Do the thing.1 Whatever the thing is you’ve been thinking about doing, go. DO IT NOW. It’s okay, I’m a blog post, a text file sitting on a server with infinite patience. I’ll still be here when you come back2.

…still here?

Then I guess you’re just the person3 I’m writing4 this for, or you ignore instructions. So here’s the deal. I’ll have this interesting5 blogpost here waiting for you, and in exchange you have to do for me the following: the thing.

No, not reading someone else’s guide to doing the thing. No, not reading wikipedia on the thing or passively watching YouTube tutorials of doing the thing6. Do the thing. Do it.

If you’re new at the thing, you’ll probably be bad at it. That’s okay. Your thing might be some important task you’ve been putting off, or learning a new craft or skill, or making plans for a drastic life change that requires taking a big risk.

It’s very easy to get stuck in the mindset of always putting things off, or putting yourself out there, or not feeling ready yet to have a proper go, because the current option is easy, and safe.

So I’m doing you a solid here by provide a strong hard-to-weasel out of excuse to go do the thing.

Set an alarm for 10 minutes, pause all your notifications or whatever, and spend a solid uninterrupted measured-by-the-clock TEN MINUTES on trying to do the thing.

I even went so far as to add a bunch of bloat javascript7 just to hide the rest of the blogpost under a checkbox so your eyes won’t sneak ahead, and you can’t bullshit yourself that you’ll do the thing after reading the rest of the post. You will have to actively choose to cheat yourself out of an opportunity to do the thing, and miss out on the sense of satisfaction of having made a good-faith attempt. If none of that stops you, well, there’s nothing else I can do. I’m a blogpost, not the police.

Why am I writing this?

There’s a few reasons that I tell myself. I leave it to you to decide which are genuine, and which are post-hoc rationalisations.

  1. It’s fun! That’s sufficient justification alone for a lot of actions.
  2. It’s worth some signalling points to write down a list of achievements and put it on a website1.
  3. Having a list of things that I did helps to remind myself that I am making forward progress when it feels like I’m not.
  4. I want to stop past-me from making the mistakes he made (mostly a function of poor trade-offs between exploration (not enough) and exploitation (too much)). Can’t do much about that. Maybe someone else can learn from the things I should have done earlier, and crib from my list to avoid learning it the hard way.
  5. Writing something makes writing more things easier. Look at all the links below that 404! They’re like little seeds for future posts that I think about at the time when writing something else (like this document), and so by adding a (currently broken) link, it commits me to writing that post eventually.

The things I did

My thing(s) that I’m glad I did/learned, in some loose off-the-cuff decreasing order of magnitude-of-effect-it-had-on-my-life2, were:

Drastic Career Shift

Having the courage3 to leave a job as a teaching assistant at my home university (that paid reasonably well and provided much4 job satisfaction) to start a PhD, which opened the doors many other such opportunities, such as

  • co-authoring a textbook.
  • working as a research assistant with KASL.
  • working as a recurring tutor for ARENA (an AI safety upskilling workshop).
  • having a much better working environment5 that allows me to be surrounded by lots of interesting value-aligned people6.

In retrospect, I should have spent more time exploring other options first. I went with a PhD as it seems like the obvious “next thing to do”. If nothing else, the credentialing from being in a PhD program made it easier to leapfrog to where I am now, though taking 6 months to build skills as a software engineer7 would have been valuable.

MLAB2022

Attending the MLAB2022 program run by Redwood research. It clarified for me what a lot of technical AI safety work others were doing looks like, and gave me a set of skills in practical boots-on-the-ground ML that would have taken months to accumulate alone8.

  • I didn’t really absorb everything until I had a second go through the material during ARENA, where I was part participant/part tutor. I don’t think of myself as a particularly slow learner, but more that the other attendees were wicked clever/the pace of the material was very intense.

From Tutor to Course Coordination

Doing things in my job as a tutor that’s usually beyond the scope of the job: Being actively involved in the high-level decision-making of course content.

  • The most enjoyment I had doing this was co-designing a toy CPU ISA from scratch, still used to this day.

  • Teaching a course I had not taken before9 was also very enjoyable: I’d attend the lectures with students, and then deliver the tutorial on the content a few days later. The short recurrent deadlines forced me to absorb the material quickly not just to the point where I could understand it, but then teach it to someone less familiar so they could learn the material faster than I did.

  • I got to lecture for a week! But before I did, COVID spread to Canberra and the university kicked us all home, so I delivered it online from my bedroom. Not so fun.
  • I got to make/co-make YouTube channels for first year theoretical comp/engineering maths courses. I look back and cringe a bit at the low budget production10 quality11, but people still watch it and find it useful, so that’s good enough for me.

Tolerating risk

Moving past a strong risk aversion that got in the way of improving my life. This lead to

  • Being comfortable investing in stocks12 rather than just accruing interest.
  • Trying out online dating.
  • Being comfortable moving overseas for a long period of time13.
  • Identifying which problems in my life can be fixed by just throwing money at it (as opposed to fixed it myself) and then throwing money at it.

Narrow skills in demand implies de facto mastery of all skills

On some level I already know this to be true, but I hadn’t fully internalized it/wasn’t acting like it was true. People specialise in the skills they are good at, and that the world needs (read as: someone will pay them for it.) You then use the money to rent someone with the skills that you don’t have. It’s not laziness, nor should it be viewed that way. I get a sense of satisfaction, as I’m sure many people do, with fixing something from scratch, or being self-sufficient. But the world can’t operate like that at scale. I’ll expand on this more in a separate post.

Stop Goodhearting your life

Reflecting on when I’m optimising for a proxy of my goals.

  • Over-optimising for GPA/grades. Sure, it helps when applying for a grad job, but (I would suspect that) a portfolio of projects or a github with activity or something, anything, to demonstrate you can do stuff. I barely remember most of the physics that I battled with in my undergrad. What good was it getting a good mark in those subjects, if I didn’t retain a lot of the content?
  • Continuing to study physics just so I could have an additional major on my degree I think was a good example of this. Nobody cares that far down the line what major I did or didn’t do, and this would have freed up more breadth courses for something I would have enjoyed more/would have been more instrumentally useful later on.
  • Over-optimising for accumulation of money/not spending money when it makes sense to do so.
  • Having a safety net is good, but forgoing fun activities or opportunities, not spending money on a suitable laptop but making do with an older piece of junk, is not worth it in the long run. Large expenses can be a pain point, but consider all the time wasted on juggling around a laptop with poor battery life, and a low resolution screen that is unpleasant to work on. You spend most of your day (well, I do) interacting with the world via a laptop/phone. It makes sense to buy nice ones.
  • Huge value add in buying fancy categories of cheap things, as elucidated here. Life is too short for cheap socks.14
  • Sometimes it makes sense to buy the subscription to a premium version of software if it saves time. The biggest one of these (imho) is GPT-4 Claude 3.5 Sonnet15. I get much more than USD$20/month of utility out of these services16. At this price point, if you can save at least one hour of time a month with it (which I imagine you will), then that’s worth it. If your usage is relatively low, even the free version might be sufficient.

Minimalist Lifestyle

Trimming possessions down to what will fit in a suitcase hiking backpack + a bicycle is very freeing. I can move somewhere new with only a few hours notice if needed. I used to hang on to things just in case I would need it one day, and very rarely that did happen and I did need something I stored away for a long time, but the mental cost of having lots of junk, and sorting through the junk was not worth the one-off ~$20/few hours of my life that I had saved by keeping something. Almost always I never end up using the things I keep that I don’t have a specific use case in mind. My rule of thumb now is that if I hold onto something just in case, and case doesn’t show up in <6 months, I sell/donate/dispose the object. Requiring the constraint that I can carry all my stuff ensures I stick to this rule.17

Studying Pure Maths

Spending a large part of my studies on abstract nonsense pure mathematics, which ended up being far more useful than I could have known. Expanded to its own post.

Cycling

Picking up road cycling, and using it as my main form of transport. Expanded to its own post.

  • I didn’t know it at the time, but this is the ideal way to get around the UK. I can cycle around the city itself, and take me + my bike between cities on the national rail. It’s definitely the fastest way to get around (at least in Cambridge + central London).

Rock Climbing

Picking up rock climbing: both for the physical fitness, the social interactivity, and the mental benefits regular exercise has. I’m partial to both bouldering (climb the colourful plastic pretend rocks inside with a soft foam mat beneath you) and top-rope climbing (same as bouldering, but you wear a harness and someone belays you down the wall when you climb to the top). Bouldering is less effort to set up and doesn’t require coordinating with a belay partner, but top rope is more social (by virtue of requiring a belay partner), you can rest on the rope if required, and you can do more dangerous risky dynamic moves without concern about how you will fall safely if you mess up. Both (to me) are much more fun than doing weights in a gym, so if you find weight training a chore, I’d recommend you try rock climbing.

Limitations are often self-imposed

Trying (and somewhat succeeding) at learning Chinese + how to play the piano.

  • I was terrible at both studying foreign languages and music in school, and had always wondered if it was a personal limitation of myself. Turns out not! I’ve learned enough phrases, and rote-learned (parts) of a song or two, that I’m convinced that if I poured enough time into it, I could get reasonably good at either of these. It’s also a lot more fun to learn something voluntarily rather than being forced to.
  • The utility in learning these was more confirming for myself that I’m not too old for the neural plasticity to go away, rather than any use that Chinese or piano skills has to be directly (which, admittedly, isn’t that much.)
    • EDIT: Actually, limited Chinese skills was actually a little bit useful for a research project I’ve been working on regarding the language of the latent space of models. Still haven’t found any use for piano, but I’ll let you know.

Writing

Hopefully getting into the habit of writing more often? This blog is probably going to be garbage/raw stream of consciousness, but that’s okay. Because I’m doing the thing18, and hopefully this is the first bad post in what would be hopefully a list of progressively less bad posts19.

  1. Is it worth even more meta-signalling points to write in a self-deprecating tone rather than the tone one would use on a cover letter to present one’s best self? Ehhhh maybe? But that would take all the fun out of it for me, so even if it’s not best practice (much like my copious use of footnotes), I’m going to write it the way I want to write it anyway. ↩

  2. I’m ignoring for the moment that some of these events are/were preconditions for others, or wholly out of my control: If my maths teacher in my final year of high school didn’t recommend me to stick at the maths course I wanted to drop out of, I likely wouldn’t have taken pure maths20 at university, and my life would be on a radically different trajectory. ↩

  3. foolhardiness? ↩

  4. I was tutoring like 4-5 different courses a semester, for ~two years, pulling crazy amounts of hours teaching, doing course level design, and marking. Apart from the marking, I absolutely loved it, a little too much. ↩

  5. Both Meridian and LISA have been gracious for having me stay and cowork out of their respective spaces, and for that I’m grateful. I’ve much enjoyed working out of these spaces, and having coworking spaces like these are a huge value add: Having a hub available makes spinning up programs like CamLAB as low friction as possible. ↩

  6. I recognise I’m in an extremely fortunate position to do so, both as a combination of having both been lucky enough to be born in the “right” country to make acquiring a VISA to the UK easy, and having my brain wired the way that it is21 such that a lot of things I find enjoyable (like teaching) are things people will pay me for, which allowed me to build enough capital to make taking such a risk possible. ↩

  7. Well, maybe. LLMs write a lot of code that I use for experiments, or when I just need a bunch of boilerplate written without much thought, or when I’m trying to just get a quick and dirty script to hack something together. But SWE skills would be valuable for bigger projects and higher level decisions. Maybe I’ll just wait 6 months and Claude 3.5 Opus or GPT-5 can handle it. ↩

  8. Not to knock teaching yourself something, by the way. I’ve self-taught myself lots of things, and this is fine when the goal is obvious, and the path to get there is very clear (juggling, driving a car, piano etc.) When it’s a new area that you don’t know much about, there’s the danger of wasting a lot of time on a dead end that isn’t worth bothering with anyway. Textbooks solve this problem (someone else thought REALLY hard about how to convey part of the knowledge that took them decades to accumulate to someone in a ~semester.) AI safety is a very nascent field, which means a) it’s not clear what techniques will stand the test of time and b) there isn’t really a good, consolidated single resource22 you can point people at and say “yes this is the thing you learn from” other than a smattering of selected blog posts and papers. ↩

  9. It was a course on introductory theoretical computer science, automata, Turing machines, complexity theory, all that fun stuff. Having a background in pure maths really helped out for this one. ↩

  10. My recording setup was a document camera on campus that I’d use late at night when the place was quiet and just one-shot the videos, and upload straight to YouTube. Editing was live (would just pause the recording if I needed to think or do a jump cut). If I’d do it over, I’d probably use a graphics tablet and add a face-cam, but keep the charm of the off-the-cuff unscripted videos. They were each pretty short (~6-8 minutes, usually)23, so if I messed up I’d just chuck the recording and do it again. ↩

  11. If I agonized over every detail, either they wouldn’t have been made at all. I was making them weekly to match the pace of the course, so something had to give. ↩

  12. The absurdly lucky timing of doing this after COVID had mostly wrapped up, and at the start of the AI boom, is not lost on me. Your mileage absolutely may vary. ↩

  13. I now feel more at home in the UK than I do in Australia, though this is probably a function of the subcultures available in the UK as compared to back home. Sorry Australia24. ↩

  14. Following the advice in the aformentioned blog post, I bought nice socks to try, was shocked at how much better they are, and then immediately bought more and tossed all my old socks out. What other things in my life can be greatly improved by throwing a measly $50 at it? ↩

  15. or whatever is the best frontier model at the time you read this. Sorry GPT-4, but Claude can one-shot complex pieces of code for me, and that’s my main use case for LLMs at the moment. ↩

  16. They’d have to charge somewhere around USD$200/month USD$50?/month before I’d have to consider, but I’d probably still pay it. (EDIT: Given the availability of free reasonable good LLMs I think my breakeven price point is a lot lower than $200, maybe $50 or so?) Though open source models are getting pretty good, so at that price point I imagine other competitors would provide API access to fine-tuned open source models that, while might not be as generally performant, would still be very good in specialised areas (like writing code.) At the time I wrote this anecdote (~June 2023) OpenAI was basically the only game in town for capable LLMs accessible to the public, but lots of other companies are nipping at their heels. The gap between open and closed source models has been closing too! ↩

  17. The exception to the rule is things that have a large value-add otherwise to justify the size. I own a 32” 4K screen plus a monitor stand for it that I have attached to the desk at where ever I happen to work at every day (which is currently LISA). The added productivity of having a large screen + the desk real estate I get back from having a nice monitor arm to hold it is to me justified, and both are liquid enough assets I could quickly sell them if I needed to move again. ↩

  18. did you? ↩

  19. and it’s certainly more constructive than just scrolling brain-rot on my phone, which I’m trying to wean myself out from doing. EDIT: Still working on that one 😢. ↩

  20. I was writing an anecdote here about pure maths but it got too big and it grew to its own post. ↩

  21. I didn’t make a “decision”25 to like teaching (or mathematics, or computer science) anymore than I “made the decision” to dislike the taste of coriander. ↩

  22. AGISF somewhat solves this, though a textbook it isn’t. ↩

  23. I did my homework on this one! Studies of MOOCs on edX indicate that watch time peaks for lecture videos around 6-9 minutes, and then drops off. If you just slap a camera on the normal hour long lecture, it won’t get watched. Lecturing in person and online are much more different dynamics then you’d initially guess, and you have to treat them differently. I couldn’t find the original paper I read at the time, but this one has the same conclusion, so good enough. ↩

  24. In Australia’s defense, I grew up in a rural area far from the big city, so it’s not a fair comparison to Oxbridge/London. I wouldn’t recommend growing up in a rural area (not that one can choose the factors of their birth), the opportunity cost is not worth the nice views and lower rent. ↩

  25. In so far that I have the free will to make decisions in the first place, which I’m somewhat skeptical that I (or anyone) does, but that’s a story for another time. ↩

  1. Perhaps a better title is How I Learned agency or How I became aware of the framing of agency as a skill to practice, rather than something innate that you either have or haven’t, but that’s not very catchy. ↩

  2. and if I’m not, that means github is down, and we have bigger problems. ↩

  3. I think the real audience for this (and the subsequent posts) is me circa-2017 when I was midway through my masters and unsure what to do, or even circa-2012 me when I was finishing high school and just doing physics by default ‘cause I liked it in school. Sadly, I cannot interact with past-me, so other people in a similar situation to past-me (maybe you?) will have to suffice. ↩

  4. Also you’ll have to deal with copious footnotes. I like footnotes, especially when the footnote has the little button next to it so you can jump back into the text where you left off8. ↩

  5. No guarantees or warranties of interesting-ness to the extent permitted by applicable law. ↩

  6. It’s very easy to delude yourself that by passively reading about/watching someone else do the thing that you’ll convince yourself you’ll get something out of it. I fall for this all the time. Sometimes it’s kinda true, but the brain (at least my brain) doesn’t retain information that way. I have to beat information into my head by doing it myself. Your mileage may (but probably won’t) vary. ↩

  7. and it ruins the ordering of the footnote numbers. Look what you made me do. ↩

  8. Is it poor form to nest footnotes? Maybe. But I’m doing it anyway, cause I like ‘em. ↩