Recommended Reading: History of Prolog, Godel Escher Bach, the Maple programming language
TLDR: This is really a love letter to my brief love affair with research maths.
I really wanted to do research maths in college and I ended up at Western university doing a summer grant for symbolic computation in my second year. This was under Greg Reid’s research lab, who was one of the core contributors of the programming language Maple. At this point Greg was tenured and I was one of his only students for the summer so I was really just there to fill a seat for him (thank you for humoring me, Greg).
Greg had me rolling the dice on Maple’s Taylor expansions for different solutions to the two dimensional wave equation, which often produced intractable and hard to reason about results. I added in different boundary and initial conditions without really understanding what it was I was even doing at Greg’s prompting and he lectured me about the difference between real and complex numbers.
Meanwhile a few of the other professors in Western’s applied maths department had been big names in symbolic proof solvers in the 1990s, a field that had made tremendous strides and then hit a giant wall of intractability. People familiar with advances in symbolic logic AI in the 1980s at the MIT research lab would probably see parallels between these and the limits of what AI agents could do at the time.
Nobody told me that math, the act of abstracting concepts into symbols and subsequent “symbolic manipulation” of those symbols, was fundamentally intractable at the global level! Actually people much smarter than me are still barking up those trees, you can go take a look over at recent activity in category theory. Please don’t get mad at me, category theorists.
I was undeterred by my failures in Maple and continued barking up trees around differential equations. Basically I was searching for a theory of everything about 50 years behind my time, and I was incredibly hostile to quantum physics.
At the end of undergrad I got introduced by Adam Stinchcombe to numeric computing, who was then working on mathematical models of the neuron after 5 years of basically being the only mathematical biology PhD student at NYU. I continue to maintain that Adam is the best and has my favorite research program in all of UofT, or at least top 3. It turns out that neuronal imaging at single-neuron resolutions via MRI is both expensive and unwieldy and multi-neuronal systems are hard to model using partial differential equations at any sort of significant complexity.
We talked a little bit about me maybe working under his program but I honestly wasn’t the strongest math student and having also taken some computer science courses I decided to flake and enter industry. I did flail around for a bit cold calling random professors in the department who were also less than impressed with me however, and I wrote an independent study modeling dating apps with dynamical systems.
I knew basically nothing about hedge funds when I joined WorldQuant in NYC in 2019, but I did know that they seemed to hire a lot of smart people and that they had a lot of money (which the smart people helped them make). Also, there were more of them in New York than Toronto for some reason.
I also don’t remember the interview at all but I was one of WorldQuant’s first hires out of Toronto and they gave me a just-graduated offer of almost $200k US, which seemed like a huge deal to me at the time. Within WorldQuant I putzed around debugging different setup issues the quant analysts had with our Gitlab instance before my manager noticed I was terribly bored and gave me some computational statistics to work on. I was still terribly bored by this and made the rash decision to quit and take a different job doing machine learning infrastructure at a startup before a year of tenure, inadvertently burning a few bridges along the way.
At Spell I was continually disappointed by the lack of ML-specific training that our senior engineers had. Most of them were product engineers from Facebook and we very rarely understood the details of ML workflows well enough to create a truly great product.
I kept a lot of my opinions to myself at the time and I think I probably would have gotten ignored if I had piped up, partly because I hadn’t learned how to persuade people very well. I tended to assume that people in the room were smarter than me if they were senior and took for granted that whatever background information I knew must be obvious to my colleagues. At the time I probably had a stronger ML background than most graduating Masters students in the field but because I didn’t have the credential it was kind of difficult to convince people that they should listen to me.
Over time I let my math skills atrophy because they weren’t being used and ended up doing a lot of Devops work which felt like the useful thing to do at the time. After studying Spell’s infrastructure a ton I concluded that the Facebook engineers who’d done its overall design also had less than solid infra sense, and we had some pretty glaring scaling bottlenecks which never reared their ugly heads all that much while I was there, luckily because I guess we never saw the full scale.
Spell’s CEO used to be a big name at Facebook for implementing one of the first newsfeed engagement algorithms, and I think that had a large part to do with our acquisition by Reddit.