---
subtitle: or, views from my ascent of the shoulders of a giant tower of
  people
title: Far be it from me
---

Writing is thinking. There is a lot to think about, therefore there is a
lot to write about. Here is a collection of writings reflecting
snapshots of my progress in thinking.

While I aim for readability, I don't always aim to be read. Beware that
your time [might](https://thezvi.wordpress.com/)
[be](https://vitalik.eth.limo/index.html) [better](https://gwern.net/)
[spent](https://www.astralcodexten.com/)
[elsewhere](https://www.lesswrong.com/posts/xg3hXCYQPJkwHyik2/the-best-textbooks-on-every-subject).
If you do spend time here, I welcome comments/questions via email.

You can subscribe to new posts with the [Atom feed](/feed). "Planned"
essays are not promised.

## Essays on technology and civilisation, especially AI risk

> For thousands of years men dreamed of pacts with demons. Only now are
> such things possible.---Michèle of the Turing Police, *Neuromancer*

Risk models:

-   Simplicity and convergent instrumentality *(planned)*
-   Initial thoughts on agency *(planned)*
-   The value ambiguity bottleneck *(planned)*
-   Notes on alignment terminology *(planned)*

Move fast, break everything:

-   Two kinds of engineering *(planned)*
-   Actually aiming at alignment *(planned)*
-   Compute and sample efficiency ``{=html} *(planned)*

Research visions:

-   Whence mechanism? *(planned)*
-   Structure and scale *(planned)*
-   What is the effective objective? *(planned)*
-   Model enumeration *(planned)*
-   What are high-level structural properties? *(planned)*
-   The Honmoon agenda *(planned)* ``{=html}

Guest lecture:

-   [Ethics and the future of intelligence](future-ethics-lecture)
    *(2024.05.23)*

Open letters:

-   [Contra Zuckerberg on 'Open Source AI'](zuckerberg) *(2024.07.26)*
-   [Safe and responsible AI in Australia](responsibility-letter)
    *(2023.07.28)*

Fiction:

-   [One size fits all](one-size-fits-all) *(2026.01.02)*

## Notes on the theory of computation and learning

Goal misgeneralisation:

-   [Instrumental/intrinsic value ambiguity](value-ambiguity)
    *(2025.07.25)*
-   Geometry of goal misgeneralisation *(planned)*

Modelling computation:

-   [Turing trees](turing-trees) *(2025.06.15)*
-   Kolmogorov complexity via Turing trees *(planned)*
-   [What is complexity?](complexity) *(2026.03.16)*
-   Embedded computation *(planned)*
-   Cheating Kolmogorov, or, hacking the teacher's password *(planned)*

Epiplexity:

-   [Understanding epiplexity](epiplexity) *(2026.07.08)*

Singular learning theory:

-   [Blowing up](blowing-up) *(2025.08.10)*
-   [Schwartz distributions](schwartz-distributions) *(2025.08.17)*

Transformers:

-   [Transformers are universal in-context function
    approximators](universal-icl) *(2024.12.05)*
-   [Attention: All you need to know](attention) *(2025.03.29)*
-   [Transformers: Perceptrons in disguise](perceptrons-in-disguise)
    *(2026.03.05)*

Bayesian inference:

-   [The tempered posterior via the principle of maximum
    entropy](tempered-posterior) *(2026.02.27)* ``{=html}
-   Kullback--Leibler divergence, or, there and back again *(planned)*
-   [Variational inference but not via the evidence lower
    bound](variational-inference) *(2026.06.30)*
-   Don't get fooled by the reparameterisation trick *(planned)*
    ``{=html}

Decision theory:

-   [Minimax, three ways](minimax-three-ways) *(2025.02.23)*
-   [Approximate minimax, three ways](approximinimax-three-ways)
    *(2025.02.25)*

Expectiles are underrated:

-   [Expectiles are configurably-optimistic expectations](expectiles)
    *(2025.06.11)*
-   [Expectiles are simple to compute](computing-expectiles)
    *(2025.06.11)*

## What passes for my research blog

Goals for my graduate studies:

-   [Balanced academic orbit](balanced-academic-orbit) *(2025.05.25)*
-   [Subscribing to arXiv](subscribing-to-arxiv) *(2026.06.29)*
-   [Time to write](time-to-write) *(2026.07.26)*

Software taste:

-   [Switching to JAX](jax) *(2026.03.01)*

Research journal:

-   [Seven Saturdays with singular learning
    theory](seven-saturdays-with-slt) *(2023.07--unfinished)*
-   [Volumetric analysis of ICLR 2025 review process](iclr2025-volume)
    *(2024.12.08)*
-   [Freeing 118 GB of W&B experiment data from a broken binary
    format](free-wandb) *(2025.12.29)*

Quotation collections:

-   [On how to choose what to work on](quotes-on-what-to-work-on)
    *(2026.03.16)*

## About the author

Rules I try to live by:

-   [Smile!](smile) *(2025.07.15)*
-   Fear of missing out *(planned)*

What I have been reading (and select detailed book summaries):

-   [Generalised readings, 2019--2025](readings-2019-to-2025)
    *(2026.02.01)*
-   [Generalised readings, 2026](readings-2026) *(2026.01.31--)*
    -   [*Replacing Guilt* by Nate Soares](replacing-guilt)
        *(2026.03.07)*
    -   [*What is Life?* by Erwin Schrödinger](what-is-life)
        *(2026.03.20)*
    -   *Why Greatness Cannot be Planned* by Stanley and Lehman
        *(planned)*