[Note: backported from LessWrong]
A decade ago, I decided to save the world. I was fourteen, and the world certainly wasn't going to save itself.
I fumbled around for nine years; it's surprising how long one can fumble around. I somehow managed to miss the whole idea of existential risk and the whole concept of an intelligence explosion. I had plenty of other ideas in my head, and while I spent a lot of time honing them, I wasn't particularly looking for new ones.
A year ago, I finally read the LessWrong sequences. My road here was roundabout, almost comical. It took me a while to come to terms with the implications of what I'd read.
Five months ago, after resolving a few internal crises, I started donating to MIRI and studying math.
Three weeks ago, I attended the December MIRI workshop on logic, probability, and reflection. I was invited to visit for the first two days and stay longer if things went well. They did: I was able to make some meaningful contributions.
On Saturday I was invited to become a MIRI research associate.
[Edit to add: about a month later, I became a full-time MIRI research fellow, and fourteen months after that, I became the executive director of MIRI.]
It's been an exciting year, to say the least.
To commemorate the occasion — and because a few people have expressed interest in my efforts — I'll be writing a series of posts about my experience, about what I did and how I did it. This is the first post in the series.
First and foremost, know that I am not done with my aggressive autodidacting. I have a long way to go yet before I'm anywhere near as productive as others who do research with MIRI. I find myself at a checkpoint of sorts, collecting my thoughts in the wake of my first workshop, but next week I will be back to business.
One goal of this post is to give you a feel for how much effort is required to become good at MIRI-relevant mathematics in a short time, and perhaps inspire others to follow my path. It was difficult, but not as difficult as you might think.
Another goal is to provide data for fellow autodidacts. At the least I can provide you with an anchor point, a single datum about how much effort is required to learn at this pace. As always, remember that I am only one person and that what worked for me may not work for you.
In order to understand what I achieved it's important to know where I started from. Thus, allow me to briefly discuss my relevant prior experience.
I was born in 1989. I have bachelor's degrees of science in both computer science and economics. I started programming TI-83 calculators in late 2002. I've been programming professionally since 2008. I currently work for Google and live in Seattle.
In high school I had a knack for math. I was placed two years ahead of my classmates. I aced some AP tests, I won some regional math competitions, nothing much came of it. I explicitly decided not to pursue mathematics: I reasoned that in order to save the world I would need charisma, knowledge of how the world economy works, and a reliable source of cash. This (and my love of programming) drove my choice of majors.
During college I soaked up computer science like a sponge. (Economics, too, but that's not as relevant here.) I came out of college with a strong understanding of the foundations of computing: algorithms, data structures, discrete math, etcetera. I cultivated a love for information theory. Outside of the computer science department I took two math classes: multivariable calculus and real analysis.
I was careful not to let schooling get in the way of my education. On my own time I learned Haskell in 2008 and started flirting with type theory and category theory. I read Gödel, Escher, Bach early in 2011.
This should paint a rough picture of my background: I never explicitly studied mathematical logic, but my interests never strayed too far from it. While I didn't have much formal training in this particular subject area, I certainly wasn't starting from a blank slate.
In broad strokes, I'm writing this because I was able to learn a lot very quickly. In the space of eighteen weeks I went from being a professional programmer to helping Benja discover Fallenstein's Monster, a result concerning tiling agents (in the field of mathematical logic).
I studied math at a fervent pace from August 11th to December 12th and gained enough knowledge to contribute at a MIRI workshop. In that timeframe I read seven textbooks, five of which I finished:
- Heuristics and Biases
- Cognitive Science
- Basic Category Theory for Computer Scientists
- Naïve Set Theory
- Model Theory (first half)
- Computability and Logic
- The Logic of Provability (first half, unreviewed)
In retrospect, the first two were not particularly relevant to MIRI's current research. Regardless, Heuristics and Biases was quite useful on a personal level.
I also studied a number of MIRI research papers, two of which I summarized:
I made use of a number of other minor resources as well, mostly papers found via web search. I successfully signalled my competence and my drive to the right people. While this played a part in my success, it is not the focus of this post.
I estimate my total study time to be slightly less than 500 hours. I achieved high retention and validated my understanding against other participants of the December workshop. I did this without seriously impacting my job or my social life. I retained enough spare time to participate in NaNoWriMo during November.
In sum, I achieved a high level of productivity for an extended period. In the remainder of this post I'll discuss the mechanics of how I did this: my study schedule, my study techniques, and so on. The psychological aspects — where I found my drive, how I avoid akrasia — will be covered in later posts.
I estimate I studied 30-40 hours per week except in November, when I studied 5-15 hours per week. On average, I studied six days a week.
On the normal weekday I studied for an hour and a half in the morning, a half hour during lunch, and three to four hours in the evening. On the average weekend day I studied 8 to 12 hours on and off throughout the day.
Believe it or not, I didn't have to alter my schedule much to achieve this pace. I've been following roughly the same schedule for a number of years: I aim to spend one evening per workweek and one day per weekend on social endeavors and the rest of my time toying with something interesting. This is a loose target, I don't sweat deviations.
There were some changes to my routines, but they were minimal:
- I have many side projects, most were dropped as studying took precedence.
- The number of weeknights I took off per week fell from a little more than one to a little less than one.
- Before this endeavour I travelled for leisure about once every two months. In the past five months I travelled for leisure once.
While my studying did not affect my schedule much, it definitely affected my pacing. Don't get me wrong; this sprint was not easy. I suspended many other projects and drastically increased my intensity and my pace. I spent roughly the same amount of time per day studying as I used to spend on side projects, but there is a vast difference between spending three hours casually tinkering on open source code and spending three hours learning logic as fast as possible.
The point here is that aggressive autodidacting certainly takes quite a bit of time and effort, but it need not be all consuming: you can do this sort of thing and maintain a social life.
My methods were simple: read textbooks, do exercises, rephrase and write down the hard parts.
I had a number of techniques for handling difficult exercises. First, I'd put them aside and come back to them later. If that failed, I'd restate the problem (and all relevant material) in my own words. If this didn't work, it at least helped me identify the point of confusion, which set me up for a question math.stackexchange.com.
I wasn't above skipping exercises when I was convinced that the exercise was tedious and that I know the underlying material.
This sounds cleaner than it was: I made a lot of stupid mistakes and experienced my fair share of frustration. For more details on my study methods refer to On Learning Difficult Things, a post I wrote while in the midst of my struggles.
Upon finishing a book, I would immediately start the next one. Concurrently, I would start writing a review of the book I'd finished. I generally wrote the first draft of my book reviews on the Sunday after completing the book, alternating between studying the new and summarizing the old. On subsequent weekdays I'd edit in the morning and study in the evening until I was ready to post my review.
It's worth noting that summarizing content, especially the research papers, went a long way towards solidifying my knowledge and ensuring that I wasn't glossing over anything.
Impact on Social Life
The impact on my social life was minimal. I decreased contact with some periphery friend groups but maintained healthy relationships within my core circles. That I was able to do this is due in part to my circumstances:
- I live with two close friends. This meant that social contact was never out of reach. Even when spending an entire day sequestered in my room pouring over a textbook I was able to maintain a small amount of social interaction. If ever I had a spare hour and a thirst for company, I found it readily available.
- My primary partner was, up until early 2014, going to school full time while holding down a full time job. Thus, her schedule was more restrictive than my own and we had been working around it for some time. Our relationship was not further constrained by my efforts.
- My core friend groups knew and respected what I was doing. I was more tense and exhausted than usual, but I had warned my friends to expect this and no friendships suffered as a result.
Impact on Work Life
The additional cognitive load did have an impact on my day job. I had less focus and willpower to dedicate to work. Fortunately, I was exceeding expectations before this endeavour. During this sprint, with my cognitive reserves significantly depleted, I had to settle for merely meeting expectations. My performance at work was not poor, by any means: rather, it fell from "exemplary" to "good".
I'd rather not settle for merely good performance at work for any extended period of time. Going forward, I'll be reducing my pace somewhat, in large part to ensure that I can dedicate appropriate resources to my day job.
It's not like I was working from dawn till dusk every day. There was ample time for other activities: I had a few hours of downtime on the average day to read books or surf the web. I participated in a biweekly Pathfinder campaign and spent the occasional Sunday playing Twilight Imperium. In September I went camping in the Olympic mountain range. I spent four days in October visiting friends in Cape Cod. I spent a day in December hiking to some hot springs. I entertained guests, went to birthday parties, and so on. There were ample opportunities to get away from math textbooks.
Most important of all, I had friends I could call on when I needed a mental health day. I could rely on them to find time where we could just sit around, play with LEGO bricks, and shoot the breeze. This went a long way towards keeping me sane.
All that said, this stint was rough. I experienced far more stress than my norm. I lost a little weight and twice caught myself grinding my teeth in my sleep (a new experience). There were days that I became mentally exhausted, growing obstinate and stubborn as if sleep- or food-deprived. This tended to happen immediately before planned breaks in the routine, as if my mind was rebelling when it thought it could get away with it.
The stress was manageable, but built up over time. It's hard to tell whether the stress was cumulative or whether the increase was due to circumstance. Doing NaNoWriMo in November while continuing studying didn't particularly help matters. The weeks leading up to the workshop were particularly stressful due to a lack of information: I worried that I would not know nearly enough to be useful, that I would make a fool of myself, and so on. So while the stress surely mounted as time wore on, I can't tell how much of that was cumulative versus circumstantial.
I tentatively believe that someone could sustain my pace for significantly longer than I did, so long as they were willing to live with the strain. I don't plan to test this myself: I'll be slowing down both to improve performance at work and to reduce my general stress levels. Five months of fervent studying is no walk in the park.
So you want to follow in my footsteps? Awesome. I commend your enthusiasm. My next post will delve into my mindset and a few of the quirks of my behavior that helped me be productive. For now, I will leave you with this advice:
- There is no magic to it. If you study the right material, do the exercises, and write what you've learned in your own words, then you can indeed learn MIRI-relevant math in a reasonable amount of time.
- Learning fast does not need to dominate your life. There can be time for social activities and even significant side projects. You will have to work really hard, but that work does not have to consume your life.
- If you're going to do something like this, let people know what you're doing. This is much easier if you have people you can turn to for support who don't mind you being extra snappy, people who can drag you away for a day every week or two. Also, stating your goals publicly helps to stop you from giving up.
The difficult part is making a commitment and sticking to it. Akrasia is a formidable enemy, here. If you can avoid it, the actual autodidacting is not overly difficult.
As for specific advice, if your background is similar to mine then I recommend reading Naïve Set Theory, Computability and Logic, and the first two chapters of Model Theory in that order, these will get you off to a good start. Feel free to message me if you get stuck or if you want more recommendations.
Following posts will cover the other sides of my experience: how I got interested in this field, where I draw my motivation from, and the dark arts that I use to maintain productivity. In the meantime, questions are welcome.