Now, recall (ha) that our main worry was around LLMs bypassing our learning and memory processes. Fortunately, we have decades of research on how to design education technology to promote, not disrupt, learning. In fact, some chatbots now offer a "study mode" option that incorporates design principles from this research -- think LearnLM, StudyGPT and Khanmigo.Well, the last one is more like a study mode offering a chatbot, not the other way around. These include tricks like scaffolding learning so difficulty increases over time and fading guidance as students get better. If these modes continue to develop, it's reasonable to think that LLMs could be pretty effective tutors going forward.

But these modes are designed for the student context, in which clear curricula exist and the main objective of the student is to learn. It's the other contexts I'm worried about. The Sovereignty Trap grows teeth when I am not a student, when I am dealing with information outside established curricula and sometimes can't even tell what the right questions are. These are contexts in which turning on "study mode" may be a luxury orthogonal to my main goals.

It's starting to feel inevitable that machines might get so good at thinking for us in these contexts that we are no longer allowed to do so ourselves. Like growing your own food, thinking for yourself may seem like an archaic and artisanal choice that has no place in the productive world. That fear has two parts, one of which I'll toss immediately:

1. Machines will get very good at thinking. (Regardless of where you predict they'll end up, they are already useful enough to replace some parts of our thinking.)
2. We stay the same at thinking, and our tools do not help us get better.

Point 2 we can have a productive noodle on. And we must, because if we want someone or something to save us -- if we want policy made to protect knowledge workers, if we want unions and big companies to commit to some alternative, if we want knowledge tools we can use without signing away our cognitive sovereignty -- we need a better ask than, "Everybody stop using AI!!!" You and I both know that's not likely. It is likely, however, that the ways in which we interact with LLMs can evolve past the oracular paradigm. As my friend Rob says, "AI is the worst it's ever going to be, right now" -- so I'd argue that we need some alternatives to the oracular paradigm pretty urgently.

Part of the problem with LLMs right now is that they're not really tools. Tools are built for specific problems, often ones that we can't solve ourselves.Well, I once watched someone try to open a bottle of wine with a shoe and a butter knife, but...neither here nor there. Without constraints, LLMs are eager oracles -- leaping in to do everything they can, whether or not we want them to, and bypassing our cognitive abilities instead of extending them. But before we can ask the engineers at tech companies to build us tools,Or, in the age of Claude, build them ourselves we have to know which problems we like to solve ourselves and which we're fine getting locked out of. Then we need to point resources -- experiments, design patterns, software, hardware, policies, boycotts -- at building tools that help us solve these problems instead of doing it for us.For those still stuck on the last footnote: It's like, we want a bottle opener, but we don't want a robot to drink the bottle for us.

Because almost all knowledge work is in the crosshairs (or at least, that's what the AI optimists claim), identifying and protecting the things we value will probably require a sector-by-sector or even person-by-person reckoning. But I'm not taking that on right now. This series focuses on the blurry, frustrating processes that move us from overwhelmed novices to insightful experts.

Here's my goal: I want to design tools for knowledge work that let us think our own thoughts, maintain our own knowledge schema and recognize our own decisions. These tools will come in many combinations of design patterns, software and hardware, but I'm not ready to commit to one yet, so I need a word for all of them.If you're someone who needs to know specifics up front, check out this example.

I'm going to call them poietic tools. Poiesis means creation or bringing-forth. It's often used to discuss the process of revealing something that is hidden, the way that a sculptor reveals a form. It reminds me of the productive struggle and iteration inherent in knowledge work, from the early days of chipping away at a block of information to the finishing touches on a new piece. Hence, "poietic."

By "tool," I mean "the system that we use to interact with knowledge." This can include the hardware that links me to the digital workspace I do this in, plus the design patterns and software that define and create that workspace. I'm also going to include the digital companions present in that workspace, like your Claude or some personal AI assistant ("AI assistant" going forward). We're not concerned with whether these assistants are good or bad, or if they will get better. Here we assume progress in AI; the questions are about how we work and what we want.

What do I want from poietic tools? A good poietic tool does the following for me:These points are collected from psychology, cognitive science, education technology design, other design, what we know about intelligence analysts, and people who write and build about how we interact with knowledge, including Andy Matuschak, Maggie Appleton and Amelia Wattenberger.

1. It helps me cope with a lot of new information.
2. It helps me explore information without biasing my path.
3. It links new information to old so that I strengthen and draw on knowledge as I learn.
4. It helps me move between levels of abstraction (e.g., from direct quotes to summaries.)
5. It helps me transition from using established schema (the way others see the world) to constructing my own in a way that persists and can be audited.This one is a little abstract -- put another way, "It lets me develop and record my own opinions as I learn."
6. It helps me cope with the complexity of the work I do and the arguments I produce.
7. It helps me tolerate uncertainty and iterative thought in the face of outside pressures.
8. It gives me low-friction opportunities to correct my AI assistant's behavior.
9. It protects unformed parts of my thinking from the eager oracle.Again, a little abstract. Put another way: LLMs can and will jump to fill in the blanks before I'm ready (the "eager oracle.") But it's hard not to be prematurely biased by a polished answer, and even harder to justify thinking for yourself if an answer appears too early.

You probably noticed some contradictions in that list above. We don't want these tools to bypass our own learning and memory, but they need to decide what information to remind us of. They shouldn't skew or derail us as we explore new information, but they can't hide useful tidbits from us. They need to help us cope with the complexities of arguments without jumping in to fill in the blank before we're ready. That is one fine line! But we've got to get it right. Let's start with the doozies.

"Exploring information freely" is a doozy. There's always a massive imbalance between the information I know about and the information I don't. And let's remember that we bring machines in to help us tackle problems of scale; if there were only ten things to know in the world, we would never have built LLMs. But chatbots exacerbate this imbalance. They're like walking up to a library and, rather than being allowed inside, submitting a question through a slot and receiving a set of books. I don't know what else there is to look at, or what is influencing the decision of the entity behind the slot. When I go to my AI assistant for resources and it brings me a selection of twenty documents out of a million, are those the ones I would choose if I could see all of them? Am I really exploring "freely" the way I would if I could read everything?

Here's another doozy: If I want to get a sense of what's in a million documents, I need to work at a higher level of abstraction than the documents themselves. For text, that typically means summarizing, which is not a neutral process. In fact, it's a series of choices: Choosing words to represent a set chosen by others; deciding what is gist and what is detail; clustering and constructing concept hierarchies; throwing information away. Here we have the same problem: How do I know that my AI assistant keeps the information I would keep? How does it know what details might be important?

I don't think there's a clean answer here. Sometimes, even I don't know what information I'm looking for. But it's not in my best interest to be locked out of the decision-making process by the eager oracle. To tip control back towards me, I would like to know in broad strokes what information is out there; I would like to audit and adjust how my AI assistant selects and abstracts it; and I would like my AI assistant to adopt my own decision-making strategies as I evolve them. So the challenges around selection and abstraction really come down to whether my AI assistant knows enough about what's in my head to make the decisions I would make, when the scale of information is such that I can't or don't want to make them myself. That should be easy, right? Can't I just tell it?

Maybe you can, but I think it would be difficult and disruptive to describe what's in my head. Anyone who has worked on a long-term project knows that thinking evolves: Certain pieces of information come to center stage or fade into the background, the story changes, terms are coined and tossed. Sometimes it helps to articulate these changes in writing, but sometimes that just gets in the way. And god forbid I have to interrupt what feels like my main work to update a resource for my AI assistant to use. So what's a girl to do?

My friend has a story about the woman who caught bin Laden. The CIA had been getting nowherePutting it mildly when this analyst decided to print the entire dossier out and revisit it with fresh eyes. She commandeered a room, organized all the intelligence into physical piles and walked around those piles for about ten months, Eventually she figured out that two people were actually the same guy, and the rest is history, etc. But the thing I love about this story is the piles. It's not like the CIA didn't have computers! There was just something about physical piles that made a complicated, arduous investigation tractable.

Let's think about why. If I can visualize and organize information in a room, then pressure on my working memory is relieved. I can "hold" as many thoughts as I like, right where I can see them. I can cue information with a glance instead of clicking through tabs. I can organize things according to whatever factors matter, e.g., by topic, by timeline, and so on. I can draw connections between groups without having to nail down exactly why I think they're related. I can even quarantine parts of my work I'm not sure about yet.

Our memories are tied to physical locations, too -- that's why tricks like memory palaces and the Method of Loci work. Think about the way many people keep books on their desks, not because they're actively reading them, but because they act as cues for ideas and trains of thought. And if we're tracking the points in the Sovereignty Trap, the effort of moving information out of its initial arrangement into one that reflects my own mental model also forms stronger memories for that information.

We see the effects of space on cognition in Andrews et al.'s paper on "Space to Think."00.Christopher Andrews et al. "Space to Think: Large, High-Resolution Displays for Sensemaking." CHI 2010, April 10 - 15, 2010. They "construct[ed] large display environments in which space [had] real meaning" by wiring together eight LCD panels in a gentle curve around the user. Then they ran two studies: A comparative study with grad students who either used a normal monitor or the big display; and an observational study of professional analysts. Both groups were asked to sift through hundreds of documents to "solve" a scenario. I won't summarize all the results here, but two things jumped out at me. The grad students using the big display had much better memory for articles they had read previously, and they got "lost" less often. And the analysts used the display to externalize complicated semantic information through the spatial organization of documents. Take a look at this quote:

Most of the organization tended to take the form of clustering, or rough categorization... Interestingly, these categories ranged greatly in specificity. Some were quite general (e.g., “potentially interesting” or “background”), while some were quite specific (e.g., documents that discuss a particular individual or organization). Occasionally, we observed these categories evolve over time as more information was acquired. For example, subject A1 started a cluster that she referred to as “background” information. However, as she learned more and placed new documents in the cluster, the meaning shifted to “critical documents”. The contents did not change - her interpretation of them shifted.

Having "space to think" allowed the analysts to visualize their knowledge schema, using space to define topical clusters and timelines. As Andrews et al. note, "It is...far easier to arrange objects in space and use the perceptual system to recognize categories and properties present...than it is to try to memorize all of the characteristics of every object and internally compute relationships." The interface also allowed analysts to shift strategies without friction. The analyst in the quote above changed the meaning of a cluster by adding more information, something that I suspect an LLM could pick up on if it were watching.

And that brings me back to our doozies. Organizing information in space is not only a thing I do to clear my own head, it's also an implicit definition of a knowledge schema -- specifically, the knowledge schema that my assistant needs to make the right decisions about how to select and abstract information. By moving information around, I create a kind of "data exhaust" or "computational wear" that my assistant can look at to see how my schema is changing.00.William Hill et al. "Edit wear and read wear." CHI 1992. There's even evidence that human-generated schema derived from spatial organization can improve the accuracy of LLMs doing sensemaking.00. Xuxin Tang et al. "Steering LLM Summarization with Visual Workspaces for Sensemaking." arXiv preprint arXiv:2409.17289, 2024. So taking advantage of physical space not only lowers my cognitive effort, it provides more information to my assistant about how I want things done without forcing me to explicitly define it. All of these advantages tip control of the process back towards me.

At this point you might be thinking, cool, but what about existing options for spatial organization like Notion, Obsidian and Muse? This is why I was loath to specify form factors for poietic tools too early. I agree that Notion, Obsidian, even the chat interface can be considered poietic. But I also think the poietic tools that best fulfill our earlier goalDesigning tools for knowledge work that let us think our own thoughts, maintain our own knowledge schema and own our decisions. will take knowledge work beyond the monitor. Andrews et al. are quite clear that space needs to be big enough to "have real meaning." For small information, like one or two Claude chats or a simple spreadsheet, our monitors are great; for big information, we need to stretch out. We need something that helps us see what we think.

Getting more serious about the physical world opens up other ways of tipping control back towards us, too. Let's go back to that earlier wish list. If we're serious about the physical world, then we've committed to multiple input modalities, including gesture. That gives us low-friction opportunities to correct AI assistant behavior with the slash of a pen or the wave of a hand. Similarly, I can circle parts of the landscape that my AI assistant should not touch, making it easy to protect myself from the eager oracle.

Lastly, externalizing my thoughts does more for me than relieve my working memory. It also helps me tolerate uncertainty and iterative thought in the face of outside pressures towards the fastest answer. It provides visual indicators that we can use to monitor and justify our progress. Done well, these tools won't just provide gamified indicators like "you read 5 papers today!" but beautiful visualizations of structure emerging from chaos. And this brings me to the second soapbox I feel justified in clambering up on: I want tools that make thinking for myself feel joyful and productive. Sometimes that's a chat interface or a Kanban board, but other times it's a desktop digital garden or a magic gumdrop that writes in the air.

Between you and me, I think it's the joy of using poietic tools that will put them in every office. Sure, they have to help us think better. But they also have to help us feel better than the oracular paradigm does. They have to help us regain our cognitive agency, our tolerance for uncertainty and the delight we feel when we do really good work. They have to be designed for what we want, not for maximum engagement.

Coming back to my earlier point: This probably looks different for everyone. Think about selection and abstraction. When I start a project, I care a lot about the latter. The words I use are part of my process, and to be given a blanket term that doesn't fit can derail or extinguish an idea altogether. I care more about whether an eager oracle infects me with its terminology more than I care about the information it brings me, at least at the start.Past a threshold of accuracy and quality, I mean. I figure, I'll start somewhere, and as I talk to people I will make my own way through the information space and begin to know better what to ask for.

Other people are different. I bet at least some of you are thinking, "I don't care what you call what I'm working on, I just want you to bring me everything that's relevant." People prioritize different behaviors from their tools depending on what they value, how they work and even what task they're doing. And this is true across many facets of knowledge work, from how we take care of our memories to how we write to how we discuss with each other. My point is that poietic tools that take advantage of the physical world help me select for a specific type of control. They help keep the eager oracle out of tasks I want to do myself by articulating those things in physical space. The chat interface doesn't tick those boxes for me. So why is it the only one we reach for?I could say more on the chat interface, but I think Amelia Wattenberger does it better.

Instead, we could encourage a vibrant ecosystem of tools that prioritize different things on that earlier wishlist. This ecosystem would need to draw on work from a whole list of fields and communities, some of which I've already previewed in this post. Next time, I'll take a look at the ideas, experiments and prototypes I think could launch a wave of tools that go beyond the oracular paradigm. Then I'll identify the bottlenecks that might be preventing this progress and think about what we could do to break them.