Chapter 1: A PhD in Computational Linguistics.

Death Certificate

Deceased: A PhD in computational linguistics.

Date of conception: August 2019.

Date of death: April 2023.

Cause of death: Pending investigation.

Summary: Siddharth Rao, twenty nine, enrolled in a PhD programme at a well regarded Indian institute in August 2019.

His research area was computational approaches to low-resource Indian languages. It is a field that focuses on developing natural language processing (NLP) and machine learning techniques for Indian languages that lack sufficient digital resources. Resources like large text datasets, annotated corpora, or pre-built language models. The goal is to build tools like machine translation, speech recognition, or text analysis systems 

for these languages despite having limited data to train on.

His advisor considered him one of the strongest students in the department. He passed his qualifying exams in the first attempt. He published a few papers in his first two years. By the end of his third year, he stopped making progress.

His interest in the research was still there. His personal life was stable. His intellectual direction was clear. He simply could not make himself sit down and do the work. He withdrew from the programme in April 2023, more than three years in, with a half finished dissertation and no explanation he found convincing. 

Case Narrative

By any measure, Siddharth Rao was the kind of student that PhD programmes are built around. He arrived with a strong master's thesis, a strong recommendation letter, and the kind of curiosity that is needed for serious research. He asked questions that were both interesting and answerable. A perfect PhD student if you asked anyone.

His first two years in the program went well. He completed coursework with high grades. He identified a research direction that excited him. He was building computational models for languages with very little digitized text. The kind of languages that mainstream natural language processing research ignores because the data doesn't exist in convenient quantities. He found the problem beautiful. He talked about it at dinner, on walks, in messages to friends. He could see, clearly, why the work mattered and what it would contribute.

In his second year, he co-authored a workshop paper with his advisor. He published a few other papers independently too. His research was always well received. A senior researcher at a European university emailed him to say she had read his work and found his approach promising. His advisor told him he was on track to produce a strong dissertation. Things were good. The program was on track.

Trouble started in his third year.

The third year of a PhD is, in most programmes, where the structure disappears. Coursework is finished. Qualifying exams are behind you. What remains is the dissertation itself. A document that will take 2-3 years to produce. It has no strict deadline. Nobody is waiting for any results from you. Your dissertation will be evaluated whenever you finish it by a committee that does not yet exist.

From now until the dissertation defense, the student has to generate their own schedule, set their own milestones, and sustain their own momentum across hundreds of working days, each of which looks exactly like the one before it.

Siddharth knew this so he made a plan. He blocked off morning hours for writing and afternoon hours for experiments. He was organised, methodical, and serious. For the first two months, the plan worked. He wrote a section of his literature review. He ran a set of preliminary experiments. He met with his advisor every two weeks and reported progress.

Then things changed.

He sat down at his desk one morning, opened his laptop, looked at the document he was supposed to work on, and felt nothing. The task in front of him, which was writing the next section of his methods chapter, was clear. He knew what it required. He knew how to do it. He knew it mattered. Yet he did nothing. He sat there for twenty minutes, then opened his browser, checked his email, read a paper that was tangentially related to his work, then looked up and two hours had passed. He told himself it was a bad day.

The next day was the same. And the one after that.

Over the following months, Siddharth developed this pattern and could not seem to break it. He would arrive at his desk with the intention to write. He would open the document. He would look at the section he needed to work on. He would understand clearly and completely what the section required. And then he would do something else. He would read papers. He would tinker with code that was already working. He would reorganise his reference library. He would reply to emails. He would volunteer to review a colleague's work. He would attend talks that had nothing to do with his research. Maybe in between he would work on the thesis, but nothing substantial.

He spoke to his advisor: "I understand what I need to do. I want to do it. I believe in the work. I sit down to do it and my brain just won't go. It's like trying to start a car with no fuel. Everything is in place except the thing that makes it move.”

His advisor suggested discipline. Siddharth tried discipline. He deleted social media apps from his phone. He installed website blockers. He left his phone in another room. He tried the Pomodoro technique. He tried writing first thing in the morning before checking anything else. He tried working in a library instead of his office. He tried accountability partners. Each intervention worked for a few days, sometimes a week, before the pattern reasserted itself. By the end of the year, his advisor's tone had shifted from encouragement to concern. The department began asking questions. Siddharth could not come out of his pattern.

In April 2023, Siddharth withdrew.

He left the programme believing something was wrong with him. He used words like lack of discipline, low willpower, laziness, self-sabotage. He believed his failure was his character defect.

The Autopsy

Look at the two categories of tasks that filled Siddharth's days. The small tasks: reading a paper (thirty minutes, clear endpoint, sense of completion). Replying to an email (five minutes, done, visible result). Reviewing a colleague's draft (two hours, tangible output, social approval). Tinkering with working code (satisfying puzzle, immediate feedback from the machine). Each one offered a clear, contained task with a visible endpoint and a prompt sense of completion.

The dissertation: writing a methods section. The reward for finishing it was the opportunity to start the next section. The reward for finishing the next section was the opportunity to start the one after that. The final reward, a completed dissertation, a defended thesis, a doctorate that sat years away. And between today's work and that reward, there was no signal. No notification. No green checkmark. No measurable progress that the brain could register as a win.

His brain was doing exactly what brains do when presented with these two categories of tasks. It was choosing the one that provided a signal.

The Diagnosis

In 1997, neuroscientist Wolfram Schultz, together with Peter Dayan and P. Read Montague, published a paper in Science that explained a new way of looking at motivation.

The scientists monitored brain cells in monkeys and found out these cells didn't respond to rewards uniformly. An unexpected reward caused them to fire strongly. A reward that arrived exactly as predicted caused no reaction. A reward that was predicted but never came caused the cell activity to drop below its normal resting level.

The brain, in other words, was running a continuous prediction system. It was constantly forming expectations about what would happen next, and dopamine was the output of the comparison between expectation and reality. A better than expected outcome produced a burst of dopamine. A worse than expected outcome suppressed it.

When dopamine fires, the brain registers that something worth paying attention to just happened, and that the behaviour which produced it is worth repeating. This signal is what drives effort. This is why people with disrupted dopamine systems can retain full awareness of what they want and still find themselves unable to generate the motivation to pursue it.

Their finding had one additional property. They explained that the brain weighs future rewards less the further away they are. Meaning a reward arriving in ten seconds produces a stronger signal than the same reward arriving in sixty. A reward a month away produces a weaker signal than one arriving tomorrow. A reward three years out produces almost no signal at all. The steeper this discounting is in a person, the more they tend to procrastinate.

A student can genuinely want a degree but still not start the final assignment due in three weeks. The degree is too far away for the dopamine system to care, while the phone in their hand offers a reward in the next three seconds. Someone can plan to go to the gym every morning, but when the alarm rings at 6 AM, the warmth of the bed is an immediate reward and the healthier body is months away. A person can know exactly what career they want, map out every step to get there, and still spend the evening watching YouTube. In every case, the person is not necessarily lazy or broken. The dopamine system is probably doing what it evolved to do. Weighing what is available now far more heavily than what might arrive later.

This is what Siddharth was experiencing. The reward for writing his methods section was enormous: the completed dissertation, the doctoral degree, the career that would follow. He valued it deeply. He could articulate its importance with clarity and conviction. But the reward was three years away. By the time the dopamine system discounted it to its present value, the signal was too weak to compete with the small, immediate rewards of email, paper reading, and code tinkering.

The outcome of that calculation was consistent, automatic, and devastating. Day after day, the dissertation lost. The single variable that was killing his progress was the one he could not change by trying harder: the delay between effort and reward.

Prior Case on the File 

In 2017, Duolingo had a problem that its leadership described internally as a "leaky bucket."

The app was attracting millions of new users, but most of them left within weeks. The product worked. The lessons were well designed. The teaching methodology was sound. People who stuck with Duolingo for months genuinely improved at their chosen language.

The problem was that very few stuck with it for months.

Learning a language takes hundreds of hours spread across months or years. The reward, the ability to speak, read, or understand a new language, sits far in the future. On any given Tuesday evening, the task in front of the user is a five minute lesson on verb conjugation. The gap between that lesson and the feeling of ordering dinner in Spanish in a restaurant in Barcelona is enormous. And the brain, running its cost benefit calculation in real time, consistently chose to give up.

Jorge Mazal, who led Duolingo's growth efforts during this problem, wrote a detailed account of what followed. The team hypothesised that gamification could solve the retention problem. Their first attempts failed. They copied mechanics from mobile games. For example, a "moves counter", that limits mistakes per lesson. Users barely noticed. They then tried a referral programme modelled on Uber's. New user acquisition increased by three percent. Nowhere near enough.

The breakthrough came when the team started manufacturing proximal rewards. They redesigned the streak system: a counter that tracked how many consecutive days a user completed at least one lesson. They built streak saver notifications that reminded users before they lost their progress. They added leaderboards that created weekly competitions among small groups of users, with promotion and demotion between leagues. They introduced XP points for every lesson completed. They added animations, badges, and visual feedback for milestones.

None of these features changed the learning. The lessons were the same. What changed was the signal environment. Every completed lesson now produced an immediate, visible, tangible reward: a streak number ticking up, XP accumulating, a position on a leaderboard shifting. The brain, which could not generate effort for a reward sitting months in the future, could generate effort for a reward arriving in the next five minutes.

The results were striking. Daily active users grew 4.5 times over four years. The share of daily users maintaining a streak of seven days or longer tripled. Users who reached a ten day streak were dramatically more likely to become long term learners.

Duolingo solved the same problem that defeated Siddharth. Duolingo's users were dropping off for the same reason Siddharth could not write his methods chapter. The difference is that Duolingo engineered a solution. They inserted artificial proximal rewards into a process whose natural rewards were too far away to generate a signal. Siddharth's PhD programme offered no such engineering.

A Note for the Reader

Sometimes what you blame yourself for, the procrastination, the inability to start, the gap between what you want and what you do, is your dopamine system discounting a distant reward to near zero. Sometimes it is genuine laziness. Sometimes it is that you don't actually want the thing as much as you think you do. The line between these is blurry, and anyone who gives you a clean framework for telling them apart is oversimplifying.

But there are a few signals worth paying attention to. If you procrastinate on things you genuinely care about and feel confused by your own inaction, that points more toward a discounting problem. If the pattern shows up across every area of your life regardless of how much something matters to you, that points toward something behavioural. If it only shows up in certain domains, it is worth asking if you simply don't want the thing as much as you tell yourself you do.

The most reliable way to solve this is to shrink the distance between action and reward. Create intermediate milestones that are close enough to produce a signal your brain can actually use. A chapter completed by Friday. A dataset cleaned by Wednesday. A paragraph finished before lunch. Each milestone needs to be small enough to be reachable within a timeframe your dopamine system can process.

Each milestone also needs to be specific enough that finishing it produces a detectable sense of completion, and it needs to be something you actually value. A milestone you could not care less about produces no signal worth acting on.

The second step is to make the future abstract reward concrete. "Finishing my PhD" is abstract. "Walking across a stage while my parents watch from the third row" is vivid. The more concrete and sensory you can make the future reward, the less steeply the brain discounts it.