reflecting on 2024

I’ve always found it a little funny that humans place so much emphasis on certain events. We’re almost about to make it one lap around the sun on a floating rock in space, hurray!

We celebrate these ephemeral coordinates with an almost sacred weight. Funnily enough, these fleeting moments are not even singular. The cosmic notion of "being born again" scatters itself across various time zones.

We gather our joys and sorrows from the last lap, along with our restless hopes for the next. Like Gatsby’s dream, we find ourselves beating on, borne ceaselessly back into our well-worn habits.

I’ve always wondered why we choose to wait until this symbolic point in space and time to begin making our mark on reality. What if instead we had a new week’s resolution? Or a new day’s resolution? Perhaps a new hour’s resolution?

Throughout this current lap around the sun, I unknowingly set some mini resolutions. Some goals were met; others fabulously failed.

There’s a saying, “Shoot for the moon. Even if you miss, you’ll land among the stars.”

I’ve always been pretty bad at setting good goals. For a long time, my philosophy was that goals should try to be moonshots, exceptionally ambitious goals that push one’s self to the limit, the idea being that this would drive growth the most.

Even if you failed, you’d still probably end up in a better place than others en route to the goal. I adopted this mindset around the time I was in middle school, and I think it has played a large part in crafting the person that I am today.

This setup has a very sparse reward signal, though. There are times when you wonder whether you’re going in the right direction, and it’s very psychologically difficult to keep your motivation when doing something like this.

Anyway, when you think about it, new year’s resolutions are a little strange. Still, maybe they’re worth another shot, at the very least to fix my sleep schedule. Those who know me can vouch that I’ve begun to look like Rigby from Regular Show due to my dark eye circles from sleep deprivation. I think I averaged four hours of sleep a night in 2024.

I like to follow the rule of three, which I might talk about in a later blog post, so let’s set three concrete goals for 2025:

1. Get at least 1% better at something every single day

2. On Saturdays, plan ahead for the next two weeks

3. Nine hours of sleep a night

For resolution number 1, I hope it will give me a north star to follow in improving myself. This is a common strategy that people use to motivate themselves since (1.01)^365 ≈ 37.8 . That’s about a 3,680% improvement, which is kind of crazy to think about.

Of course, a 1% on day 365 is much larger in magnitude than a 1% gain on day 1, but the guiding principle stands. This resolution is open-ended on purpose. It could mean that I get really good at CUDA or really good at gardening in 2025, who knows!

For resolution number 2, I hope it gives me more structure in my life. There is the classic question of whether one should live their life spontaneously and go with the flow, or whether they should meticulously plan out every second of their day.

I find there are massive diminishing returns after 3 days (there’s the rule of 3 again…) of being lazy. I guess it kind of makes sense. Surrendering to the flow of things makes you susceptible to unexpected changes. Your life becomes unstructured, your goals wander, and your sense of purpose becomes vapid.

It might feel good to bury things away for a day or two, but these things will eventually have to be dug back up. It’s best to know what to do once you dig them back up.

I notice that I’m actually more at ease when I schedule things super carefully. This doesn’t mean that you do work all the time. You could schedule out blocks of time where you do nothing but relax. It’s nice to take days off and recharge, but eventually that short-term indulgence in not having to worry about anything turns into a dangerous aimlessness that’s easy to fall complacent under.

For resolution number 3, I hope it gives me the vitality needed for the other two goals, as well as restore my personal health.

Here’s how my 2024 went:

• Decided some time in the winter and late 2023 that I would dedicate my life to building safe AGI

• Took seven classes in one semester at Princeton: Natural Language Processing, Computer Vision, Reinforcement Learning, Distributed Systems, Theory of Algorithms, Compilers, and an independent work project - I did this not out of obligation but rather to speedrun the rest of the CS department; can’t say I would recommend doing this, but I learned a decent amount I guess

• Adopted as someone’s kitty cat

• 7 classes noises

• Snuck into a room at the top of East Pyne with my girlfriend and threw paper planes of peace out the windows

• Worked in Elad Hazan’s lab on spectral filtering for state space models
• Spent some quality time with my girlfriend before she left the states

• Got pretty decent at PyTorch
• Read a lot of ML papers

• Learned how to do distributed LLM pretraining
• Trained a 2B model from scratch (code)

• First-authored a paper and posted it on arXiv
• Fortunate to be advised by Elad Hazan and Tri Dao for my senior thesis
• Started work on Princeton AI Alignment, Hoagie Club and other evergreen apps for the Princeton community

• Hired part-time for an AI startup

• Got my first citation on a paper I first-authored
• Had Jason Wei come give a talk for Princeton AI Alignment on scaling LLMs

• Applied to some PhD programs
• Wrote my first CUDA program
• Almost met Eric Schmidt

There were, of course, lots of low lights too. I'm not one to shy away from sharing my struggles, but this blog post is already getting kind of long! Maybe it’d be better if I split them up into more modular, bite-sized pieces in future blog posts.

Here’s to another lap around the sun.

See you next year. Same time, (roughly) same place: