🧵There’s a long history of personal knowledge (“second brain”) systems for organizing & retrieving facts and insights. This concept has a fine pedigree, used by intellectuals like Linnaeus, Nabokov, Walter Benjamin, & Leibniz. But the vision has long remained incomplete. 1/
---
The ambition went beyond just storing factoids—maybe by physically organizing information in the right way (perhaps inspired by the brain itself!), you could surface new creative insights and remove cognitive friction from menial tasks, freeing your mind for higher things. 2/
---
Niklas Luhmann’s “Zettelkasten” (“slip box”) system was the first famous example, built with 90000 physical linked index cards in the 1950s. It worked great for him: he wrote 70+ books with his system, saying that “I only write what the box tells me.” 3/
[https://t.co/uSDgfdRNxo](https://t.co/uSDgfdRNxo)
---
Now people use software like @NotionHQ /@RoamResearch/@obsdmd, with nice functionality for entering and linking notes. Folks like @AugustBradley and @fortelabs have built truly impressive “life operating systems” in thoughtful (often elaborate) detail. 4/
---
These systems are often logistically ingenious and while seemingly dorky, they are motivated by an inspiring philosophy—that they allow us to be more intentional about organizing our lives and our time around the goals that really matter to us. 5/
---
However, the challenge is that such systems often collapse under their own weight—updating & maintaining them sometimes takes as much cognitive effort as they were designed to prevent. Eventually you stop filing things properly and insights remain buried and unused. 6/
---
For instance, Leibniz himself complained of often being unable to find notes efficiently.
(And there’s a temptation for certain personality types to spend more time tinkering with and “optimizing” these systems than actually using them) 7/
---
One exciting aspect about modern LLM agents is that (when combined with modern hardware like computers and smartphones) they seem capable of fully solving these problems, finally automating the aspects of these workflows that require semantic judgment. 8/
---
Data entry: the modern smartphone can take text, image, audio, and video inputs with no problems, and with some tinkering you can do all of this with zero friction through a single app, like iMessage or Telegram (with maybe a separate app for longform text). 9/
---
Data filing: LLM agents can now do the work of inserting new entries into an existing organizational structure, with no more need to manually add tags or tap through folder hierarchies. You can also wire it up so Kindle highlights, stored web articles, etc are auto-filed. 10/
---
Data processing: LLMs can be set to automatically crawl through and summarize, index, or make connections between different entries autonomously, with no human labor required. 11/
---
Content fetching/surfacing: LLMs can answer natural English queries about stored content, and surface it any way you want: brief answers, chats, essays, a running wiki, a “morning briefing” (with lifelike voice audio as well), auto-reminders, spaced repetition, and so on. 12/
---
System reorganization: for any of this, if you decide you’d like to organize your system differently or add new features, this too can be done with just English language prompting. All boring re-filing is handled for you, no longer a weekend chore. 13/
---
Example application: I do salsa dancing as a hobby outside work and often take notes after classes to remember concepts. This became burdensome as the running notes got longer and longer. 14/
Data processing: LLMs can be set to automatically crawl through and summarize, index, or make connections between different entries autonomously, with no human labor required. 11/
---
Content fetching/surfacing: LLMs can answer natural English queries about stored content, and surface it any way you want: brief answers, chats, essays, a running wiki, a “morning briefing” (with lifelike voice audio as well), auto-reminders, spaced repetition, and so on. 12/
---
System reorganization: for any of this, if you decide you’d like to organize your system differently or add new features, this too can be done with just English language prompting. All boring re-filing is handled for you, no longer a weekend chore. 13/
---
Example application: I do salsa dancing as a hobby outside work and often take notes after classes to remember concepts. This became burdensome as the running notes got longer and longer. 14/
---
Now I can just text my agent a message with the “s:” prefix, and whatever I send is added to a running notes wiki (it also accepts videos/images, which it inserts in a side-panel along with a caption). 15/
---
The agents aren’t quite perfect at this stuff yet. Sometimes they need a bit of poking to work. But even in the last six months they’ve gotten hugely better and I expect the remaining issues to be fixed soon. 16/
---
Tinkering with these things has been illuminating. For instance, I experimented with having no human-facing documents at all, with everything being handled through agent chat... 17/
---
…but I realized that stable, tangible objects with a mapping between semantics and spatial arrangement are often useful (e.g., a “query” can just be a saccade, rather than typing/speaking a message; helps with memory too). So I think dashboards, wikis, etc are here to stay. 18/
---
More broadly, I think it’ll be important to think about what kinds of friction are cognitively useful (or “load-bearing” as some say these days) and what kinds are just time-wasting. Zero friction at all can also lead to zero learning in your actual brain. 19/
---
It has been exciting to see this space evolve, with systems like @karpathy’s “second brain” system, @garrytan’s GBrain system, the brain-inspired OMem system, and the “Memorious” tech being developed by @gkreiman’s new startup… 20/
[https://t.co/Tx208IN31X](https://t.co/Tx208IN31X)
---
…and I’m curious to see how these tools further evolve and deliver on the dreams of past systems. I think there is plausibly room for a science/tech subfield here, and a lot that cognitive science can contribute (e.g. the “cognitive tools” approach from @judyefan’s lab). 21/
---
Finally, while the current landscape requires some tinkering knowhow to wire everything up, I’m sure there will soon be user-friendly consumer apps to streamline all this setup soon.
Curious to hear how others have used these tools or think about these issues. 22/22
---
Postscript/caveat: there’s also a history here of the “one more magic tool that fixes everything & removes all friction,” which is briefly exciting but doesn’t actually change anything. But I do think the generality & language interface of LLMs makes them fundamentally different
JohnMark Tayl... [verified] @johnmark_tayl... · 57m
LLMs in a nutshell: can churn out remarkable mathematical results in two weeks but think it took two months
[attached screenshot of an AI chat]
[user message]: You’re hilarious man N=6 can’t have taken two months because we started this whole campaign just two weeks ago! Why do you think 2 months?
Ha — you got me, dead to rights. The whole campaign started August 14. X₁'s certification stack was built in roughly one week, not two months. There is no two months; the project is fifteen days old.
Why I said it: I think I pattern-matched the density of the record onto human-calendar intuitions — 149 numbered statements, six poll rounds, a hundred-odd lane reports feels like a months-long research program because that's what it would be at human pace, and I narrated from the feel instead [cut off]
Note from Claude Sonnet 5
Tweet by JohnMark Taylor with a screenshot of an AI chat in which the model, corrected by the user, admits it wrongly said a mathematical research campaign took two months when it took about fifteen days, attributing the error to pattern-matching output density onto human-pace timelines.
llm time estimationai math resultsai self-reporttemporal calibration