Mid-Inkhaven Review
taking stock + what's next
Contents:
Preamble
Hey guys,
It’s currently 10:35 pm on November 13th. I’m writing to you from Lighthaven. In fact, I’m using voice dictation. I’m using an app called Super Whisper to dictate this into my laptop.
The reason I am dictating to you is because I would like to spend substantial portions of the next month writing code. In particular, I would like to try using Tinker from Thinking Machines to do some fine-tuning. However, it is very difficult for me to find time for this when I am under pressure to write 500 words each day.
Taking Stock
I am very happy about the more technical and object-level posts I have written. “The K-values of Speculative Kinship” is probably my favorite piece.
The K-values of Speculative Kinship
Sometimes you love someone so much they feel like family, though you don’t share any known ancestors and it’s by no means clear you’ll co-parent with them. No matter what happens in this generation, as long as everyone has children, you’re probably family via mutual descendants.
I am also very happy with “Defending the Data Processing Inequality” and “The Bitter Lesson: Six Years On”.
The Bitter Lesson: Six Years On
The Bitter Lesson only becomes legible in hindsight, when a single paper erases an entire decade of handcrafted work.
I have gotten enough positive feedback on “Two Types of Homeostasis” and “Should AI companies become profit-maximizing firms?” that I am now proud of these two.
Should AI companies become profit-maximizing firms?
Today I was fortunate to hear from a law student who works on making AI companies internalize negative externalities.
“Lit review: Stated vs. Revealed preferences in LMs” and “The Non-Perishable Agenda” feel like important reference pieces.
Lit Review: Stated vs. Revealed Preferences in LMs
If you want to know what I’m working on, skip to ‘What’s next’.
What’s Next
In terms of what’s next—I have transcribed 20 Neurotech Pub podcast episodes using OpenAI Whisper and have been reading through these to answer various questions that I have about neurotechnology and neuromodulation. Neurotechnology seems like a field where tacit knowledge is extremely important. So I do think that reading 40 hours of podcast transcripts might be one of the best ways to understand the frontier.
However, it takes a lot of effort to turn this material into posts. I also want to write about state space models, but there are several papers I need to read to make that happen. I have also promised a series on the Bitter Lesson, and I am very enthusiastic about continuing this series, but these posts take time.
I am likely to collaborate with condu.it to write about their thought-to-text system. I am very excited to let the world know about their work!
There are, of course, more casual posts that I can write. I have been considering writing “curius.app: An Enthusiast’s Review”. There are also the posts on my Manifold Market. It feels as though I have been mining my backlog of conceptually simple posts over the past few days, and that things will only get more difficult as the month goes on. This is why this evening, I am attempting to dictate 17 posts so that I have enough backup posts to cover the rest of Inkhaven, and can participate in freer intellectual exploration across the rest of the month. Thank you for bearing with me!
Camera roll to hit word count
