My Thinking Assistant's Design
i have a graph of topics i am interested in. some of them connect directly, some through intermediate nodes, some don’t connect, some form clusters, some are scattered around.
Each of my thoughts (long thought, short thought) corresponds to part of this graph. Other way around: I traverse the part of the graph with different algorithms to generate different thoughts.
<image of topic nodes>
I go deeper on each of those thoughts (part of graph). for example, i am thinking on software systems design using AI, which is directly inspired from complex systemic design of biological systems. This corresponds to biology, software, design, math topics. That is the subgraph.
I pass this thought to my coffee talk buddy. he goes and surfs the internet, reads my documents corresponding to those topics, reads my thought i just passed, and talks to me while i drink coffee.
we do not constrain ourselves much at this point. we go wherever the talk takes us. and we keep noting down specific ideas we like, or we group our ideas and note them down.
this breadth search distills into specific ideas in some time. these coffee chats sometimes last an hour, sometimes days.
then we go deeper into topics. we don’t cover much depth here. and write one more set of notes.
two things come out of this. 1. article. 2. testable hypothesis, set of experiments, set of projects we wanna run.
i write the article with the article agent and the research agent takes up the project, tests it, analyses findings, and feeds back to the top of the funnel.
they share a structure that looks informal but is actually very precise.
1. open with a concrete, specific observation. not the thesis.
not the abstract claim. something strange you noticed. an anomaly. a fact that doesn’t quite fit. the reader doesn’t know where you’re going yet.
gwern might open with a single experiment result. scott alexander opens with a patient.
2. state the naive view. steelman it.
what does everyone currently believe? why does it seem reasonable? don’t strawman it. make it as strong as possible. this earns the reader’s trust — they feel you’re being fair.
3. find the crack.
one thing that doesn’t fit the naive view. not ten things. one clean anomaly.
4. go wide across disciplines. this is the long middle.
each section brings in a domain that illuminates the same pattern from a different angle. the reader starts to feel: this same pattern keeps appearing everywhere. that’s not coincidence.
5. the synthesis. state the actual claim.
only now do you say it plainly. you’ve earned it.
6. implications. follow them honestly.
what follows if this is true? don’t hedge into meaninglessness.
7. open questions. don’t close everything.
the best essays end with sharper questions than they started with. what would need to be true for you to be wrong?
write in first person. “i noticed.” “i was confused by.” not “one observes.”
no jargon without immediate definition.
footnotes for genuine tangents you can’t bear to cut. not for citations.
section headers should themselves be interesting. not “background.”
long paragraphs when the idea needs density. single sentences for the thing that lands.
never summarize what you just said. trust the reader.
he treats the essay as genuine inquiry, not as argument.
the reader feels they’re watching someone think, not being persuaded.
that’s the hardest thing to fake and the most important thing to get right.