While at this point it's a little late for a full reflection on my freshman year, I thought it'd be useful to quickly write up my course reviews and general academic experience. I'm mostly writing this because I want a record of my thoughts to look back on.

My biggest academic takeaway from the year was that I'm capable of being a productive person and focusing for long periods of time. While this only happened inconsistently, it did happen, which is more than I could've said beforehand. The main reason for this was the environment – it was much, much easier to get stuff done in large coworking spaces surrounded by hard-working MIT students. Home was not an environment where productivity came naturally; MIT was.

On the other hand, I spent much of the year generally unmotivated, which of course carried over to academics. Thoughts on how to change that will have to wait for another time.

Freshman fall was also on pass / no-record, so I didn't care at all about grades. I was also rather distracted by learning more about AI, to the point where it became harder to interest myself in things that I didn't see as directly relevant. This was somewhat unfortunate in hindsight, though also somewhat unavoidable. I expect this to be less the case this year.

Fall:

6.S898 – Graduate Deep Learning

• I know a lot of people who really liked this class, but to be honest, I didn't love it. I felt like I already knew most of the content, partially because I was doing MLAB concurrently and MLAB was pretty consistently ahead of the course. But many 6.s898 lessons were also already intuitive, like representation learning or a hacker's guide to deep learning.

• Mostly felt like a non-class. I didn't learn much, stopped showing up to class after a few weeks, and once every two weeks I would grind out a problem set in the student center with friends.

• Did a final research project on sparse autoencoder feature universality with a friend, which was a blast. We started about four days before the project was due, grinded nonstop for those four days, and came out with a solid final project and a lot of new practical knowledge about ML implementation. This was my first real research experience, which was cool.

8.01 – Intro Mechanics

• Already knew around half the material going in from high school physics.

• The problems were unfortunately formulaic. It felt like to succeed, I only had to learn the types of problems they would ask, rather than the underlying concepts.

• Also didn't go to this class, because it was pass / no-record and I didn't care that much.

6.1010 – Fundamentals of Programming

• Basically just a programming grindset class. One lab per week, each one is just a ton of coding aimed at learning all the python basics. A useful and interesting class, though nothing unusual.

18.701 – Abstract Algebra

• To be honest, I should've taken this during a different semester. Group theory was way too abstract to seem interesting to me, the professor was bad, and I just didn't care about the material.

• I think either group theory just isn't that interesting or I lacked the mathematical maturity to appreciate it. I don't know which, but if it's the latter, I probably should've taken it later on in college.

• Stopped going to this class, too, though I lasted a little longer than in my other classes. I started learning better when I stopped going to class and started reading the textbook instead. But I still didn't do practice problems, which would've definitely helped me understand the subject in a more real way (and possibly be more interested in it).

MAS-245 – Kevin Esvelt's Advising Seminar

• Loved getting to know Kevin. He's great. Highly recommend. The actual seminar is also great, though I had read some of the stuff before. But even so, there were some very memorable readings I might have otherwise never discovered. Eg, happiness studies, analyses of the replication crisis, random facts about technological progress, etc.

My fall classes were pretty bad overall. But adjusting to college was hard, and given the circumstances, it could've been a lot worse. Before the fall I had probably spent <30 hours programming before in my life, and I walked away with much improved programming skills and a much more thorough understanding of AI. My main extracurricular was the MIT AI Alignment reading groups, which helped a lot with this last point. I want to have much more organized fun this coming year, so I'm planning to get involved in some other stuff, too (maybe a dance or chamber music group).

Life Things

Some miscellaneous semester highlights that I can think of at 2 AM: blood wrestling at a giant party with Van de Graaff machines and wooden amusement park rides built by students, a lot of hacking (MIT code for legally breaking into buildings at night), long night walks by the Charles, Tom Odell and Joji concerts, a weekend-long alignment workshop, and many other unforgettable moments with new friends.

January Term:

Went to Estonia. Very bleak country come winter time.

Also during this general time: saw a lot of family and high school friends, caught up with friends I hadn't seen in forever at winter sparc, began running MIT AI Alignment's introductory reading groups.

Spring:

6.8200 – Sensorimotor Learning

• Great class! Opposite 6.s898, many people I took the class with disliked it, but I liked it a lot. The problem sets were a massive waste of time, but I really enjoyed lectures. They were very information dense. This served as my introduction to RL, and it taught me pretty much all the important algorithms I needed to know. Only went to a few lectures, but I frequently watched them afterwards online.

• Final project was contrasting the OOD generalization of RLHF and DPO, with the same friend as 6.s898 + one other. This project was little more of a disorganized shitshow than the previous one, but I still learned a ton about efficient implementation and good ML research practice. We didn't get any great results, but we got all the necessary code working in time, and we could've gotten better results with a little more time.

7.016 – Intro Molecular Biology

• Amazing class! Waking up for a 10 AM was hard, so I didn't end up going that much. But alas. I thought I was terrible at biology from back when I got a B in it during freshman year of high school, but that didn't turn out to be true. This class was also taught way better. The profs focused as much as possible on intuitions rather than exact details, and we learned about some interesting niche topics like stem cells and optogenetics. Much more fun than memorizing the Krebs cycle.

6.120 – Data Structures and Algorithms

• Really fucking fun. Sort of like logic puzzles as a class.

• Pretty time-consuming, especially the problem sets, but for the most part I didn't mind.

• Most of it probably wasn't that useful (like, I'd be just as well off with a third of the content), but it was fun. I'm somewhat confused why most MIT students learn this stuff before a variety of basic programming skills, but I'm not really complaining.

• Went to maybe a quarter of these, so on the higher end. Lectures were information-dense. Reading lecture notes afterwards usually worked just as well, though.

18.600 – Intro to Probability and Statistics

• Very easy but also very useful.

• Didn't go to a single class; spent around 5 minutes each day reading the lecture slides.

• On one hand, this could've been a week-long self-study. On the other, it's important content that I'll be glad to know going forward.

By the spring semester I knew I wasn't going to go to most of my classes, so I scheduled the four of them amongst two time slots. I didn't have a single afternoon class all semester, which, given that I often wake up around noon, was not conducive to a good attendance record.

I dedicated some effort to getting good grades in the spring semester. It panned out, but I don't think it was worth it.

Late into the spring, I started working with the Algorithmic Alignment Group on preference learning and LLM generation diversity research, which was pretty fun. This was my first time working on a real academic paper.