Things I would do differently: Education

I am a remarkably unreflective person. I go months at a time without thinking about my past, or the people I once knew. And not because my past's fucked up or anything; just because the present and the far, far future crowd it out. I'm pretty happy with this arrangement.

Recently, though, I've realised some easy things I could have done to be a better writer / scholar / researcher as of 2017. (They are hardly tragedies though, just inefficiencies.)

1. Physics

Picking courses as a 17 year old in a country without tuition fees, I latched on to the most obvious sources of meaning: philosophy, music, literature. But I could have gotten into physics or stats or computer science if I'd applied (I did get in for biology). And these would serve my present purposes much more, because I'm aiming at truth, and these latter are our greatest machineries of truth.

I don't regret my MA. (Though I probably would if I were English.) Formal philosophical study - that is, seeing what knots and messes the greatest minds in history have tied themselves into, working off no data - has probably saved me from some errors people make when they slip into metaphysics unawares.

And it has probably made me less overconfident that the world can be solved by pure, solitary thought. ("The penalty for not doing philosophy is giving bad philosophical arguments a free pass.") And I have a thick layer of protective scepticism about macroeconomics.

But I would have read philosophy and poetry anyway - I have a great appetite for them, and had it before I got institutional grounding - and so would have gotten much of the inoculation against bad philosophy and the realisation of the relative shallowness of great artists even had I done something harder.

As it is, I've been scrabbling to piece together an education in scientific modelling ever since graduating, and it has taken ages on my own, and I am quite sure that I did this backwards. (Needless to say, the average 2010 economics curriculum was not scientific enough to count.)

But ooh. "Inoculation against bad philosophy and bad economics": is this is the most positive case I can make for my classes? Yes but never mind classes: the greater part of the value came from having 4 years to straighten out my head, and a hundred wonderful people from over the world to collide with, brighter than anyone I'd known before. But again, I'd probably have found them as a physics boy; it was a small university, and my nature is not so malleable.

The distinctive value of an arts degree - that it draws creative misfitting people, that it's low-intensity enough for you to have many projects and loves without constant stress, that it permanently demystifies the baroque, ridiculous world of high culture - are wonderful, but I think I'd rather know how the world works, on balance.

2. Code everything

After my arts degree I switched into software development, a viscerally satisfying career to me. Not just talk, not just interpretation: but fucking building things.

But as well as a fun career, code is an incredible way of expressing thought. You get an oracle, the compiler, tell you if it could possibly be true.

See, coding is a novel way of thinking in general. Yes, it is like maths - but testable, causal, interactive.

A programming language is "how you tell a computer what to do". But before that it's a way to express ideas and get push back from a rational oracle. (It's not reality that's pushing back, of course. You don't know if they're true, but you know if they are clear, if they could even possibly be true, if you are not completely fooling yourself.)

Consider the Bible, or Karl Marx's work, or Sigmund Freud's work. These are rammed full of invalid and unsound ideas - but they are beautiful, unified, and powerful, so they proved persuasive to billions of people. Human language offers no easy test of consistency, no way of really precisely connecting idea to idea. We have had only hard, piecemeal, irreplicable interpretation.

To see what's added by code, here's a thought experiment: Imagine the economic value of a line-by-line description, in English, of the Linux kernel. It would be nothing compared to the billions of dollars of value the kernel has created or saved.

The computability of source code is a side effect of its clarity. Code is testable thought.

I'm converting my maths notes into Python as a matter of urgency, because standard Mathscript is not good. I don't know why this took 2 years to occur to me; clearly the claws of the arts run deep.

This macroeconomics course, in Python and Julia, has crystallised a host of things I only mechanically learned before.

In philosophy, it would have let me get into the thriving and objectively progressive research programmes: philosophy of information, logics, cellular automata, and so on. Here are two great examples of coded philosophy, as proof of concept.

3. Use the blogospheres.

I have learned more about economics from reading Hanson, Quiggin, Krugman, Caplan, Dillow, Friedman, deLong, Harford, Cowen, Sumner, and Smith, than I did in two full years' worth of lectures at Aberdeen. Which is strange, because most of them are academics. But, because their readers are from broad backgrounds, the writing is vastly superior to that of papers: clearer, briefer, and more easily evaluated for both rigour and well-foundedness. In 2010 the econ 'sphere wasn't as highly developed as it is now, but was still good enough.

In stats, Andrew Gelman, Uri Simonsohn, and Cosma Shalizi's blogs have taught me what's wrong with science and how to fix it, which I didn't get a jot of in classes.

(Philosophy and maths benefit less from this, because their usual texts are more digestible and more ineliminably systematic, respectively.)

This step wouldn't have improved my grades much, because of teaching-to-the-test.* But it would give me what universities are supposed to give: firm grounding in expert knowledge about things which matter, and the ability to apply it appropriately.

* A dark implication: that one could be better-off, in finances but also in knowledge, without uni altogether. (Since they distract you with password learning and rote crap.) We rely on the spiritual and psychological gains of 4 years of relative leisure. And at the micro level, this is a clear good deal.

4. Focus

Over the past 4 years, out of uni, I've read an average 102 books a year. They have been about everything, and it has been wonderful. A four-year cruise on about £300.

But I am persuaded that this isn't how you contribute to human knowledge. The absurdities of siloed scholarship - economists and anthropologists and sociologists and psychologists and all talking about the same thing, but wholly ignorant of each others' insights - are large, and can't be fixed except by people who own several hats. But everything else is done by specialists, because the coalfaces of knowledge are very far from common sense, in several different directions, and anyone who tries to reach several of them is likely to end up near where they started.

One of my resolutions this year is to read fewer than 25 books, but to make them all count. I have a folder, "Spoilers for Reality", with textbooks and serious crap to get through. (In each of those hundred-book years I was supposed to be studying maths, and you can imagine how much I actually did.)


notable mental methane vents

  • SOC (n.): system on a chip. Previously known as a microcontroller. What we now call just 'a computer', but integrated boards were an enormous deal, a revolution within the digital revolution.

  • UUOC (n.): Useless Use Of Cat (Award). Surprisingly mean retort to StackOverflow answers which use the UNIX tool cat where a pipe would do.

  • abience (n.): the urge to withdraw. Usually used to mean pathological avoidance, but to me it is also the plain, sacred joy of missing out.

  • hardtack (n): A very basic cracker, just baked flour and water. Staple of navies and Tudor explorers.

  • HARKing (v.): Hypothesizing After the Results are Known. A particular problem in social science, where pre-registration of studies is a tiny minority of work.

  • merchantable (UK legal n.): Good enough to be sold.

  • technical steer (n.): Input from expert staff, AKA 'knowledge'.

  • whitespace damage (n.): subtle but breaking changes to source code performed by ordinary text processors, e.g. line wrapping, hidden characters, odd apostrophes. This phrase is a shibboleth for being A Very Serious Person, e.g. a kernel dev.

  • moving up the value chain (phrase): performing work further away from physical extraction, processing, and manufacturing. Supposedly insulates you from competition because your outputs are less easily evaluated as they become less physical. Economic abstraction.

  • FANG: Facebook / Amazon / Netflix / Google, particuarly when their stocks are used as a bellwether.

  • DGP: data generating process. This took bloody ages to google.

  • rebranding (v.): "a euphemism for 'euphemism'" - Jonathan Meades

machines inside

PSA: It took me many years to internalise the formal methods I know now.*

I use "internalise" as distinct from "learn", because, let's face it: we all "learn" statistics in uni, in the sense of briefly knowing a tiny set of teacher passwords, of knowing what a mode is, and of knowing how to dumbly apply two canned tests of inference.** But almost no-one with that badge on their resume actually remembers, actually uses, and was actually changed by contact with it, the driest and most nutritious method.

My measure of internalisation is if you use the method, without prompting by school or advisor, in your investigations. Internalisation requires some understanding, but I'm not saying that I have any deep grasp of these things. I just appreciate their power, and use them as well as I can where I can.

  • First contact with algebra: 2000
    Internalised algebra: 2012.

  • First contact with Analysis: 2003
    Internalised Analysis: Not yet.

  • First contact with formal logic: 2008.
    Internalised first-order logic: 2010. (pic above)

  • First contact with proof: 2010
    Internalised proof: Not yet.

  • First contact with statistical inference: 2011 **
    Internalised statistical inference: 2017.

  • First contact with Bayesian / cognitive / decision science: 2010
    Internalised decision science: Not yet.

  • First contact with full-blown probability theory***: 2012
    Internalised probability theory: 2017.

  • First contact with (imperative) computational thinking: 2014
    Internalised computations: 2015.

  • First contact with functional programming: 2016
    Internalised computations: Not yet.

  • First contact with machine learning: 2015
    Internalised ML: Not yet.

* This strikes me as worth stating, because the rigorous fields are so demoralising to tackle alone, and take so long for even very intelligent people to get comfortable with. In a standard mathematical education, we don't get to see the cockups or the thousands of fruitless hours that Jacobi or Germain had to put in, to win as they won. (The painstaking labours of Wiles and Zhang are at least a bit more available to us.)

** I have tried to learn statistics (that is, higher statistics, data analysis and inference) four times in my short life:

  • 1) 2010: in the standard, cursory Research Methods module in undergraduate economics (I find myself not guilty).
  • 2) 2011: to catch up with discussions on LessWrong
  • 3) 2013: through a formal stats degree
  • 4) 2016: on the job.

Only the last is sticking.^ But I was extremely lucky to get a statistically demanding job without credentials in the first place; I snuck in because my profession confuses people and my programming ability was so far beyond the spec that they halo-effected me in.

So where does everyone else actually learn stats?

^ Again, I'm using a strict definition of "learn": in the others I learned many terms and followed many formal derivations and clicked many buttons in many statistical packages, but I did not actually gain the statistical mindset, modelling ability.

*** Not even measure theory...


efficient transcendence

  • Trust is efficient: more trust means less expenditure on vetting and surveillance. "Just give and see".
  • Honesty is efficient: more honesty, less expenditure maintaining a big diff web of lies. "Just say what you think".
  • Empathy is efficient: more perspective means a better comprehension of the actual situation, and so better outcomes for yourself and others. "Just imagine".
  • Keeping promises is efficient: enables future dealing. "Just do as you say".
  • Charity is efficient: able to address market failures, and able to extract the most value from cash. "Just get over yourself".

but also

  • Gullibility (misplaced trust) is costly. Lurid individual examples are easy to find, but more generally something more than 8% of the entire world economy is consumed by misappropriation. ($1tn corruption, $7tn fraud, $0.12tn shoplifting.)
  • Discoverability is costly: with millions of anonymous people potentially encountering you, the rarity of psychopathy and ideological violence isn't much protection.
  • Empathy is totally subject to our worst biases, moreso than reasoning.
  • Deontology can be wildly inefficient.


qualitative quantum paradigm disruption

  • "Qualitative difference": a change so obvious even unaided human perception can spot it.

But what about differences that are only obvious to people in the know? The difference between this
and this

is obscure to my parents, but clear to me, for instance.

A simple and I think honest answer is that knowledge is indeed a perceptual aid. Call qualitative differences which require knowledge to detect, Simonian differences. (This is after Herbert Simon's studies of the excellent acquired intuitions of e.g. firefighters and chess grandmasters.) Thus:

  • Nas vs Slightly Remastered Nas: quantitative
  • Nas vs Rakim: Simonian qualitative
  • Nas vs Big Pun : qualitative


the dust of life

How much of human psychology could aliens reconstruct from our art?

Let's say they have every work of fiction, but no documentaries, no datasets, nor even any archaeological digs to go on. Nothing uncontrived.

So, postmodern media would be very useful to them: as a proof of our self-consciousness, but also since it gives clear indication of what we regard as markers of fiction. A knowing wink to camera is data about fiction given within fiction. Anything which is spoofed, anything which a loud narrator makes reference to, is a trope and can thus help to file away convention in favour of mimesis.

But I think the answer is "surprisingly little". As Picasso says, "art washes off the dust of everyday life": i.e. it is wilfully unrepresentative and heightened, and not psychologically real. Even if we gave the aliens all good art ever, there would still be large and systematic inaccuracies.

The biggest missing things are repetition, our biological overheads, and inarticulacy. Also, watching our great tragedies, which are our greatest works don't you know, they would probably not infer how rare meaningful suffering actually is. To make the art hoard an accurate picture, we would need to include 1,000,000 copies of every mumblecore film, to balance out the unrealistic coincidence, existential transcendence, and wit.

My mate thinks that they'd learn a lot about us exactly via our deluded fiction: it would show our ideals and our prejudices, and that, since these are where we live in our heads, this goes a long way. And that is true, for a human reader scoping out another new group of humans. But you can't make inferences about delusions if you don't know the truth, and I see the aliens as quite likely to mistake mere tropes for real behaviour. For instance, 5% of films are horror genre. I don't know exactly what proportion of lives are horror lives, but it is somewhere south of 0.006%.

(One delusion they could easily skewer, starting from no knowledge, is the rate of coincidences our stories rely on. All you need is probability there.)

Another friend comments, after the initial question but before my further spraff:

If we found another civilized world, the greatest challenge would suddenly change from the difficulties of sailing the cosmos to the impossibility of understanding the literally alien.

I don't believe convergent evolution will apply in intelligent aliens: and so how could we understand a culture of [nonindividual hiveminded] mindworms? Aliens may well suck our atmosphere, drink the oceans and move on without any interest in our culture whatsoever.

So, respectfully, a better way of phrasing your question might be: what could a distant, far-future human learn about Earth culture, with only postmodern texts to hand?


notable embodied cognitive sewage

  • lavalier (n. and v.): Pendant necklace. Particularly one with your frat's three-letter name on. Central to an old saccharine ritual in American frats and sorors: you swap lavaliers with your partner as a sort of pre-engagement ritual with your bros and sisters watching.

  • TINLA (init.): This Is Not Legal Advice. To go with my new fav initialisms of epistemic humility, IANAL and IANAD.

  • WLOG (init.): Without Loss Of Generality

  • Taleb's demon (init.): Probabilistic equivalent of Maxwell's demon: a demon fucks with the usual urn metaphor for statistical inference, making you realise that we are never really justified in thinking we understand the sample space, since you can only understand the sample space by sampling. Actually due to Peter Taylor.

  • rate raiding (n. / adverbial gerund): To systematically hit an API to learn the company's ruleset or regression ("rates"). Elsewhere called a model extraction attack.

  • to threadshit (v.): to take over someone's discussion with a series of disgusting images until everyone goes away. A risky word - prone to devolve as "fake news" has devolved into "I don't think that's important" rather than its more useful meaning, "made up bullshit".

  • madaline (n.): Many Adaptive Linear Elements. An early neural net, physically wired up.

  • dindu (n.): A new racial slur, short for dindu nuffin. (That is, 'didn't do' in pseudo-ebonics.) Apparently a dogwhistle from Reddit.

  • Evropa (n.): Europe. A shibboleth for white supremacists.

  • grigri (n.): One brand of brake thingy used in climbing. Produces that cool-looking regular, slow descent. A genericised trademark, which I am usually against.

  • bama (black East Coast n. and adj.): uncool person; a hick. From Alabama. Applied to other black people.

  • aegrotat (UK n.): Degree awarded to a student too ill to sit examinations. By extension, a sick note or a student invalid. Latin for "he is ill", present tense.

  • 5:01er (pej. n.): Person (especially a developer) who leaves work at a fixed time every day. Connotations are laziness and nonseriousness. In reality, of course, leaving work at the contractual time just means you have other things going on in there. 5:01er is sometimes retconned to mean "someone who doesn't actually care about coding and doesn't want to improve". But the etymology gives their intention away.

  • Woodley effect (n.): A putative multi-generational fall in intelligence levels. Has to face up to the empirically confirmed multi-generational rise in intelligence, but people I respect think both are real. (e.g. "Flynn is due to nutrition and unleaded fumes, Woodley is due to genetics, Flynn is larger for now".)


me throughout the ages

Era Job Morals Prospect
(fl. 2015)
Data scientist

Effective altruist
(fl. 1975)
Computer programmer *
Consequentialist Extropian
(fl. 1870)
Logician or Inventor
or compiler of
mathematical tables
Utilitarian Positivist / Fabian
/ Nietzschean
(fl. 1800)
Pamphleteer /
Power loom
Hutchesonian /
Late Humean
Universal Reason
Early Modern
(fl. 1650)
Dutch liberalism
/ Leveller / Quaker
Baconian optimism ***
(fl. 1400)
Humanism Republican Humanism.
Middle ages
(fl. 1200)
"gramarien, retoriki,
filofer, geometrer,
logissian" **
by default
by default
... ...
... ...
Middle kingdoms
(fl. 800 CE)
Śāntideva Buddhist Bodhicitta
... ...
... ...
Warring States
(fl. 400 BCE)
Mohist Sheng (聖) / Junzi (君子)

* Maybe "expert system designer".

** More likely lay clergy. If we're going by birth rather than affinity I would be a "turnip herder".

*** The object of which is more or less our present day.