The living system, part two: when you start serving the system

The system generates possibilities. The human decides what matters. That sounds simple, and it is the hardest part of the arrangement.

Part one described what changed once my Editorial Intelligence repository held enough accumulated context to give me back connections I hadn’t put into it — an old field note explaining a new observation, several unrelated pieces turning out to share an argument. The short version is that a store of material stopped behaving like an archive and started behaving like an environment for thought, and that the surprise this produces comes from recombination rather than from anything resembling intelligence.

This part is about what that costs, which took me considerably longer to notice than the effect itself.

When the system starts creating work for you

This worries me considerably more than whether AI is somehow thinking, because a sufficiently generative knowledge system starts creating its own workload.

Every idea creates three other ideas.

Every framework reveals another framework.

Every experiment generates a backlog.

Every connection looks like something that ought to be documented.

You end up with a beautifully organised machine producing an endless list of things for you to do, and at that point you are no longer using the system. You are serving it.

There is a familiar productivity trap underneath that. We assume organising work gives us control over it, when sometimes organisation just makes the infinite amount of possible work more visible. AI accelerates the problem, because its great strength is that the marginal cost of producing another possibility is close to zero while human attention stays expensive. So a system like this needs something beyond memory and generation. It needs editorial judgement.

The human remains editor-in-chief

This has become one of the principles I want Editorial Intelligence to keep:

The system generates possibilities. The human decides what matters.

That sounds simple and isn’t. It means resisting the assumption that everything interesting deserves development, being willing to leave good ideas in the repository unused, distinguishing between something that strengthens the core argument and something that merely extends it, and deciding when there is already enough.

It also means recognising that the final scarcity in an AI-rich environment is unlikely to be content. It will be attention, judgement, priority and taste — the ability to say that this matters more than that, which is an editorial function rather than a technical one.

This is bigger than writing faster

Most discussion of AI and knowledge work still focuses on production: can it write the article, summarise the meeting, turn the report into ten social posts, reduce the time something takes to make. Those are useful applications and I suspect they are the least interesting part of this.

The larger change comes when AI operates against persistent organisational memory — when customer interviews, research, sales conversations, support tickets, strategy documents, market observations and previous content stop sitting in separate systems and become an interpretable evidence layer. At that point the AI isn’t only being asked to produce. It can help ask:

What keeps recurring?
What has changed?
Where is the evidence accumulating?
Which assumptions are weakening?
Which customer language is spreading?
What have we already argued?
Where are we contradicting ourselves?
Which ideas have become more important than they appeared six months ago?

That looks less like content automation and more like organisational sense-making, which is much closer to what Editorial Intelligence has turned out to be about.

What this doesn’t prove

I should be careful about how far I push this, because the evidence behind it is one repository belonging to one person.

I don’t know how much of this depends on conditions that are specific to me: that I built the system, that I remember most of what is in it, that I am the only person adding to it, and that I already know roughly where the interesting material sits. A system I designed will inevitably feel legible to me in ways it might not to somebody else.

I also can’t separate the effect of the system from the effect of simply having worked on the same subject for a long time. Some of the connections I’m describing might have arrived anyway.

The honest version of the claim is narrow: with enough accumulated context, retrieval starts producing connections I did not consciously make, and that changes how I work. Whether it holds inside an organisation — where knowledge is contributed by many people, most of whom did not design the system and cannot remember its contents — is a different question, and one I would want practitioner evidence to answer rather than my own experience.

That distinction matters more than it might appear. A knowledge system that works because its owner already knows what is in it has not solved organisational memory. It has just made one person faster.

A living system without pretending it is alive

I’m comfortable calling Editorial Intelligence a living system, but only metaphorically. Living systems evolve, absorb new information and change in response to what happens to them; connections emerge that were never designed, old parts become less important, new parts become central, and the architecture shifts as the environment does. That is increasingly how the project behaves.

What I don’t want is to confuse that dynamism with autonomy. AI doesn’t replace the person responsible for deciding what the system is for, and the more capable the system becomes the more that responsibility matters — because once knowledge becomes generative, the challenge stops being how to get the machine to produce something and becomes how to keep control of what deserves to survive.

The published layer is inside the system

There is a boundary I assumed existed and doesn’t.

Most knowledge systems are live on the inside and frozen at the edge. Material accumulates, connects and changes internally, and then something gets extracted, published, and stops participating. The published version becomes an artefact the system produced rather than a part of the system.

That is not what has happened here. Field Notes runs on the same principle as everything sitting behind it. A published piece is still a node — it gets extended when the argument develops, absorbed into another piece when it turns out to be a restatement, occasionally contradicted by something written later. Several essays on this site have changed materially since the first people read them.

Which sharpens the point about judgement rather than softening it. If publication is the moment a thing is judged to have survived, that decision gets made once and then defended. If the published layer stays live, what deserves to survive is a question the system keeps asking about work it has already shipped. That is a harder discipline to hold, not an easier one.

It also carries a cost the private layer doesn’t. An internal store can be messy, contradictory and half-finished, because only one person reads it. Revising in public means revising in front of people, and the temptation there is to smooth rather than to extend — which would produce a body of work that looks increasingly settled while quietly becoming less honest about how much is still unresolved.

The real opportunity

I started building Editorial Intelligence as a way of preserving and developing my own thinking. I’m increasingly convinced the more important idea is what happens once enough thinking has been preserved.

At first a knowledge base helps you remember.

Then it helps you retrieve.

Then it helps you connect.

Eventually it starts helping you discover.

That progression changes what a repository is for. It stops being where finished thinking goes to live and becomes somewhere thinking continues — which may be one of the more consequential ways AI changes knowledge work. Not because machines have begun thinking for us, but because it is now possible to build environments in which human judgement, persistent memory and machine inference interact continuously.

Such a system can surface possibilities you would otherwise have missed, hold a thought long enough for its relevance to become visible, and connect parts of your own experience that working memory cannot hold at the same time. Occasionally it can surprise you. That surprise isn’t evidence the machine has come alive. It’s evidence that the relationship between people, knowledge and tools is becoming a great deal more interesting.