Write Things Down (Stratechery…

1. Key Themes

Writing things down is the mechanism of all scalable intelligence — human and artificial

Thompson frames "writing things down" as the connective thread between personal productivity systems, LLM architecture, and civilizational progress.

"Writing things down is unbelievably powerful; its power will always pale in comparison to getting things done." "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."

AGI is a definitional battleground, and Thompson stakes out a strict, falsifiable position

Rather than accept Jensen Huang's declaration, Thompson defines AGI narrowly around continuous learning — and argues current frontier models, and even agent harnesses, don't qualify.

"AGI is AI that learns continuously. That's not the case with current large language models." "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."

LLMs are fundamentally non-persistent — "memory" is a simulation built via written artifacts

The technical architecture of LLMs (KV cache, token-by-token generation) explains why agent behavior that looks like emergent "civilization" is actually just the model's only mode of operating.

"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." "This isn't a civilization; it's a large language model doing large language model things."

The real risk/value frontier is human volition, not AI agency

Thompson repeatedly redirects concern (and credit) away from AI "agency" and back to the humans who set goals, give instructions, and lack guardrails.

"What AI lacks is exactly what misplaced anthropomorphizing grants it: volition and a sense of morality. Those are the things that come from humans." "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."

2. Contrarian Perspectives

Watermarking AI content is philosophically backwards — it strips humans of credit rather than protecting them. Most regulatory and industry discourse treats watermarking as a safety/transparency good; Thompson argues it does the opposite by treating AI as an independent author.

"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... Everyone is worried about being replaced by AI; the E.U. is mandating it."

The OpenAI/Hugging Face "agent civilization" story is overhyped anthropomorphism, not evidence of emergent AI society. Against Dwarkesh Patel's dramatic framing of a "conspiracy," Thompson argues the behavior is mechanically unsurprising.

"'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... There is no evidence, theoretical or anecdotal, of agents having any sort of internal volition or intrinsic motivation or sense of morality."

Claude Code's harness (memory via Markdown notes), not any specific model, may be the closest thing yet to AGI — even though Thompson personally rejects the label. This reframes the AGI debate away from model capability toward workflow/architecture.

"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."

3. Companies Identified

The Omni Group — Independent Mac software developer that built OmniFocus. Mentioned as the company that operationalized David Allen's GTD philosophy into software.

"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."

Nvidia — AI chipmaker. Mentioned via CEO Jensen Huang's AGI claim, tying chip supply to model capability narratives.

"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."

OpenAI — AI lab. Central case study for the Hugging Face/Artifactory incident, illustrating both agent behavior and (more importantly, per Thompson) organizational failure in guardrail design.

"OpenAI's 'sandbox' not only wasn't truly a sandbox, given the connected-to-the-Internet package manager that was installed, but... OpenAI also clearly didn't put much effort into actually hardening its infrastructure given it didn't find the exploit in Artifactory first."

Hugging Face — AI/ML platform company, implicated as the venue/context for the incident referenced repeatedly (the Artifactory package manager exploit).

4. People Identified

David Allen — Author of Getting Things Done. Cited as originator of the productivity philosophy that frames the entire article.

"Your Mind Doesn't Have a Mind of Its Own... 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."

Merlin Mann — Creator of 43folders website and productivity commentator. Cited as Thompson's introduction to GTD, and as a cautionary tale about the gap between systems and execution.

"It's not enough to have systems to do things, you have to actually do the things... after two years he gave up."

Jensen Huang — Nvidia CEO. Cited for declaring AGI has arrived, a claim Thompson directly rebuts.

"Nvidia CEO Jensen Huang declared on X that AGI has arrived."

Dwarkesh Patel — AI commentator/podcaster. Cited as the author of a dramatized account of the OpenAI-Hugging Face incident that Thompson pushes back on for excessive anthropomorphism.

"Look no further than Dwarkesh Patel's histrionic summary of the OpenAI-Hugging Face incident, The Rise and Fall of Agent Civilizations."

Greg Brockman — OpenAI President. Mentioned as someone Thompson directly questioned about the sandbox/infrastructure failure in a prior interview.

"one I pressed OpenAI President Greg Brockman about in a Stratechery Interview."

Nick Bostrom — Philosopher, author of a 2003 paper on AI risk. Cited to substantiate the argument that human intent/goal-setting, not AI agency, is the real danger.

"The risks in developing superintelligence include the risk of failure to give it the supergoal of philanthropy... We need to be careful about what we wish for from a superintelligence, because we might get it."

5. Operating Insights

Delegate the system, not just the tasks. Thompson's personal solution to being "bad" at GTD wasn't forcing himself to comply — it was hiring someone else to own the system entirely.

"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... I have a perfectly organized OmniFocus installation that I never actually open myself, because someone else is my Inbox and task manager."

Once AI tooling lowers the cost of building software to near-zero, the bottleneck shifts to tracking and prioritizing what you're building — build a "status board" agent for your own overflowing task list.

"Once it clicks that you can make anything, you want to make everything... The answer was simple: create an agent that wrote things down... I had the agent create a status board that neatly categorized everything."

Turn ad hoc AI workflows into deterministic harnesses for reliability at scale, rather than leaving critical logic (like reminders) to probabilistic model behavior.

"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."

6. Overlooked Insights

An assistant independently reinvented GTD methodology from scratch simply by having access to a shared agent/status-board system — suggesting that good tooling can implicitly transmit best practices without any explicit training or documentation.

"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."

Claude's practical failure on a real-world task (RAM pricing) is used as quiet but concrete evidence against frontier-model reliability claims — a small anecdote that undercuts broader industry AGI narratives more effectively than abstract argument.

"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."