Environments for thinking
Hi friends,
(welcome to my email dispatch! You can sign up for these or read the archive at buttondown.email/thesephist 💌)
What I read
I’ve been trying to get back into a regular habit of reading / listening to books. I listened to Gladwell’s The Bomber Mafia yesterday, which is written so compellingly and clearly even a relatively amateur to history like me could appreciate the story. Unlike Malcolm Gladwell’s usual books, this one is a more true-to-history documentarian book and less a drawn-out ted talk, and it’s quite short! I recommend it.
I also started Neal Stephenson’s Seveneves, which opens with one of my favorite first sentences in any novel,
The moon blew up without warning and for no apparent reason.
Bill Gates also has some nice things to say about it. I’m generally a fan of hard science fiction set in space, but in the past have found Stephenson’s writing a bit drawn out. We’ll see if I feel differently this time around.
Besides books, I also read and thought a lot about creating good environments for generative, creative, frontier work:
Michael Nielson’s Principles of Effective Research, which was my favorite
John Schulman's An Opinionated Guide to ML Research
I also skimmed Richard Hamming’s “You and Your Research”
Relatedly, environments for learning and discovery:
Andy Matuschak’s Exorcising us of the Primer
Bret Victor’s Communal computing for 21st century science, whose PDF copy I can’t locate online right now but is available in talk and Substack form here. (If you’d like my full PDF, reply and I can send it to you.)
All of this has me thinking a lot about communal and organization containers in which great research work is brought into the world, and brought to bear on hard problems at scale. Ambition, community, clarity of vision, technical taste, elevated by better tools. It pushed me to think about how to better shape my own enabling environment and container for my own work.
I even wrote about similar topics 3 years ago, in Towards a research community for better thinking tools, though I was quite naive then and I think I’ve learned a lot more now.
This passage from Richard Hamming’s talk stayed with me:
I have to get you to drop modesty and say to yourself, "Yes, I would like to do first-class work." Our society frowns on people who set out to do really good work. You're not supposed to; luck is supposed to descend on you, and you do great things by chance. Well, that's a kind of dumb thing to say. I say, why shouldn't you set out to do something significant? You don't have to tell other people, but shouldn't you say to yourself, "Yes, I would like to do something significant."
A public service announcement 📣
Val Town is looking for a DevRel! You’ll be creating new tools, libraries, ideas, and building blocks on Val Town and thinking about how to build new kinds of software. They’re some of my favorite hackers in NYC. If this sounds like you, you can reach out on their Notion page.
What I’m working on
I wrote three pieces since my last email:
Synthesizer for thought — For most of the history of music, humans produced sounds out of natural materials — rubbing together strings, hitting things, blowing air — until we improved our understanding of the physics of sound, and built devices to record sound and decompose it into its constituent fundamental parts using the mathematical model of waves. Then, we could imagine new kinds of sounds as mathematical constructs, and then conjure them into reality. Not only that, backed by our mathematical model of sound, we could systematically explore the space of possible sounds and filters. The instrument that results is the synthesizer. Could machine learning models do something similar for ideas?
The quotes on my wall, about the collection of screenshots of quotes that fill my desktop wallpaper.
Epistemic calibration and searching the space of truth, about the "ChatGPT voice", "Midjourney style", why they happen, and some ways out that I can imagine. Lexica’s Sharif Shameem generously called it a “must read” if you’re working on tuning base models. Thanks Sharif!
I’m also still chipping away at some research directions:
I’ve started to sketch out what some concrete interfaces for better reading, searching, and understanding large scale knowledge may look like based on my work on latent space based interfaces. I’ll be showing it at an upcoming Demo Night in NYC, and writing about it soon. If you’re in the city, come out and say hello!
I’m still training some improved versions of models for future experiments, though it’s been a bit stalled by some infrastructure issues…
I’ve been very closely following my peers’ work on applying ideas from Prism to other modalities like images. Gytis and Yondon have been sharing some very neat demos.
A reminder: you can reply to these emails :-) They go straight to my inbox.
Wishing you a happy and safe week ahead,
— Linus