Steven Johnson has spent his career on what he calls tools for thought, and one of the more interesting ones to arrive lately is the research assistant he helped Google Labs build, NotebookLM. Google just dropped the “LM” and folded it into the wider brand as Gemini Notebook. Johnson used the occasion to write, in his Adjacent Possible newsletter, about why the tool works at all, and he reached for a piece of philosophy to explain it: the extended mind.
The idea comes from a short 1998 paper by the philosophers Andy Clark and David Chalmers, “The Extended Mind,” published in the journal Analysis. Their argument runs through a character named Otto. Otto has Alzheimer’s, so he keeps a notebook of everything he needs to remember, and he carries it everywhere. When he wants to get to the museum, he looks up the address he once wrote down, the same way you would pull it from memory. Clark and Chalmers make the radical claim that the notebook is not a stand-in for Otto’s memory. It is his memory, a genuine part of his mind that happens to run on paper instead of neurons. Johnson’s point is that a notebook you can talk to, one that reads your sources alongside you, is the most literal version of Otto’s notebook anyone has built. I think he’s right, and I think the tool is genuinely good. I also think the argument has a second half he doesn’t reach, the half that applies to teams and companies rather than to individuals.
The extended minds that last are personal #
Look back at the whole lineage and one feature holds all the way through. The Renaissance commonplace book, a personal volume where a reader copied out the passages worth keeping. The diary. The box of index cards a scholar spent decades filling. Otto’s notebook. Now NotebookLM. Every one of them is a single person’s, curated by hand, and useless to anyone who wasn’t the one keeping it.
That isn’t a flaw. It’s the reason they work. Otto’s notebook is reliable because Otto tends it, corrects it, and trusts it the way he trusts his own recall. NotebookLM is powerful for the same reason: you choose the sources, you decide what goes in, and the model reasons over a corpus you assembled on purpose. The curation is the value. An extended mind you didn’t shape wouldn’t be an extension of your mind at all.
So the personal extended mind is a solved problem, and Johnson’s tool is about as good a version of it as exists. The trouble starts when you try to make one for more than one person.
No one keeps the company notebook #
Every team accumulates knowledge about how it actually runs. What we told that customer last week, why the launch date moved, which of us actually knows how the billing edge case works. The instinct is to do for the team what Otto does for himself: keep a notebook. Stand up a wiki. Start a shared drive. Point everyone at the same NotebookLM.
It never holds. The wiki is accurate for a month and stale for a year, until someone gets a bee in their bonnet and fixes it, and then it starts rotting again. The shared drive becomes a place files go to be lost. Not because people are lazy, but because the model that makes a personal extended mind work is exactly the thing that can’t scale. Otto’s notebook stays current because one person owns it. A company notebook’s upkeep belongs to everyone, which means it belongs to no one. That’s the tragedy of the commons, and a mind nobody is keeping isn’t an extended mind. It’s an archive, and a rotting one.
I got a firsthand education in what an organization does instead. My first month at a large tech company was spent in the engineering bootcamp everyone went through, working real tasks against real systems. Most of those tasks required learning something new or digging up some details and sample code, and finding a single source of truth was hard enough that I drew a conclusion about the place: it had an oral culture, not a written one. There was no shortage of text. But you weren’t expected to discover things on your own. You were expected to know who to ask, and half of onboarding was building your network.
That is what a company does when the written record can’t be trusted. It doesn’t stop functioning. It routes around the gap by keeping the knowledge in people’s heads and moving it by conversation, which works fine as long as the experts are still there and have time for you. The cost never shows up in a budget. Every answer spends two people’s attention instead of one person’s search. Every departure takes part of the record out the door.
You can throw people at the gap. Big companies do. They pay someone to keep the CRM clean and someone to write the status update that spreads what changed. What that buys is a team fighting entropy by hand, staying a step ahead of fragmentation. It never becomes a system that keeps itself. Hand-curation was always the mechanism, and hand-curation is precisely what doesn’t survive contact with a group.
A collective mind has to assemble itself #
A knowledge base that keeps itself current has been a dream for as long as we’ve had wikis. It never arrived because the technology wasn’t advanced enough. What’s changed is that AI can finally do the tending. And the material it needs is already there, produced just by the work itself. The knowledge of how the place actually runs is moving through its email, its chat, and its meetings every hour of every day. Nobody has to sit down and write it into a notebook, because the team is already writing it, to each other, in the course of doing the work.
That is the shape a collective extended mind has to take. Not a notebook you write in, which is the individual pattern and doesn’t survive being shared. A brain that builds itself from the stream of messages the team is already sending. The upkeep problem that kills every wiki disappears, because there is no upkeep step. The maintenance is a byproduct of everyday activity.
This is the move I don’t think the notebook framing can make, and it’s not a knock on Johnson or on Gemini Notebook. A personal tool should be personal. Otto’s notebook would be worse, not better, if it tried to fill itself from everyone’s calendars. The collective case is a different problem with a different answer, and it opens up right where the individual one runs out.
What this actually takes #
I want to be careful about where the argument ends and the product begins. We’re building the brain at this+that toward exactly this: a knowledge layer that assembles from the messages a team already exchanges, so it stays current without anyone tending it. Parts of that run today. The people and relationships in your world keep themselves current, and our workflows write what they learn back into the brain. The full version, where your whole message history populates a brain that then keeps itself accurate across the company with no one wiring it up, is what we’re building toward, and we’re piloting it with a design partner rather than claiming it at scale.
There’s a real catch worth naming, too. A brain that builds from messages still needs the messages to accumulate before it knows much. It starts closer to empty than a notebook you’ve spent a year curating, and it earns its accuracy over weeks of real traffic, not on day one. The bet is that automatic, even from a slow start, ends up ahead of diligent-by-hand, because diligence at team scale is the thing that always slips. We’re also working to shorten that slow start rather than just wait it out. The plan is to seed the brain from the messages you already have. You tell us what matters, say that customers are important and hand us a few examples, and we comb back through your history to propose the rest: the other customers, the products you sell, the competitors you track, with you reviewing the list before any of it sticks. It won’t erase the slow start, but it means you don’t begin from nothing.
Johnson is right that the extended mind is the frame for all of this. Otto’s notebook, the commonplace book, NotebookLM: that lineage is real and it’s still being extended. The part still open is the collective one, and you can’t get there by handing everyone a better notebook. You get there with a mind the team never has to keep. That’s the one we’re building, and you can see where it’s headed here.
Key takeaways #
- Steven Johnson frames NotebookLM, now renamed Gemini Notebook, through the extended-mind theory from Clark and Chalmers’ 1998 paper: Otto’s notebook is literally part of his mind because he keeps it current.
- The extended minds that survive are personal and hand-curated, from the commonplace book to the diary to NotebookLM. The curation by one owner is what makes it work.
- That model breaks at team scale. A shared notebook’s upkeep falls on everyone and therefore on no one, which is the real reason wikis rot and shared drives become graveyards.
- When the written record can’t be trusted, companies default to an oral culture where finding anything means knowing who to ask. It works, but it spends two people’s attention per answer, loses part of the record with every departure, and turns half of onboarding into network-building.
- A collective extended mind can’t be written in by hand. It has to assemble itself from the email, chat, and meetings a team is already producing, so the keeping is a byproduct of the working.
- That self-assembling company brain is what we’re building toward. Pieces run today; the full, self-maintaining version is in a design-partner pilot, and even then it needs real message traffic to accumulate before it’s smart.