise scale it is Evolving Decision Systems: the recognition that the right answer today is quietly the wrong one tomorrow, that AI freezes the moment it ships while the world keeps moving, so the systems that matter are the ones that never finish evolving. At individual scale it is coocBoard — the interface I am actually building and putting in people’s hands, a proof of concept meant to test one hypothesis: that helping a person cut the noise, find the one thing worth their attention, and build a real arc (rather than collect isolated hits) is what actually moves them further. And at the scale of a single mind it is this: the living curriculum, ten tracks held not as a checklist to complete but as channels a single evolving understanding flows between.

The deepest version of the idea is the one I worked out for agents as complex adaptive systems. An agent harness, done right, is not a wrapper that constrains — it is a developmental substrate that enables. You do not encode the solution; you provide the Lego blocks — memory, tools, feedback, a place to keep what works — and let capability self-organize around the problem, under selection pressures you design with care. But self-organization is not magic: an elephant does not grow wings, because the substrate and the pressures do not favor them. If you do not provide the blocks, the capability cannot emerge.

Now look at the parallel. A curriculum is a developmental substrate for a person. A harness is a developmental substrate for an agent. The Fabric is a developmental substrate for an enterprise. Same architecture, three altitudes: sense the world honestly, keep a steerable and evolving model rather than a frozen portrait, compose what is needed just in time, and never finish. The reductionist checklist fails at all three scales for the same reason — it treats a complex, evolving thing as a complicated, finished one.

The résumé trap, and its research twin

You can watch this same mistake play out in the most consequential place of all: how people try to build a career.

The standard move for breaking into AI engineering is to build projects and dress up a résumé. But look at how it is usually done — you copy-paste a job description, pick a project that will look like a match, and over a Friday-to-Sunday you prompt a coding agent until something runs. It gets you shortlisted. It also leaves almost no mark. The projects all converge on the same handful of component-level demos, because they were reverse-engineered from job descriptions rather than from any real problem — and real problems, especially inside enterprises, live in specific, messy contexts that no generic demo ever touches. Taste does not develop, because taste comes from wrestling a problem you actually care about, not from assembling completeness for a checklist someone else wrote.

The living-system version is different. A project should emerge — as an idea or a hypothesis that surfaced while you were reading a book or a paper and thought, “I want to test whether this is true.” Then the project is an experiment, and an experiment leaves a canal: it changes what you understand, sharpens your taste, and — almost incidentally — makes a far better story than anything built to match a JD. Same three days of effort, completely different residue.

And this is not only a beginner’s problem. The research world has its own version of the fork, and Arvind Narayanan names it well: deadline-driven work, organized around papers and acceptances and an external clock, versus value-driven work, organized around a long arc and the honest question of whether it made anything better — a switch he says took him twenty years to fully make. I read it as the same reductionist–holistic split at the research altitude: the confined, deadline-focused, roadmap-centric, artifact-optimizing mode on one side, and the living, self-directed, value-first mode on the other. What a friend of mine bluntly called “Claude-maxing” papers for a workshop is just the first mode with the volume turned up.

Both cases collapse to one thing. A résumé project and a workshop paper can each be built two ways — as completeness-theater assembled to satisfy someone else’s checklist, or as an experiment that emerged from a question you genuinely hold. Only the second one leaves a mark, and only the second one compounds.

Utility over theory — which is why it has to be built

I have to be honest about the trap on my own side of this, because I wrote about it on the last night of 2025. Theory does not belong in the storefront. People, teams, ecosystems — they do not receive sound theory; they receive utility. Water, food, something that quenches the actual thirst. My own failure mode has been holding a rigid, half-baked system in my head and refusing to ship until it is perfect by that private standard, which means it never becomes anything anyone can use. The correction I committed to was simple: even if it looks rough, start building, let the world use it, and learn from what comes back. Consistency, not perfection, is the thing that compounds.

That is precisely what coocBoard is for. The living-curriculum idea, left as an essay, is just theory in the storefront. Compiled into an instrument a stranger can pick up and run, it becomes utility — an interface that raises agency instead of a manifesto about raising agency. The philosophy is the map; the instrument is the water.

And I keep coming back to an image from that same night. Life is like a river — it starts fresh, gets contaminated, faces disturbances, and then, as it keeps flowing, it settles and runs clear. Learning is the same. You do not clarify it by scheduling clarity in advance. You let it flow, let it carve its banks, and trust that the channel forms. A living curriculum is never finished, and that is the point. It is not a thing you complete. It is a thing you keep alive.

A note on this document’s own making: none of this was clear a few days ago. The idea was somewhere in mind, but unformed; it took the interaction to draw it out, and each pass pulled it into sharper shape — learn, experiment, interact, repeat. This memo is itself an instance of the thing it argues for: understanding self-organizing through interaction, canalizing over successive passes into a channel it can finally flow through. And it is not only this document. Read through the musings and you will not find a fixed position being defended; you will find a person becoming one — the thought reorganizing itself in public, which is the whole claim, demonstrated rather than asserted.