The living system, part one: when a knowledge system starts to feel alive

Most knowledge systems are archives. Something changes once AI can work across enough accumulated context — and it isn’t that the machine starts thinking.

I’ve reached a slightly unsettling stage with Editorial Intelligence: the GitHub repository has started to feel alive. Not literally — it isn’t conscious, it doesn’t have intentions, and no AI has secretly become the author of the project. But the experience of working with it has changed, and the change seems worth describing carefully, because I think it points at something more general than one person’s repository.

At the beginning it was mainly somewhere to put things: frameworks, field notes, research, article ideas, observations from work, experiments, half-formed thoughts that might become useful later. GitHub gave those persistence and AI made them easier to retrieve, but I was still doing almost all of the connecting myself.

Something different has started happening. There is now enough accumulated context that the system gives me back connections I didn’t explicitly put into it.

One observation connects to something I wrote weeks earlier.

That exposes a gap in a framework.

The gap suggests an experiment.

The experiment creates another observation.

That observation becomes an article.

The article changes how I understand the original framework.

And the new understanding goes back into the repository.

The system is no longer just storing my thinking. It has started to participate in the conditions that produce new thinking, and that distinction is the whole subject of this piece.

From archive to thinking environment

Most organisational knowledge systems are treated as archives. Documents go in, meeting notes go in, research and presentations go in, and someone might search for them later. That is useful and fundamentally passive: the assumption is that the intelligence stays with the person doing the searching, while the repository just holds the material.

Once an AI system can work across enough persistent context, the archive becomes more active. It can compare an observation with earlier work, identify repetition, notice contradictions, and suggest that two apparently separate ideas are versions of the same argument. It can show that something treated as a one-off is becoming a pattern, point out where an idea is still unsupported, and retrieve something you had forgotten at the moment it becomes relevant again.

The underlying material is still human and the interpretation still requires judgement. What changes is the relationship with stored knowledge: the archive starts behaving less like a cupboard and more like an environment for thought.

The difference is accumulated context

There is a large difference between asking AI a question in an empty chat window and asking it to work inside a body of knowledge that has been accumulating for months.

A blank chatbot knows nothing about why an idea matters to you, what you tried before, which arguments you rejected, which concepts keep returning, or the difference between something central to your thinking and something you mentioned once. So you keep recreating the context, which is one reason AI can feel impressive and strangely shallow at the same time. It can generate. It doesn’t know where the generation belongs.

Persistent context changes that, because the system develops a history — not memory in the human sense but something more practical, a retrievable record of previous decisions, observations, evidence and ideas. That history lets AI do more than produce a plausible answer to the latest prompt. It can work against what came before, which is where the interesting behaviour starts.

Thinking outside the head

I used to think of knowledge work as something that happened primarily inside the individual. You read, notice, remember, connect, decide and write; the notebook or the content management system comes afterwards and records the result.

I’m increasingly unsure that’s the right model for AI-enabled work, because some of the thinking can now happen across a system rather than inside a head.

I notice something. The repository preserves it. AI retrieves related material. I judge whether the connection is meaningful. That judgement changes the knowledge base. The next interaction begins from a richer position.

The loop might look something like this:

observation → interpretation → memory → connection → judgement → new observation

And then back around again.

None of those components is remarkable alone, and the interesting behaviour appears in the interaction between them. I bring lived experience, curiosity, taste and judgement; the repository provides persistence; AI provides retrieval, recombination and a very cheap ability to test whether a possible connection holds. The result is a kind of extended editorial cognition, in which thinking no longer has to happen entirely inside my head.

Why it starts to feel alive

This is where it becomes psychologically strange, because a system starts feeling alive at the point where it can surprise you — where you put something in and get back something meaningful you didn’t consciously intend. An old field note suddenly explains a new observation. Several unrelated articles turn out to share an underlying argument. Something you’d filed as an interesting anecdote turns out to support a framework.

Nothing supernatural has happened; the surprise is produced by accumulated context and recombination. But from inside the experience that distinction blurs, because you know you created the ingredients and you also didn’t consciously create the connection. That is different from note-taking, and different from ordinary generative AI. The system has become generative because its history has.

The danger of confusing emergence with intelligence

There is an obvious temptation to anthropomorphise this — to say the repository is thinking, that the AI understands the project, that the system knows where it wants to go. I think that would be a mistake. The system has no point of view independent of the material and instructions it is given. It does not care whether Editorial Intelligence succeeds, cannot decide that an idea matters, and cannot distinguish an intellectually interesting possibility from a productive direction without a human standard to judge against.

That matters more, not less, as the system improves, because generative systems are extremely good at creating more. More connections. More frameworks. More article ideas. More experiments. More things that feel worth exploring.

Abundance of that kind can look like intelligence. Sometimes it is useful intelligence. Sometimes it is just possibility multiplying, and possibility is not the same thing as judgement.

That is where this part stops, because what follows is a different problem. A system that generates possibilities faster than you can evaluate them doesn’t simply make you more productive. It starts producing work for you, and deciding what to ignore turns out to be the harder discipline.

Part two picks that up: what a generative system asks of you, why the scarce resource stops being content, and what happens when the published layer stays live too.