Prologue: Two Years Early
I was riding in a car on the highway, watching a World Cup match on my laptop.
Today that sentence sounds ordinary. At the time, almost no one could have said it. There was no iPhone yet, no real mobile browser, almost no YouTube. People sent texts and a little email from tiny screens. The match was streaming over a phone we had built ourselves at Qualcomm, and nothing like it was in anyone's hands yet.
Inside the team, we had been asking one another, honestly, who would ever need this. In that car, I had my answer.
I had that experience about two years before everyone else, and it changed my sense of what was coming. The future stopped being an equation and became something I had lived. A technology can look unnecessary right up until it becomes inevitable.
Moments like that are rare. At almost every turn of my life, I could see only a few steps ahead.
Roughly every five years, the work I do for a living has changed shape. I began as an academic researcher, became a signal-processing engineer, then a quantitative trader, then a data scientist. Today I lead a machine learning team at Google.
From the outside, the path looks planned: METU, Cornell, Qualcomm, a trading firm, Apple, Google, each name handing off neatly to the next. From the inside it never felt that way. What carried me was not a plan. It was curiosity, a habit of studying in the evenings, people who knew more than I did, and a willingness to stop being the person I had been five years earlier.
Only looking back do I see that these were not separate careers. They were one thread. In every job I was looking for the same thing: a faint, meaningful pattern buried in a great deal of noise. The signal never went away. It only changed form.
My parents could not have predicted my path, and I cannot predict yours. So instead of a map, I want to leave a record: what I did, why I took each turn, what I got wrong, and the few principles that persisted. I am writing it for my sons, and for any young person standing where I once stood, unsure which door to open.
The Ceiling
Ankara, 1996–2000
The first year of engineering bored me.
I had finished Atatürk Anadolu Lisesi in 1996 and entered METU, the Middle East Technical University, to study electrical and electronics engineering. Much of that first year repeated high school. I concluded, a little naively, that if engineering was this manageable, I could study mathematics alongside it.
So I joined the double-major program. It was far harder than I had expected. I finished both degrees in 2000, near the top of both departments, but the grades were not the real lesson. The difficulty was.
Keep raising the bar and you will eventually find your limit. Was I an Einstein? No. A solid scientist? Perhaps. Everyone has a ceiling, and the only way to learn where yours sits is to push until you touch it.
Signal processing caught me there, because it was where engineering and mathematics overlapped most. I read signals textbooks in my spare time, for fun. That habit of reading what no one had assigned would pay for almost every turn that came after.
Between my third and fourth years I spent a summer at Northeastern University in Boston. I had dreamed of MIT, just across the river, but I knew no one there. Northeastern was the door that was actually open, so I walked through it.
It would not be the last time.
Five Hours from Manhattan
Ithaca, 2000–2004
I thought I was moving to New York City.
Much of my class at METU applied to American graduate schools together; I remember us drafting our statements of purpose side by side in the lab. Stanford admitted me without funding, which for me meant no. Cornell was the best school that offered a full scholarship, and I chose it over Georgia Tech and the rest.
Cornell was in New York, and to me New York meant the city. I didn't know New York was also a state, or that Ithaca sat five hours from Manhattan, and there was no real internet to check.
Ithaca turned out to be beautiful, with two rivers running through campus and Cayuga Lake below the hills. It is a wonderful place to give yourself entirely to science, and a hard one if you love big cities.
I spent four and a half years there on a master's and a PhD in telecommunications and signal processing. I describe what we did as "big data without data." Today you gather enormous datasets and let a model learn from them. We chased the same goals with probability theory, stochastic processes and statistical models, mostly on paper. My advisor took me on largely because of my mathematics degree, the one I had added because my first year in Ankara bored me.
Much of our funding came from U.S. Army research labs. Between 2000 and 2004 we worked out how unmanned aerial vehicles could form communication networks, years before most people had seen a drone. Our sponsors told us this was where everything was heading within ten or twenty years.
They were right.
I did well, but I didn't want to be a professor. Research was hard, and even an excellent paper might reach only a few hundred readers. At METU I had built practical systems and loved watching them work. Writing theory without ever learning whether it held up in the real world did not satisfy me. I wanted to build things and watch them work again.
A Billion and a Half Pockets
Qualcomm, 2004–2009
Design this receiver block for the modem that will reach the next 1.5 billion people.
That was the assignment I was handed almost as soon as I arrived at Qualcomm in December 2004. I formally received my PhD a month later. I was in my mid-twenties. In academia, a few hundred readers is a good outcome. Here, a billion and a half people would carry my work in their pockets. That difference is why I chose industry.
It was a strange moment to arrive. The dot-com bubble had burst, and 3G had disappointed everyone. We were building 3.5G.
A phone hears every transmission many times over, as echoes bouncing off buildings and hills, tangled with the noise of everyone else's calls. We were replacing rake receivers, which gather those echoes and add them up, with equalizers and interference cancellation, which try to undo the tangle itself. It was the problem I would keep meeting, in one form or another, for the rest of my career: pull a faint signal out of a great deal of noise. Peak data rates rose from 3.6 megabits per second when I joined to 21 when I left.
As the numbers climbed, we asked one another who would ever need this. You already know where I found my answer: in a car on the highway, two years before everyone else.
Qualcomm's leaders saw three to five years ahead. Often they were simply extending a trend line with a ruler, then finding the courage to build what the line demanded. Engineers like me could not always see why.
I have tried to remember that ruler ever since. Years later, I would lay it across a chart of my own firm's profits.
The Evenings
2004–2009
Around 2004 an idea took hold of me: if you could understand the brain, you could solve almost everything else. If I ever went back to research, that was what I wanted to work on.
In 2006 or 2007 I picked up a book on machine learning, The Elements of Statistical Learning, and studied it purely for pleasure. I was stunned. Problems we had wrestled with in signal processing, or solved in our own clumsy ways, had cleaner answers there.
I came to believe that the core theory of signal processing, the field I loved, had largely been settled by the 1970s, and that since the 1990s the frontier had moved to machine learning.
Three forces then pulled me toward finance at once.
The first was money, and I won't pretend otherwise. Back in 1999, a friend choosing between Bilkent and METU, both on scholarship, was given blunt advice: pick whichever pays more. It stuck with me. Years later I noticed classmates with the same degrees earning one and a half to two times as much for the same work, and I asked myself why I wasn't one of them.
The second was that finance ran on machine learning. The third was that finance hired pure researchers: people who could take a messy practical problem, express it in mathematics, and produce a result in the real world. That was my strongest skill.
So for two to three years, my days and my evenings belonged to different careers. By day I worked on modems. In the evenings, after work, I studied for the next one.
In 2007 I was close to an offer. Then the financial crisis hit. Banks collapsed, and nobody was hiring.
Perhaps I was lucky to miss that job. You never learn the odds of the road not taken.
By 2009 the high-frequency trading firms had come through the crisis strong, and they were hiring people who could build mathematical models. Even then it wasn't easy. Those firms recruited olympiad medalists and prize-winning physicists, and on paper I was a notch below them. Years later I would manage an International Mathematical Olympiad gold medalist. But first I had to get through the door, and patience got me there.
People notice the day someone joins a new company or enters a new field. They rarely see the evenings that came first, spent studying something no one required of them yet. Almost every career change I have made began years before the job change itself.
Bread from Different Ovens
Austin, 2009–2015
Years later, a manager at Google asked what I had done in finance. I said: competitive model building.
You build a model, the person beside you builds one, and if yours is not better, you are out. It is a brutal place to prove yourself, and a clarifying one.
I moved to Austin in 2009 to join a high-frequency trading firm. Every new kind of data demands its own expertise; in Turkish we say you must first eat bread from a few different ovens. It took me one to two years to become genuinely good, and about four to make partner.
My biggest lesson there was about method. My PhD had trained me to create things from scratch, and at first that is what I did. The instinct failed me. What worked was taking the best existing strategies and adapting them carefully: a signal from here, a technique from there, five or ten good ideas combined with judgment. The moment I made that switch, my models began to win.
Don't confuse originality with value. Nobody gives you extra credit for rediscovering what the world already knew. Learn from what works, then build on top of it.
When I left six years later, the firm traded roughly 5% of all NASDAQ volume, and another researcher and I had built about half of the models behind it. We traded nearly everything: stocks, orange juice, sugar, oil.
People asked what value that creates. My answer: we made the market. When you want to sell a share, someone has to buy it right now. Without market makers you would phone a broker and wait half a day, or pay a dealer a wide spread between buying and selling prices, as at an airport currency booth. We earned about a cent on a hundred-dollar share and held positions for minutes. Warren Buffett plays the same role at the far end of the time scale. In 2008 he bought when everyone else was selling, and held for years.
Austin was also where I first met the cloud: thousands of virtual machines training models for 5,000 stocks in parallel. But the firm was falling behind. It moved too late to FPGAs, chips that can be rewired to run trading logic directly in hardware, and to faster hardware in general. Its leaders saw trends one to three years late, the mirror image of Qualcomm.
Eventually I plotted the firm's annual profit.
The best year was $130 million, against about $30 million in costs. Then came 60. Then 50. Then 40.
I picked up the Qualcomm ruler and drew a straight line through those points. It ran to zero.
Even in good years the firm had paid almost no bonus. Meanwhile we had moved back to Silicon Valley for my wife's job, and tech companies now paid more than finance. So I left, following the old advice from 1999 once more: go where you are paid more.
About two years later, the firm where I had made partner was sold for close to nothing.
A Heartbeat on a Wrist
Apple, 2015–2016
It took courage to hire a finance person to build AI for a watch, and I am grateful my managers at Apple took the chance. Someone, in effect, said: I think you can do this, go. Careers hold more of those moments than people admit.
Before going back to tech, I had rebuilt myself once more. Finance had shown me that strong programmers turn ideas into results far faster than I could. So in 2015 I studied software engineering, data structures and algorithms, and data science until I could answer almost any interview question. Every job change in my life has taken six months to two years of preparation on the side. If you don't renew yourself, you can't answer the questions.
The brief on the Apple Watch was simple: imagine things nobody has built, then build them. I modeled people's exercise habits, predicting, for example, how many calories you would burn between four and eight in the evening. The watch could then nudge you: a ten-minute walk would hit today's goal; would you like to swim tomorrow at three?
That same year I read Ilya Sutskever's sequence-to-sequence paper and nearly fell off my chair. A neural network could read one sequence of words and write another. I wanted to do exactly that on the watch: predict the next word, and draft three replies to an incoming message. The paper later won a test-of-time award at NeurIPS, and Ilya's work was my greatest inspiration in those years.
After six to eight months on the exercise project, we presented it to a director.
He said he would not have done it our way.
It was a shock. Perhaps I wasn't mature enough yet; perhaps I didn't want that fight. Apple is a design company, where thirty people may debate the color of a single icon, and AI was never at the front. I concluded that serious AI would run on servers, not devices, and that Google would get there long before Apple.
The Steakhouse
San Francisco, 2016–2017
A hedge fund owner in San Francisco took me to a steakhouse where a steak cost at least $100. Over dinner he told me he loved paying people a million dollars in their first year, so they would get used to the money.
It wasn't true. He churned his staff every six to eight months to avoid honoring their contracts. I was head of research there for less than a year.
Normally I weigh many factors before a decision. That time I let one of them, the money, decide for me, with help from a smooth talker and a good dinner. It was a detour worth learning from rather than repeating.
The detour did leave me one gift, though not the one he had promised. The glowing review he gave me helped Google hire me a level higher than I might otherwise have started.
The Long Projects of a Life
Google, since 2017
A wedding begins with a ring. Then comes a venue, then invitations, then dresses. Buying a house, planning a vacation and learning to cook unfold the same way: not as single searches, but as long projects of a life, carried out a step at a time. Each of us has somewhere between ten and a hundred of them under way at any moment. If we understand which stage you are in, we can show you what actually helps.
At Google we call these tasks. Modeling them is the work I joined in October 2017.
I came in as a tech lead at level 6, a step above my Apple level. I had offers from Apple, Facebook and Google, and I chose Google because I believed the future of AI would run on servers and be built there. By then I could solve programming questions the way we once solved ÖSS questions for the Turkish university exam, almost without thinking.
Once again I had to master a new field, information retrieval (the science behind search), with another two years of study on the side. The scale is hard to picture. We cluster some 30 billion search queries into hundreds of millions of groups, then cluster those again. One problem was clustering a graph, a network of 30 billion points joined by trillions of links.
It took me a year to build an algorithm better than the one our researchers had. It was the best on five of six academic benchmarks, and it was published at KDD, a major data-mining conference, in 2021. This time I was the practitioner and the researchers wrote the theory, the reverse of my Cornell years.
I tell my sons: if you see a bad ad on Google, or a good one, I may be the reason. I believe this work keeps much of the internet free. Every free article and website you read is paid for by someone, and good ads, shown rarely and at the right moment, are how.
After about three and a half years as a tech lead, I became a manager as well. I lead a team of eight and think of myself more as a professor than a boss. In Ithaca I had decided not to become one; the job found me anyway. I meet each person every week and help shape their projects. Every year we write a plan, much like a professor applying for a grant. Sometimes the goal is half a percent of revenue, which at Google means roughly a billion dollars.
Two things I learned to value late: verification and demolition. At Qualcomm I found checking each block tedious; in finance I saw what happens when nobody does. And every year we try to retire last year's models and replace them with better ones. Replacing five models with one is as valuable as building something new. If you never tear down, you cannot build.
I am level 7 now, and people ask about level 8. I don't chase it. My goal has always been to be useful. If that brings a promotion, good. If not, also good.
The Next Word
What I believe about AI
In 2016, when AlphaGo played Lee Sedol in Seoul, I stayed up through the night to watch every game. In the second game the machine played a move, the thirty-seventh, that commentators first took for a mistake. It wasn't. AlphaGo already contained the core recipe behind today's best models: first imitate humans, then search through your own options to find moves better than any human taught you.
Many people dismiss language models as "just predicting the next word." I never did. On the watch, predicting the next word was exactly what I had wanted to do. Before that, I had spent years predicting what a market would do in the next minute, and I learned that prediction, done well, is understanding. If the next word you must predict is the answer to a hard math problem, predicting it means solving it. These models will move from the next word to the next concept, across many horizons at once, just as we predicted one, ten and thirty minutes ahead in finance.
AGI, artificial general intelligence, is a spectrum, not a finish line. To me, GPT-4 was the first model close to human general intelligence, something like an IQ of 50 or 60. AI will climb to 100, then 130, then 150. It will also be jagged, brilliant at some things and oddly weak at others. An airplane flies faster than a bird but is not a bird; a submarine is not a fish. And AI is already being used to build better AI.
Still, the last few percent take a long time. In 2015 we all thought self-driving cars were a year away; a decade later they are only now arriving. Matching every human ability may take another decade, and the help along the way will be enormous.
To see where this goes, I still reach for the Qualcomm ruler. The cost of AI per token, a word or piece of a word, has been falling about fourfold a year: sixteen times in two years, sixty-four in three. What is impossible today becomes child's play in two or three years. When models respond in real time, inside a robot that sees, hears and answers instantly, it will feel like magic and open uses we can't yet imagine.
On jobs: people who work with AI will move ahead, and those who cannot adapt will fall behind. But production and consumption rise together and life speeds up, so I don't expect mass unemployment.
On safety, I believe the most capable models should not be released as open weights, which anyone can download and run. Someone accountable must always be able to pull the plug: several responsible companies, not a single monopoly, and certainly not anyone at all.
So I wouldn't tell you to prepare for a particular job title. Many of today's titles will disappear, and new ones we can't yet name will appear. Prepare instead to be hard to make obsolete. What follows is my best attempt at saying how.
What I Hope You Take with You
If you remember only one thing, remember to learn how to learn. School is only a simulation of learning. Real learning happens outside it, when you pry knowledge loose from books, papers and videos on your own. Every turn in this story was paid for that way, in the evenings, and everything below follows from it.
Build strong foundations. Mathematics, programming, clear writing, articulate speech, sound thinking, and now the ability to work alongside intelligent machines. A mathematics degree I added because my first year of engineering bored me is largely why my advisor at Cornell took me on. Foundations like these carried over into every field I entered, and they will carry into fields that don't exist yet.
Stay curious. My move into AI began with a book I read in the evenings for fun. Read well beyond your own field; that is usually where the next door is.
Be flexible. What you do will likely change every five years. "Engineer," "mathematician," "trader": each label was true of me once and wrong ten years later. Don't let your identity become a prison.
Stand on the shoulders of giants. My models in Austin began to win only when I stopped building from nothing. Gather the five or ten best ideas around you and put them together well.
Work with great people. The mentors ahead of me shaped my path, and I still look for them. Secure people share what they know; insecure people hoard it. Work among the first kind.
Treat pay as a signal. I have never been shy about following money; that blunt advice from 1999 stuck. In my experience pay tracks value, not perfectly but closely. If people with your skills earn more elsewhere, ask what they are doing that you aren't. Ask how many people your work touches and how important a problem it solves.
Love your work. The luckiest people are paid for what they would happily do as a hobby. I read signals textbooks for fun long before anyone paid me to.
Weigh more than one factor. Remember the steakhouse; that is what happens when one factor decides. And the offer I lost when the crisis hit may have been the best thing that happened to me. Ask of every decision whether it is a stepping stone, and judge it on more than one thing.
Keep looking for better options. Some of my classmates got into METU and acted as if life was settled. Getting in is only the start. I could have joined Google in 2004; I arrived in 2017. The detours made me who I am, but keep your eyes open.
Choose your partner wisely. Your partner shapes the quality and success of your life as much as anything else. My wife studied the same engineering and the same mathematics double major; she understands both my work and my worries. Choose someone who lifts you up.
The Signal, for You
For me the signal was a radio wave, then a stock price, then a heartbeat on a wrist, then a search query. I don't know what it will be for you. I do know that the curiosity to follow it, and the willingness to become a beginner again every few years, will matter more than any degree.
If the companies and job titles in this letter someday become footnotes, that will only prove the point. Qualcomm, Apple and Google are not the important part of this story, and neither is any title, salary or promotion. They are coordinates: they tell you where I happened to be standing. What matters is how I moved between them.
When I was young, I thought the challenge was to discover how far I could go. Now I think the better question is how many times you can become someone capable of going further. You don't need to see the whole road, only what the next stretch of it requires. And always find something better.
Be curious, stay flexible, and add value to whatever you touch. The signal will change. Learn to hear it.
— Gokhan Mergen
I first told much of this story aloud, in a long conversation with my old METU classmate Oğuz Ergin, "Yapay Zekayı Arayan Adam" (The Man Searching for AI), on his YouTube channel in January 2025. Writing it down is my way of making it outlast a video.
Comments