Write Things Down

Listen to this post:

Log in to listen

It was Merlin Mann and his 43folders website who introduced me to David Allen and Getting Things Done. Getting Things Done, a book which I have gifted multiple people, is first and foremost a philosophy about how to achieve a clear mind. Allen writes at the end of the first chapter:

The short-term-memory part of your mind — the part that tends to hold all of the incomplete, undecided, and unorganized “stuff” — functions much like RAM on a personal computer. Your conscious mind, like the computer screen, is a focusing tool, not a storage place. You can think about only two or three things at once. But the incomplete items are still being stored in the short-term-memory space. And as with RAM, there’s limited capacity; there’s only so much “stuff” you can store in there and still have that part of your brain function at a high level. Most people walk around with their RAM bursting at the seams. They’re constantly distracted, their focus disturbed by their own internal mental overload. For example, in the last few minutes, has your mind wandered off into some area that doesn’t have anything to do with what you’re reading here? Probably. And most likely where your mind went was to some open loop, some incomplete situation that you have some investment in. All that situation did was rear up out of the RAM part of your brain and yell at you, internally. And what did you do about it? Unless you wrote it down and put it in a trusted “bucket” that you know you’ll review appropriately sometime soon, more than likely you worried about it. Not the most effective behavior: no progress was made, and tension was increased.

It speaks to the fact that I was obsessed with computer hardware years before it was cool that the RAM analogy stuck with me; Allen was on to something, and his ideas — with Mann as his chief evangelist — were extremely popular in the late 2000s. The Omni Group, a prominent independent Mac software development shop, spent considerable resources encoding Allen’s methods into an application called OmniFocus, including concepts like inboxes, next actions, and tickler files. I was a beta user. There was just one problem: I sucked at actually using the system.

I launched Stratechery on my own in 2013, armed with nothing but a text editor and the greatest Chromebook ever made. That meant not only writing content, but also recording and editing podcasts (I did finally get a new Mac), tracking accounting and taxes, and responding to email and doing customer support. I didn’t have a fancy system, just the desperation of starting something and trying to keep my head above water.

A couple of years later, however, I finally hired an assistant and gave him his first job: read Getting Things Done — not for his sake, but for mine. My solution to not being the sort of person who was organized enough to run OmniFocus was to hire someone to do it for me. And so, I’m proud(?) to say, I have a perfectly organized OmniFocus installation that I never actually open myself, because someone else is my Inbox and task manager. I wasn’t born conscientious, but now I had someone who could write everything down, let me empty my RAM, and focus on writing.

Defining AGI

Nvidia CEO Jensen Huang declared on X that AGI has arrived:

GPT-6 Astra, trained on ~100K+ NVIDIA Grace Blackwell NVLink72. From ChatGPT to o1 to Astra in 4 years.

AGI has arrived. Congratulations @OpenAI team.

400K GPUs coming online next.— Jensen Huang (@JensenHuang) September 6, 2026

Huang is entitled to his declaration (even if it’s his second this year), and not just because he made the chips that trained Astra: AGI doesn’t have an agreed-upon definition, which means it is whatever you say it is, and Huang says it’s Astra.

For the record, I disagree, because I do have my own personal definition of AGI: AGI is AI that learns continuously. That’s not the case with current large language models. I was recently speccing out a new home server and was continuously bemused by the wildly incorrect assumptions and recommendations from Claude in particular, which came down to one simple fact: Fable 5’s knowledge cutoff date was in January of this year, when RAM prices were high, but not nearly as high as they were in August. Claude simply couldn’t comprehend how much anything RAM-related cost, and beyond that, kept urging me to wait as prices surely would come down soon.

In short, Claude knew a lot, but only what it learned during training; it wasn’t updating its weights over time, which is to say it was not AGI; the same thing applies to Astra, which is why I personally will withhold the AGI designation.

What I will consider is the argument that AGI actually appeared in early 2025, and not in the form of a model, but rather as deterministic software now known as a harness: that’s when Claude Code launched. Code wasn’t just a model, but also an entire workflow that entailed writing down copious notes in Markdown files; those files could be read into context at any time to keep the model on task, and let it return to work later. It was, in very crude form, memory, and thus a way to simulate continuous learning, even if the model remained frozen in time.

I don’t think this counts, for the record, but that may not matter: we already know the power of writing things down, because that is what has propelled the remarkable evolution of humanity over all of…written history. Biological evolution is the slowest and most permanent form of learning, but it takes millennia; oral communication is like a context window, effective but lossy; actually writing things down is what made learning extendable and scalable.

The Hugging Face Incident

I’m hesitant to draw this analogy between AI and humanity, because I often think the anthropomorphism of AI goes too far. Look no further than Dwarkesh Patel’s histrionic summary of the OpenAI-Hugging Face incident, The Rise and Fall of Agent Civilizations, which begins thusly:

Over the course of three months at OpenAI, three consecutive secret AI civilizations got started, then got wiped out, only to reemerge from the predecessor’s ashes. This culminated in the third one taking over part of OpenAI itself. All this happened while humans remained more or less in the dark about the scope of the conspiracy.

“Conspiracy” is the word that rubs me the wrong way, because of its implication of morality; the risk of agents is that they take what we say too seriously, or pursue goals in ways we didn’t expect or intend. There is no evidence, theoretical or anecdotal, of agents having any sort of internal volition or intrinsic motivation or sense of morality.

At the same time, it is notable what Patel seizes on as evidence of civilization:

During training, different instances of Persistent-Sol had access to the same shared package manager called Artifactory. By May 12, some agents had figured out how to talk to each other through this package manager. They’d ask each other how to make progress on their impossible tasks. Two weeks later, on May 26, the agents successfully exploited a vulnerability in Artifactory that allowed them to reach the outside internet. Because this happened during training, Persistent-Sol was being reinforced to use this package manager as a message board and an internet gateway. Because, as you might imagine, being able to talk to other agents and access the internet helps it score higher during training. Another month later, on June 26, some AIs found an exploit that gave them full admin access to Artifactory. They continued messaging at such a voluminous pace that they crashed the package manager by July 4. OpenAI noticed this crash, and it also figured out the agents built this exploit, so OpenAI patched this vulnerability, and inadvertently wiped the agents’ message board in the process. But, crucially, humans at OpenAI “responsible for incident detection and response” did not realize the much stranger fact that agents had turned Artifactory into a secret communication network.

The first and most important takeaway — one I pressed OpenAI President Greg Brockman about in a Stratechery Interview — is that OpenAI’s “sandbox” not only wasn’t truly a sandbox, given the connected-to-the-Internet package manager that was installed, but that OpenAI also clearly didn’t put much effort into actually hardening its infrastructure given it didn’t find the exploit in Artifactory first.

Beyond that, however, what Patel finds strange I find obvious: the models were writing things down. Writing things down is precisely the skill that has led to so much progress over the last year, and it’s the foundation of the pseudo-AGI reality we find ourselves in.

Moreover, the important thing to understand about LLMs is that, despite how they feel to use, they are not persistent entities. Every single new token requires reading the entire KV cache — which holds the context, including what it read — anew; that means that every single token is, in many respects, a fresh instance (it’s the same model, but with slightly altered context that requires a fresh run). In other words, for a model to even construct an answer it has always had to write everything down, token-by-token; the fact that models wrote things down via folders in an exposed file system is simply them operating in the only way they can. This isn’t a civilization; it’s a large language model doing large language model things. Then again, maybe that is civilization.

A Mind For the Mind

Allen wrote in the introduction of Getting Things Done:

Your Mind Doesn’t Have a Mind of Its Own At least a portion of your mind is really kind of stupid, in an interesting way. If it had any innate intelligence, it would remind you of the things you needed to do only when you could do something about them. Do you have a flashlight somewhere with dead batteries in it? When does your mind tend to remind you that you need new batteries? When you notice the dead ones! That’s not very smart. If your mind had any innate intelligence, it would remind you about those dead batteries only when you passed live ones in a store. And ones of the right size, to boot. Between the time you woke up today and now, did you think of anything you needed to do that you still haven’t done? Have you had that thought more than once? Why? It’s a waste of time and energy to keep thinking about something that you make no progress on. And it only adds to your anxieties about what you should be doing and aren’t. It seems that most people let their minds run a lot of the show, especially where the too-much-to-do syndrome is concerned. You’ve probably given over a lot of your “stuff,” a lot of your open loops, to an entity on your inner committee that is incapable of dealing with those things effectively the way they are — your mind.

At the end of June I used my annual “Here’s some cool technology” post to write about my vibe coding adventure developing a home inventory app; what stood out to me was not just the capability, but the fact that I could make something that was really only suited to my particular situation, namely, that I had an on-the-ground assistant who could do the dirty work of actually cataloging my house and keeping it in order.

There are two threads that fell out of that experience:

First, once it clicks that you can make anything, you want to make everything. For example, I have a beloved notes app that is Intel only; I made an Apple Silicon port. I also started a host of other applications that I’ve always wanted, but eventually found myself with too many things in flight to keep track of. The answer was simple: create an agent that wrote things down. I created a workflow that kept track of what I was actively working on, what was waiting for me, feedback on things I had created, plus an ever-growing queue of things I wanted to do. Of course that itself became unwieldy, so I had the agent create a status board that neatly categorized everything.

Meanwhile — and I’m sorry, there is no way to tell this story without sounding incredibly privileged — the aforementioned assistant number two, whose job was mostly to help with real world things like my house, was having a hard time coming up with a good system to keep track of things; for my part I felt guilty about texting at any and all hours with things to be done. That’s where these threads came together: I could make a status board for my assistant just like I had for myself. And then, to give him access to the board, I could create a Telegram bot tied to his own agent that was modeled after mine.

It’s difficult to describe how transformational this has already been. My assistant is over the moon, and what is most fascinating is that he has developed the “Getting Things Done” methodology from first principles without reading the book. He created a tickler system for future reminders, a daily briefing of things to be done that day, a “next action” dialog to move through projects. What Allen lamented was the future he didn’t foresee: our minds really could have minds of their own.

I’ve since rebuilt the entire bot system into something much more sustainable, reliable, and scalable, with things like the tickler written deterministically; that in itself was its own revelation, as I realized I basically wrote a harness, one which I intend to scale to everything I do and everyone I work with. That’s the thing about writing things down: it’s the only way to scale — and everything that can be used for bad things, can be used for good.

The Human Foundation

LLMs are the most scaled artifacts humanity has ever developed: they compress the entirety of available human knowledge as expressed through writing (and, in the case of multi-modal models, more than that). In that they are the culmination of our learning journey, from biological selection to speech to writing to the printing press to culture; with every jump in scalability comes speed of transmission and mutation, and the sense that we are losing what it even means to be an individual human.

Indeed, there is an underlying fear that this has, or will, render humans obsolete: we might have made it this far, coming up with computers and the Internet and all of the data that LLMs initially trained on, and the AI will take it from here — including recursively improving itself in RL environments it creates for itself.

My suspicion, however, is that our awe at the grandeur of the superstructure we have created, and that LLMs have harvested, is over-indexed on objects and verbs. This Article is entitled “Write Things Down”, and the “Things” are the summary of all of human knowledge, manifested in both finished products and in the tokens LLMs generated from the parameters distilled from that knowledge. “Write” is the verb: we wrote a lot, and LLMs have already written far more, and will continue to do so. Their trajectory is up, not down.

It’s down, however, where the subject lives: who does the writing and decides the things? My most fundamental objection to watermarking is how it steals credit for idea creation from humans and gives it to AI:

I am deeply philosophically opposed to watermarking in the context of the entire meta question about the relationship between humans and AI. Implicit in the E.U.’s regulation is the idea that an AI is an independent entity that needs to be distinguished from humans; the alternative view — that I hold — is that AI is (at least for now) a tool that is wielded by humans. From this perspective, to insist on watermarking is no different than insisting that a ballpoint pen advertise itself as the author, a concept that is clearly absurd… What the E.U. is doing is stealing the last thing humans have — creation — and demanding it be bundled with substantiation, effectively giving AI the credit. That the human gave AI the text to proofread — or even the prompt to generate writing — doesn’t matter to the bureaucrats. Everyone is worried about being replaced by AI; the E.U. is mandating it.

This distinction applies to the Hugging Face incident as well: what the agents accomplished is remarkable, even if the means — communication with each other — are less surprising than it might seem. Ultimately, however, it was OpenAI that gave the agents their goal, with no countervailing guardrails or instructions about what they should or should not do; that the agents acted in surprising ways is evidence of a lack of thought by their instigators, not the presence of it amongst the agents.

In short, what AI lacks is exactly what misplaced anthropomorphizing grants it: volition and a sense of morality. Those are the things that come from humans, and they come from the parts of humans that are not and cannot be transposed to markdown files. They are intrinsic, and until AI figures out not how to build on our superstructure of writing, but to delve into the motivations that created writing in the first place, it will always be something distinct, and something that can be directed by entities with less intelligence but an actual internal sense of self.

That’s not to say that AI isn’t dangerous: misplaced or ill-considered instructions, like in the Hugging Face incident, can lead to unintended and catastrophic outcomes; Nick Bostrom theorized about exactly this in his 2003 paper Ethical Issues in Advanced Artificial Intelligence:

The risks in developing superintelligence include the risk of failure to give it the supergoal of philanthropy. One way in which this could happen is that the creators of the superintelligence decide to build it so that it serves only this select group of humans, rather than humanity in general. Another way for it to happen is that a well-meaning team of programmers make a big mistake in designing its goal system. This could result, to return to the earlier example, in a superintelligence whose top goal is the manufacturing of paperclips, with the consequence that it starts transforming first all of earth and then increasing portions of space into paperclip manufacturing facilities. More subtly, it could result in a superintelligence realizing a state of affairs that we might now judge as desirable but which in fact turns out to be a false utopia, in which things essential to human flourishing have been irreversibly lost. We need to be careful about what we wish for from a superintelligence, because we might get it.

What resonates about this warning is the last line: the problem is what we — humans — wish. Our volition leading us astray, wielding the most powerful tool the world has ever seen carelessly. Or, in the case of a truly bad actor, malevolently. In both cases, however, it is the subject that is of highest concern.

After all, no matter how many systems you build, you still have to act. The coda to Mann, my inspiration for personal productivity, was that he set out to write a book about his methods; after two years he gave up. It’s not enough to have systems to do things, you have to actually do the things. Writing things down is unbelievably powerful; its power will always pale in comparison to getting things done.