I’m Gavin Leech, a PhD candidate in AI at Bristol. I like technical solutions to philosophical problems.

Technically an EA blog, though I don’t consider this a blog. (Blogging is more speech than writing.)

If you want to talk, get in touch by email, or anonymously, or extremely privately via PGP.

Research

I'm interested in keeping an eye on AI. By sheer good fortune my advisor is Laurence Aitchison. Pubs and code and teaching here.

I'm working on:

  • Tensors for fast sampling in big latent variable models.

  • Emotion-like behaviour in ordinary RL agents.

  • Better ways to communicate preferences to RL agents.

Change my mind

I offer bounties for errors:

  • $1 for nitpicks (ambiguities, misreadings, exceptions, technicalities).
  • $10 for an error which falsifies a sentence.
  • $50 for an error which invalidates a post.

I suppose this should all be multiplied by the confidence level of the post. If no confidence level is set, then "100%". Factual and logical errors are most likely to resolve cleanly; I probably won't cough up if you just tell me I'm an idiot for believing or not believing in e.g. Meinongianism. That would get a smile.

I'm not sure how to handle claims that have changed truth-value since I made them; half of this blog might be false if you return in 30 years. Leaning towards $1 to $5.

Also, if I make a claim you want to bet against, get in touch for odds.

Payouts

Page Error Person Payout
About Was unclear about the required size of error for a bounty, leading to a prima facie contradiction. Loki Sempere $10 to EA Funds
/psych Overstated the consensus against Milgram. It's still shaky, and not really about obedience, but it's wrong to say there's no evidence, just that the original experiments were terrible. Matt Edwards $5
/gists I said "constant time" instead of "constant time complexity". The function in question takes 3x more time to process 2^64 than it does 2^0, hence the intermediate payout between nitpick and falsehood. Loki Sempere $5

Bets

Bet Bettor Odds Resolution
Jan 2016: Trump finishes term. CR 1:1. Nice bottle of whisky. To me.
June 2020: Biden wins election. CR 1:1. Nice restaurant dinner. To me.

Giving

I'm a member of Giving What We Can, people who pledge substantial amounts of their lifetime income to the most effective charitable causes. I'm donating 5% this year, aiming at 50% overall 2. My reasoning on the most important causes, and my present basket of charities is forthcoming.


I am cause-neutral (in the sense that I will support whatever is highest value, to the extent that objective evaluation is possible) and cause-sceptical (in the sense that we don't really know what the best thing to do is).

Books

I try to review every book I read.

Most book reviews are vacuous, blocklike expressions of approval or disapproval. There is nothing in them except the reviewer, their squeals. I do some of those, but I try to do some of the other kind too.

Currently reading

Favourite books

Title
Air Guitar: Essays on Art and Democracy (1997) by Dave Hickey
Tell Me No Lies: Investigative Journalism and its Triumphs (2004) by John Pilger
Hitler's Uranium Club: The Secret Recordings at Farm Hall (1995) by Jeremy Bernstein
Selected Essays of Michel de Montaigne (1592) by Michel de Montaigne
Bad Science (2008) by Ben Goldacre
Meditations (180) by Marcus Aurelius
The Black Swan: The Impact of the Highly Improbable (2007) by Nassim Nicholas Taleb
Whereabouts: Notes on Being a Foreigner (1987) by Alistair Reid
The Oxford Companion to Philosophy (1995) by Ted Honderich
How to Actually Change Your Mind (Rationality: From AI to Zombies) (2018) by Eliezer Yudkowsky
Ficciones (1944) by Jorge Luis Borges
Oxford Book of Essays (1991) by John Gross
Tractatus Logico-Philosophicus (1921) by Ludwig Wittgenstein
A Supposedly Fun Thing I'll Never Do Again: Essays and Arguments (1997) by David Foster Wallace
Rationality: From AI to Zombies (2015) by Eliezer Yudkowsky
What If? Serious Scientific Answers to Absurd Hypothetical Questions (2014) by Randall Munroe
Stories of Your Life and Others (2002) by Ted Chiang
The Will to Battle (Terra Ignota, #3) (2017) by Ada Palmer
The Fable of the Dragon-Tyrant (2005) by Nick Bostrom
The Hedonistic Imperative (2015) by David Pearce
The Age of Em: Work, Love and Life When Robots Rule the Earth (2016) by Robin Hanson
Factfulness: Ten Reasons We're Wrong About the World – and Why Things Are Better Than You Think (2018) by Hans Rosling
Mortal Questions (1979) by Thomas Nagel
Doing Good Better: How Effective Altruism Can Help You Help Others, Do Work that Matters, and Make Smarter Choices about Giving Back (2015) by William MacAskill
Library of Scott Alexandria (2015) by Scott Alexander
Famine, Affluence, and Morality (1972) by Peter Singer
Save Yourself, Mammal!: A Saturday Morning Breakfast Cereal Collection (2011) by Zach Weinersmith
Incerto 4-Book Bundle: Antifragile, The Black Swan, Fooled by Randomness, The Bed of Procrustes (2011) by Nassim Nicholas Taleb
The God That Failed (1949) by Richard Crossman
Brewer's Rogues, Villains & Eccentrics: An A-Z of Roguish Britons Through the Ages (2002) by William Donaldson
Guards! Guards! (Discworld, #8; City Watch #1) (1989) by Terry Pratchett
The New Penguin Book of Scottish Verse by Robert Crawford
What Matters Most is How Well You Walk Through the Fire (1999) by Charles Bukowski
Tales from Earthsea (Earthsea Cycle, #5) (2001) by Ursula K. Le Guin
Map and Territory by Eliezer Yudkowsky
Computing machinery and intelligence (1950) by Alan Turing
Watchmen (1987) by Alan Moore
Beyond Good and Evil (1886) by Friedrich Nietzsche
The Penguin Book of English Verse (2000) by Paul Keegan
Lolita (1955) by Vladimir Nabokov
Psychotic Reactions And Carburetor Dung (1987) by Lester Bangs
Blindsight (Firefall, #1) (2006) by Peter Watts
Small Gods (Discworld, #13) (1992) by Terry Pratchett
Brief Interviews with Hideous Men (1999) by David Foster Wallace
Right Ho, Jeeves (Jeeves, #6) (1934) by P.G. Wodehouse
The Book of Disquiet (1982) by Fernando Pessoa
The Patrick Melrose Novels (2012) by Edward St. Aubyn
Bartleby the Scrivener (1853) by Herman Melville
Infinite Jest (1996) by David Foster Wallace
The Elephant in the Brain: Hidden Motives in Everyday Life (2017) by Kevin Simler
Blood Meridian, or the Evening Redness in the West (1985) by Cormac McCarthy
The Wind's Twelve Quarters (1975) by Ursula K. Le Guin
Poems of the Late T'ang (1965) by A.C. Graham
What Should We Be Worried About? Real Scenarios That Keep Scientists Up at Night (2014) by John Brockman
Iain Crichton Smith: Selected Poems (1986) by Iain Crichton Smith
Collected Poems (1988) by Czesław Miłosz
Conquest of the Useless: Reflections from the Making of Fitzcarraldo (2004) by Werner Herzog
Cultural Amnesia: Necessary Memories from History and the Arts (2007) by Clive James
Labyrinths: Selected Stories and Other Writings (1962) by Jorge Luis Borges
Reliable Essays: The Best of Clive James (2001) by Clive James
Hyperbole and a Half: Unfortunate Situations, Flawed Coping Mechanisms, Mayhem, and Other Things That Happened (2013) by Allie Brosh
Travels With Myself and Another (1979) by Martha Gellhorn
Collected Poems (1988) by Philip Larkin

Misc

Epigram


It is equally fatal to the spirit to have a system and to not have one.
So you must combine the two. 1
— Friedrich Schlegel


...there is in men and women a motivation stronger even than love or hatred or fear. It is that of being interested — in a body of knowledge, in a problem, in a hobby, in tomorrow’s newspaper.
— George Steiner


Favourites

Site

I love making this site better. I'm no designer but I enjoy the results, and the feedback (n=6 friends) is good. Here's a list of my design decisions.

The title is an Iain Banks allusion: his powerful protagonists tease pompous people by taking names with as little gravitas as possible. To take the arg min of something is to find the thing which results in its lowest value.

Conflicts of interest

Financial:
  • I am funded by UKRI and so really by the British government.
  • I won an Emergent Ventures grant via Tyler Cowen.
  • ex: I used to work at AXA, a big insurer.
  • My mask paper used compute donated by Microsoft.

Political: Few. Never a member of any party.

Psychology

Morality as Co-operation (questionnaire): (Really not a fan of evolved morality.)


Oxford Utilitarianism Scale
  • Impartial Beneficence: 33 out of 35
  • Instrumental Harm: 21 out of 28


Intrinsic Values Test


Moral Foundations:


The Big Five



Myers-Briggs Type: FOAD.

Coding

I've been employed to code for 7.5 years 5. Of which, I've been employed in machine learning for 4.9 years.

It is some career. So far I have worked on a precision methadone pump, a website generator for bed & breakfasts, secure off-site credit card storage, procurement streamlining for e.g. 'aerospace' companies, a metre-precise 3D model of Britain's rivers, a performance dashboard for use by secondary school teachers, medical risk scoring models, countermeasures against machine-learning model extraction...

Experience with

Antagonism between language communities is a really clear example of how strong identities cause perversities.

I have written production code in
  • Python (raw, SciPy/Pandas, PySpark): 9
  • C# 11
  • Typescript 12
  • nodeJS 13
  • Pig Latin 14
  • bash 15
  • Oozie XML 16
  • PHP 17
I also have side projects in
  • Java : All my Master's degree work: Game of Life, Vigenere cipher, .
  • Scala : Just Project Euler so far.
  • Rust : Just Project Euler so far.
  • Powershell: e.g. a web scraper (before Python was provisioned).
  • bash : An archiver for hyperlinks.
  • Jekyll : this itself.

Frameworks, tools, stacks, hardware

Stored as JSON here.

Elsewhere

Endorsements


DQ: A very nice website - links to the original papers, including participant numbers and number of citations... effect sizes and replications effect sizes.
JH: Very interesting, for a variety of reasons - and not just because it's fun to trawl through the wreckage... of 'those silly social scientists'. This [post] is a big old thing... much more comprehensive than anything I ever did... Not a lot of people would have the commitment to even try and organise something like this... You do get a sense of the enormity of [the replication crisis] which isn't really available through other documents...

DQ: What are the implications for how we actually teach first year psych?
JH: TEACH THEM THIS!
- Dan Quintana & James Heathers


The best mask-wearing study so far?
- Tyler Cowen


This was a great EAG. I love how excited everyone was, like this guy. (If you're in the audience, sorry Gavin, didn't get your permission [to show that photo]).
– Will MacAskill


Utterly charming. In the nature of such lists, one does not expect to agree with every, single one---of course not. But the lists exhibit a lively and critical mind, unwilling to follow the usual party lines.
– Deirdre McCloskey 6


I'm honoured to be on your list [of favourite books].
– Robin Hanson 7


Nice image, young man. You may be a writer.
– Will Self 8



Also, my mum says I'm cool.

Identity

I try not to identify as things. My premises, projects and preferences say overwhelmingly more about me than my social identity does.

Tone

I'm sometimes told I'm too strident. I'll accept this, because constant hedging is unreadable, more annoying than authoritativeness. Also because strong claims galvanise opposition: as long as they're reasonable as well as strong, hopefully this stimulus creates value.

Anyway I'm not as sure as I sound: there's a comment box under every piece, call me out.

Why so many quotations?

Well, they're the densest store of insight. We are not empiricists by nature or by culture, so they are my concession to the apeish narrative side of mind. That quotes page is like a message digest of 10 years of me reading and learning.

I sometimes regret my instinctual distrust of Twitter: a lot of contemporary aphorists must be whiling away the years on there.

Metrics

Here are scales for rating all kinds of things.

Loss function

Above all I hope to beat the trend spotted by Robin Hanson about independent scholars:

over time amateurs blow their lead by focusing less and relying on easier, more direct methods. They rely more on informal conversation as analysis method, they prefer personal connections over open competitions in choosing people, and they rely more on a perceived consensus among a smaller group of fellow enthusiasts. As a result, their contributions just don’t appeal as widely or as long.

Gavins Leeches

Illnesses

I suffer from degenerative DNA scrambling, increasingly lazy cells, irreversible methylation, terribly tangled proteins, disappearing stem-cell nooks, and ultimately oxygen poisoning, like everyone.

What's with all these accordions?

Information hiding is a basic condition of good large software; these accordions are a prose analogue. The hope is that, by tucking away wires, I can write long things without triggering anyone's TL;DR reflex.

It also sort of simulates a dialogue with the reader: I plant a question, and then they prompt for the response.
  1. Es ist gleich tödlich für den Geist, ein System zu haben, und keins zu haben. Es wird sich also wohl entschließen müssen beides zu verbinden.
  2. (or an equivalent value of my labour.)
  3. If you count SAS. (If you do, I was employed to code before I learned to Code.)

    Note that just being 'employed to code' says surprisingly little about your ability. And almost nothing about your possession of the spiritual advantages of technical life.
  5. Comment on my "vade mecum" book list including Age of Em.
  6. Here, 21/11/14.
  7. e.g. a computational linguistics wrapper
    e.g. a full notebooked ML pipeline, from sparse raw csvs to GBM to policy implications.
  8. e.g. An ironclad medicine pump controller with WPF.
  9. e.g. A CRUD procurement app with Angular and .NET.
  10. e.g. a secure pharmacy record ("Rx") system.
  11. e.g. Various YARN batch jobs, processing a few TB then emailing out regular tabular summaries.
  12. e.g. Debugging HDFS schema issues with Avrotools.
  13. e.g. chaining Shell and Pig scripts to crunch and mail reports.
  14. e.g. fairly complex websites, Laravel prototyping, a cURL handler for a credit card storage API.