Tl;dr: AI didn’t replace my job. It replaced the easy parts. And that made everything harder.

First, the time AI saves becomes time for more thinking.

If AI reduces a coding task from 2 hours to 30 minutes, you don’t get 90 minutes of rest. You get 90 minutes of the next problem, either delivering three similar tasks or deciding what to tackle next. The cognitive load doesn’t decrease. It concentrates.

Second, AI takes the fun parts of the job.

Writing code, for example, used to have rhythm. You’d get in “the zone”—a meditative state where your mind wanders productively and your fingers follow. It is work, but it is also rest. That’s why most software engineers love coding. When AI writes the code, that rhythm disappears.

Third, AI hallucinations mean constant verification.

In coding, it’s easy to test if AI got it right. Worst case, you verify it in production. (Please don’t do that.) In people management, as contrast, verification is more complicated. AI provides confident answers, but you can’t just trust them—not yet.

In sum: you’re accountable for the next cognitive heavy lift, AI took the enjoyable part, and you have to verify everything it produces. This is a recipe for burnout.

But here’s the thing: we’ve been here before.

Think about the shift from pencils to keyboards. Writing by hand was slow, but satisfying. Typing killed that ritual and opened up something better: faster editing, versioning, and document retrieval.

Or more dramatically, film to digital. Every shot mattered when you had 24 exposures. You waited days to see your photos. Digital destroyed that magic and created something entirely new: instant sharing, smart albums, and storytelling.

Each transition inevitably killed a ritual we didn’t know we loved, yet created possibilities we couldn’t have imagined.

I don’t know what the new rhythm looks like yet. But we should be excited (and maybe a little scared) about the adventure ahead.

Happy 2026.

If you wonder how AI is changing the world, Anthropic has just released the first comprehensive analysis of how AI is used in 150+ countries and all 50 US states.

Insights?

• Task delegation to AI jumped from 27% to 39% in just 8 months!

• Each region adapts AI uniquely, where DC at 4x population share driven by policy/legal makers, Massachusetts dominates life sciences, and Brazil leads language learning at 6x global average.

The future is now.

After my Google internship, I landed at Square1 as a new-grad software engineer. My first project was to rebuild their hardware-selling website from scratch.

Why did we need a new website for selling hardware? Scaling.

In the early days, Square had just one piece of hardware: the Square Reader. A little white plastic device that plugged into a phone’s headphone jack and read the magnetic stripe on a credit card. To seed its payment ecosystem, Square gave it away for free. The first website was a hackathon project built to ship out those free readers.

But when Square started selling more hardware products—terminals, stands, registers—they needed a real eCommerce site with common features such as preorders, payments, returns, and my favorite shopping cart. Those features were hard to build on top of the original site.

So we started over. Instead of using third-party service, we chose to build on Solidus, an open-source eCommerce platform on Ruby on Rails. Our goal was to reduce costs and to keep the flexibility to scale as Square’s business grew.

This project took six engineers and two PMs for six months from prototype to launch. Even with a powerful framework, it was still a lot of work.

(No wonder software engineers were in such high demand. But hey—if you hit the endpoint today, you can still see the website alive.)

My biggest realization at the time was that school hadn’t taught me the practical skills I needed at my first job. I had learned algorithms, data structures, databases, networks, but not “Systems That Can’t Go Down 101”.

In school, if your program ran once in a demo, you got an A. In production, if your code failed once—say, during a real customer’s order—you got a page, woken up at 2am, and the company still lost money.

I was lucky (or unlucky) to learn this lesson early in my career. But I’m not sure if that kind of tactical knowledge—or even the job itself—will be as relevant in the future.

In today’s AI-powered era, with tools like Claude Code, the sheer amount of engineering time required is shrinking fast. A project that once took six engineers six months could probably be done in a month. Also, many technical details will be hidden underneath the “conversation” between engineers and AI in nature languages.

I wonder will future engineers learn less from the work and more from the classroom again?

I rethink it from both sides:

• The practical lessons from the job, especially knowledge tied to a particular framework, may not be as relevant.

• The theoretical lessons from school are intellectually fun, but they also have to evolve—because AI itself is reshaping what foundations mean. For example, should AI be considered part of programming itself?

If there’s one thing I’ve learned across both worlds, it’s this: the only thing that doesn’t change is the need to keep learning. Or, as Steve Jobs once said:

“Stay hungry. Stay foolish.”

This wasn’t the plan.

Back in school, I was fascinated by AI’s potential to change the world. My first machine learning class was taught by Prof. Hsuan-Tien Lin at National Taiwan University. We covered everything from Occam’s Razor to the curse of dimensionality. We built toy neural networks from scratch (this was before PyTorch or TensorFlow were even a thing). Eventually, I became a teaching assistant for the course and later contributed some problem sets to Learning From Data - a Caltech Course hosted by Prof. Yaser S. Abu-Mostafa. Working closely with these fabulous thinkers, the path seemed obvious: get a PhD, become a researcher.

Years later, I landed a PhD internship at Google. It was my first glimpse into AI at scale. The Google infrastructure blew me away. All I needed was a Chromebook—yet behind the scenes, I was orchestrating MapReduce jobs across data centers.

The algorithm itself? It didn’t quite feel like the frontier I had imagined. (Especially now, looking back from the LLM era we live in.)

Then came the moment I met Jeff Dean—legendary engineer and head of Google Brain. His talk was electrifying. But it also made something thing painfully clear: the most groundbreaking work wasn’t necessarily happening inside PhD programs. It was happening wherever three things converged: massive data, advanced algorithms, and serious computational power.

That’s when it hit me: a PhD wasn’t a golden ticket. What mattered more was what you worked on, and how you worked on it—regardless of whether you were in academia or industry.

By then, I had lost interest in my own research in medical imaging. On top of that, my wife and I were expecting our first child. Life was moving, and so was I. I told my advisor I'd finish my thesis remotely for the master's degree I needed to work legally in the U.S.

Starting out in the Bay Area wasn’t easy. We were preparing for a newborn and needed to find housing for a family of three for the first time. During the internship, we had rented a unit with a shared bathroom—something that clearly wouldn’t work going forward.

In the Bay Area, rent was high even for a single-room unit, and most apartment complexes either asked for a six-month deposit or proof of a full-time job. I had neither. All I could offer was: I interned at Google, I’m actively interviewing, and I promise I’ll pay rent on time.

I was doing everything I could—sending out applications, interviewing—but time was running out. As my wife’s due date approached, our savings were almost gone. In a moment of despair, I even stood in line to apply for a “white card” (public assistance). But looking around, I saw others who needed the help more than I did. I walked away.

On the day my child was born, I got a phone call from a fintech startup. I got the job.

I was thrilled, but I asked to delay my start date by a few days to focus on our newborn. When our $2,400 rent came due, I had only $240 in my account. I started work just in time to cash my paycheck and cover the rent.

That’s how my career began—not in a research lab, but in the real world, where systems are messy, constraints are real, and every line of code counts. I started as a full-stack engineer. Not the path I expected, but exactly the one I needed.

A colleague at Anthropic recently told me that Learning from Data—the same NTU class where I was a TA—was his very first machine learning course. He still remembers grinding through those SVM problem sets. He encouraged me to take his GPU class at MIT. Like me, he also left his PhD behind. Funny how those early foundations ripple forward in ways you never expect, connecting paths that seemed completely different at the time.

This is Liu’s Log—a space where I write about the edge of tech and the stories that stick. If that resonates, I’d love it if you subscribed.

Subscribe now

I still remember my first one.

They handed me a laptop, a blue notebook embossed with “facebook,” and a sticker that said: This is now your company.

I did the math—some vanishingly small fraction of the company was now “mine.” Not much, really. But over time, I came to own something more enduring: the work, the mess, the weird surprises, and lessons stuck around long after the stock vested.

I didn't understand it then. But through Composer, Instagram, and PyTorch, I've come to appreciate just how true it was.

I started out as a full-stack engineer on Facebook Composer (the little box at the top of your feed where you post text, photos, videos, and more). At the time, over 30 product teams were building on top of it, but only two backend engineers were supporting the platform. I was one of the two.

To avoid getting buried in endless support requests, I learned to write detailed internal wikis—so I could protect my time and focus on the problems I was genuinely curious about. What amazed me was that, even this year, someone found one of those old wikis and gave it a thumbs-up. Eight years later, it’s still doing its job.

So what was I up to? I proactively asked to work on the Sprouts ranking system—the engine that determined the order of post buttons built by 30+ other teams.

I wasn’t trying to invent the next ranking algorithm. I just wanted to understand how data moved—from your phone to the backend—and how that data powered machine learning. That curiosity cracked open the door to machine learning systems at scale and quietly shaped the rest of my career.

As the Composer team grew, I joined a small team at Instagram as its founding engineer. We spun out from the Instagram Ads Delivery team to ensure the accuracy of Instagram metrics like impressions and link clicks. My background in data and logging gave me the confidence to jump into a domain where data quality mattered as much as the metric products themselves.

My manager gave me two onboarding options: one small-scope but urgent, the other large-scope but long-term.

I said, why not both?

That spirit led me through some long weekends and deeper into both problems. The urgent one—refactoring InstagramAdsLogger—ended up cutting global CPU usage by 0.6% and saved the company $12M/year during the COVID-19 capacity crunch. The long one—migrating over 40PB of organic metric data—set the foundation for a new team that I helped hire and grow. Eventually, I became the Uber TL for Instagram's Ads and Organic metrics and co-founded the Instagram Metrics Council to align 100+ metric products across engineering, product, data, legal, and GTM.

After several years working across Facebook and Instagram, I assumed a new role: building and leading a small Infra team within PyTorch.

The team was known for its homebrew inference engine, AITemplate. This time, my manager gave me just one onboarding option: kill AITemplate.

It wasn't just a technically challenging pivot, but an emotional one as well. AITemplate had been the core of the team's identity. The decision to switch to PyTorch 2.0 led to significant attrition and many unknowns ahead.

We could have folded. Instead, we rebuilt.

We successfully pivoted from AITemplate to AOTInductor and made it the default lowering solution in the company, with numerous model launches and hundreds of millions in annual infra cost savings.

As our scope expanded, we rebranded as GPU Techniques—a team with a broader mandate in the new AI era. We helped launch Llama 2, 3, and 4, contributing to industry-leading performance milestones and sharing key innovations to push the boundaries of what's possible.

At our most recent offsite, 16 of us gathered—a group of performance engineers with a high bar and high trust, brainstorming what's next.

Stepping away now means I won't be part of writing the next chapter. That weighs on me. But I'm proud of where the team is, and I'm confident in where they're going.

Looking back, I realize how much I learned from the people around me—brilliant engineers who taught me to think differently, leaders who gave me room to grow, teammates who pushed through impossible problems alongside me. Meta became the place where I discovered what I was capable of.

I'll never forget my son took his first steps the same day I joined Meta. Eight years later, he is now a passionate baseball player. He has grown a lot. And so have I.

Thank you, Meta.

Welcome to Liu's Log

This is where I’ll be writing stories from the edge of tech—behind-the-scenes looks at scaling systems, building teams, and navigating the weird, fast-moving world of AI.

Whether you’re an engineer, a tech lead, a manager, or someone figuring out your next move—I hope you’ll enjoy the ride.

If you’d like to follow along, please subscribe.

Subscribe now