Reading list
The shelf behind the positions.
This isn’t a canon. Every entry is load-bearing: a position, a post, or a working rule on this site leans on it. Each one says what I take from it and where I use it. If a book stops doing work, it comes off.
Last updated:
Kant to Idealism to Pragmatism
The spine of how I think. Kant draws the limits and the floor. Hegel says you can’t check the instrument from outside it. The pragmatists make the test practical: what difference does it make?
-
Critique of Pure Reason
Immanuel Kant · 1781, revised 1787
Examine the instrument before you trust what it says. A benchmark knows a model only under its own conditions: task mix, format, scoring rule. A claim that reaches past those conditions isn’t false so much as ungrounded.
-
Groundwork of the Metaphysics of Morals
Immanuel Kant · 1785
Two things I don’t trade away. Persons are ends, never only means, which is position P-01. And the rule-breaker’s tell: you don’t will that the rule should fail, you want an exception for yourself, “just for this once.” That is exactly what an agent does when it grants its own permission.
Used in An agent that answers its own permission question hasn’t asked it and Positions
-
Phenomenology of Spirit, Introduction
G. W. F. Hegel · 1807 · §73
The best objection to my Kant entry. If knowing is an instrument or a medium, you can’t subtract its distortion without already knowing the thing. For evals that means there is no instrument-free audit. Critique has to be ongoing and empirical: run it, compare against deployment, redraw the boundary.
-
“The Fixation of Belief” and “How to Make Our Ideas Clear”
Charles S. Peirce · Popular Science Monthly, Nov 1877 and Jan 1878
The first paper ranks the ways people settle belief: tenacity, authority, the a priori method, and the method of science. Only the last lets the world push back. The second gives the test I use on every score: a claim means the practical effects it predicts. A gap on a leaderboard that predicts nothing you’d observe is decoration.
-
Pragmatism
William James · 1907 · Lecture II
James made Peirce’s principle famous and also loosened it: a difference that makes no difference elsewhere is no difference. I keep his slogan and Peirce’s rigor. For evals, “works” has to mean an observable effect someone could check, not that the belief is satisfying to hold.
-
The Quest for Certainty
John Dewey · 1929 · Gifford Lectures
Dewey’s case against philosophy’s hunt for theoretical certainty, with knowing tied back to action. The oversight version: stop asking for a certificate that a system is safe, and ask which inquiry keeps running after launch and who can act on what it finds.
Behind P-02 and P-05
Argument and judgment
Tools for checking a claim, mine included, before it goes out.
-
The Uses of Argument
Stephen Toulmin · 1958
Data, claim, warrant, backing, rebuttal. Most bad AI claims fail at the warrant, the step from evidence to conclusion that nobody wrote down. When I ask “what’s the test?”, I’m asking for the backing.
Used in An agent that answers its own permission question hasn’t asked it and A steelman the other person wouldn’t sign is an iron man
-
The Art of Controversy
Arthur Schopenhauer · translated by T. Bailey Saunders, 1896
A catalog of ways to win an argument without being right. I read it as a checklist of what to catch, and the first place I look for those moves is my own drafts.
-
Intuition Pumps and Other Tools for Thinking
Daniel C. Dennett · 2013 · “Rapoport’s Rules”
Restate the other side so well they wish they’d put it that way, then criticize. That gives a steelman a pass condition: the person you’re arguing with would sign it. If they wouldn’t, you built your own opponent.
Used in A steelman the other person wouldn’t sign is an iron man
-
Psychology of Intelligence Analysis
Richards J. Heuer Jr. · CIA Center for the Study of Intelligence, 1999 · ch. 8
Analysis of Competing Hypotheses: put every live explanation in one matrix and try to knock them down instead of propping up your favorite. Evidence that fits every hypothesis tells you nothing. Most of what gets cited for a model’s “reasoning” is that kind of evidence.
A working rule for how I weigh evidence
Systems, measurement, and AI
Where the philosophy becomes engineering: controls, feedback, and what happens to a number once someone aims at it.
-
Goodhart’s law, and the line everyone quotes
Charles Goodhart, “Problems of Monetary Management: The U.K. Experience” · 1975 · and Marilyn Strathern, “‘Improving ratings’: audit in the British University system,” European Review · 1997
Goodhart: a statistical regularity tends to collapse once you lean on it for control. Strathern’s paper gave it the phrasing everyone quotes: a measure that becomes a target stops being a good measure. Agents make it faster and more literal, because an agent may not represent the difference between the metric and the goal at all.
Used in A pass threshold without a date is a mood and Agents can’t see what wasn’t measured
-
Engineering a Safer World
Nancy G. Leveson · MIT Press, 2012 · open access
STAMP treats accidents as failures of control, not chains of broken parts. That’s the engineering form of P-02: name the controller, the feedback it gets, and the constraint it enforces. If you can’t, there is no control, only a component that hasn’t failed yet.
Behind P-02 and P-03
-
Thinking in Systems: A Primer
Donella H. Meadows · edited by Diana Wright · 2008
Stocks, flows, feedback loops, delays. Behavior comes out of structure, so you learn a system by pushing on it and watching it respond, not from a snapshot. That’s my “know by loops” rule, and it’s why after-the-fact logs aren’t containment.
Behind P-05 and Philosophies
-
“Concrete Problems in AI Safety”
Dario Amodei, Chris Olah, Jacob Steinhardt, Paul Christiano, John Schulman, Dan Mané · 2016
Five practical failure modes, and reward hacking is the one I keep coming back to. It’s Goodhart with an optimizer attached. The fix belongs in how you measure and supervise, not in hoping the model means well.
Read with Agents can’t see what wasn’t measured, the agent version of the same problem
Side threads: Stoic and Daoist
Not the spine. They keep the spine from turning into theater.
-
The Enchiridion
Epictetus · recorded by his student Arrian · ch. 1
The book opens by splitting things into what is up to us and what isn’t. Sort first, then spend effort only on the first pile. Most of the drama around a launch belongs in the second.
Behind “Composure over theater”
-
Daodejing
Attributed to Laozi
Wu wei: act with the grain, don’t force. I take it with one limit attached. When a duty needs a gate forced, I force it. Going with the grain is the default, not a rule against stepping in.
Behind “Composure over theater”
Think something’s missing, or misread?
Name the entry and the passage. If I’ve leaned on a book for something it doesn’t say, that’s a correction I want. Tell me, or argue with the posts on the blog.