My 2023 in NYC, and what I’m thinking about for 2024

Note: this update was originally sent as a newsletter. Sign up for timely updates.

--

2023 has been one of the most consequential years of my life, but it came with a hefty price tag.

In the first 5 months of last year, I got my American visa, married the most beautiful soul in the entire world, sold everything I owned, packed a suitcase, left my family, my parents, my siblings, my friends, and moved to a different continent. Oh, and I turned 30 too.

In the following 7 months, I began setting up our life in the new world. Apartment (for two, without credit history, on a single salary, in New York), credit cards, health insurance, savings management, driving license, social security, banking, documents, green card, EB2, NIW, HSA, (Roth?) IRA, 401k, deductibles, copays, commuter benefits, finding groceries that don’t give you cancer or diabetes…

It’s been intense – to say the least.

And stressful: my wife can’t legally work with our current visa, so making sure money is in order and bills are paid on time has been our primary concern. All this taking minimal time off and holding tight to a demanding (albeit exciting) job.

One thing that prevented me from going insane in this mayhem has been, interestingly, training very heavily with a personal trainer, a former US marine that I met by random chance. When I say very heavily, I mean it. A couple of liver enzymes dangerously spiked under the new regimen, only to revert to normal levels after a few weeks. The guy is no joke.

If I look at my last year from 30,000 feet of altitude, I see a meaningful change in how I live my physical body, and I see a complete overhaul of my environment and context – from the city, to the country, to friendships, to people around me, down to the way I spend my time (and my money!).

I overhauled my life, turned it upside down, but that came at a high price. See, one thing that really bugs me is that last year I significantly slowed down in my intellectual, "professional", growth.

When was the last time I shipped something outside of work? When was the last time I single-handedly shipped a product actively used? When did I learn something new – a programming language, a new skill, a new concept, a new anything? Heck, when did you guys read anything new from me in the last year? Yeah. Exactly. Back when I was still in Europe.

In 2023 I stopped shipping. And that's giving me acute FOMO.

Now, don’t get me wrong. At Replit I'm shipping plenty enough. But the thrill of building a product end-to-end, of kick-starting a learning project to master a new topic or skill, of greenfielding a new subject, is something that I simply couldn’t afford pursuing in 2023, and I miss dearly.

Truth is, if it wasn’t for the incredibly talented colleagues — friends — that I’m surrounded by, and that inspire and push me every day, I would’ve probably stopped growing altogether.

I do think that moving to the US, and to New York specifically, was the right thing: by every single metric my wife and I are better off and are living a significantly better life than the one we had in Europe. We’re happier. I also believe it was the correct strategic decision from a professional standpoint. A significant investment, clearly, but it did set me up to play in a radically different league. I lifted my surroundings by at least two orders of magnitude. You know, small fish in a big pond and all that. This will pay off immensely in the decades to come.

But I feel like most of last year was just like clinging to a boat being rocked by a stormy sea. Not much room for anything other than survival.

I crave more growth. More tinkering, more learning, more shipping, more building, more doing things for the sake of doing them. More poking life and seeing something popping out the other side.

There are a few things I’d like to explore for 2024. Interesting ideas I have been thinking about but haven’t had the energy, focus, strength, stamina, presence to go deeper into.

Theoretical learning – LLMs architectural improvements

The most fascinating thing about LLMs is that – architecturally – they’re both simple and elegant. I’m far from having a PhD in AI, but I can still almost grasp their underlying ticking. That “almost” bugs me. I want to understand more deeply how they work, and where the latest research is going. A couple of examples of where my mind is at: 1. Why linear-time sequence modeling with selective state spaces (Mamba) enjoys a 5x higher throughput than the Transformers architecture (the T in GPT)? 2. How is it possible that compressing prompts by selectively using perplexity can accelerate inference without sacrificing signal?

My plan here is to rebuild my understanding of how LLMs work by studying all the key papers with Andy Matuschak’s learning playbook, and taking it from there.

Do I plan to become an AI engineer? Or to found an AI company? No. Not necessarily. I don’t know. And I don’t care. My primary objective here is to nurture my intellectual curiosity. And LLM architectures do indeed scratch that part of my brain.

Practical learning – build products again

A harrowing realization I recently had is that I forgot how to build products. I used to build mobile apps natively. In Java. That stuff was hard. Today, I can hardly put together a mobile hello world. And I still don’t feel super comfortable with highly complex web-apps. Before zero-ing in on data, at Replit I mostly focused on shipping high-impact, low-complexity features to drive growth. I now want to swap the two, and to ship a product this year that truly challenges my technical skills.

Projects

There are at least a couple of projects that might spin off this learning work.

First, Madhav and I have been working on a new open source library to more effectively augment a dataset for feature engineering by using LLMs. I’ve been working heavily on this topic in the last 6 months at Replit, when I built the infrastructure to tag our code at scale (from code quality to use cases to complexity). It’s fun and, interestingly, also very impactful.

Second, I keep thinking that there’s a spatial element to learning. When trying to learn a new concept, lately, I’ve kept using ChatGPT to query parts of the source material I don’t understand. But sometimes those very answers trigger entire new branches of learning. This recursive branching and go deeper and deeper in the concept tree can, probably, be represented visually, in an infinite canvas à la Figma. There might be a cool tool in there, somewhere, waiting to be built. I want to explore trying to build it myself. Hardly monetizable, clearly, but that’s not an element I’m particularly interested in at the moment (for visa reasons I can’t have any side income beside my primary occupation).

Other things that I might not be able to get to

I haven’t learned a new programming language in a few years. It’s just Python and Typescript these days. It would be fun to pick up something new. Mojo is fascinating, but I don’t have a practical use case to deploy it yet. Sure, there’s always Rust. But likely I won’t get to learn either.

I also will likely not be able to learn Spanish, nor low-level GPU programming, although I’m keeping both in my wish-list.

I’m interested in the first for no other reason than that it’s cool. Learning a new language (a human one!) is mind-bending and mind-expanding. It’s an external and internal journey, an intellectual and emotional roller-coaster. And it’s fun!

I’m interested in the latter because it’s clearly where computing is going. AI requires massively parallelized workflows, sure, but let’s not forget that a non-trivial part of crypto does too. I’m not a huge crypto guy (although I do love the cryptography part), but I still recognize that in a world of AIs, privacy-first systems will be more and more needed. Vitalik presents a compelling case in this essay.

Themes

Thematically, there are a few things I’ve been thinking a lot about recently, and I would love to explore more.

A fil rouge you might notice in the previous paragraphs is language. Human language, programming languages, language models…

I was listening to a podcast recently on the history of how proto-indo-european language spread all over Europe 5,000 years ago, and the devastating cultural impact it had on the society of that time, with echoes that are still here today.

I have this nagging thought that language is more than just a symbolic representation to convey thoughts. It’s clear that language is contributing to making those very thoughts, in a circular feedback loop. It’s not a novel idea, Greek philosophers noticed this too thousands of years ago, but there’s a new interesting angle when it comes to the emergent behavior of LLMs.

I want to spend more time reading about this.

I'm also still extremely interested in the topic of geographic arbitrage of global taxation. I was writing about this two years ago (“the most fascinating topic I’ve spent time thinking about lately is the fiscal impact of global remote work. We live in an era where people think increasingly borderless, while nations are getting more isolated”). Now that I live in the US but have still, somewhere in me, the desire to retire back in my home country, I keep thinking that this continues to be extremely relevant. Not sure what shape this will have, but I can definitely tell this is going to be a massive topic for many people in the next few years.

Closing thoughts

And finally, top of mind for me has been thinking about what my 30s will look like. And I realized that I don’t have a single clue. That’s probably ok.

Sometimes I feel like I’m falling behind. People around me are doing amazing things and I’m here struggling to expedite my wife’s work permit. But then I remind myself that it’s fine. To each their own battle. Immigrants have it tougher, we know it. Probably tougher skin too, I guess.

I’m just happy, now, that I’m re-focusing myself on what brings me actual long-term joy: back to growth!