Why Do Research

This track teaches the how of research: how to find the frontier, how to bet on what scales, how fast to run. But the how is worthless if the why underneath it is wrong. Method mounted on a bad motive does not repair the motive — it just delivers you to the wrong destination with excellent technique. So before the first principle, we settle the prior question, the one that never appears in a methods track because it sounds like a life question rather than a research question: what are you working for?

It has stopped being a life question. In 2024, The AI Scientist ran the entire research loop with no human in it — generated the ideas, wrote the code, ran the experiments, drafted full papers, and reviewed them — at a cost of under fifteen dollars per paper. Set the quality debate aside; the direction is what matters. The ordinary paper — known problem, incremental delta, competent write-up — is becoming an artifact that machines produce in bulk at near-zero cost, and cheaper every quarter. In that world, what am I working for is not a question you answer once in a personal statement. It is a question you must keep asking, continuously, because every year the set of answers that survive it gets smaller.

If the Answer Is “the Degree,” Quit Early

Run the audit now, before the sunk costs accumulate, and start with the most common answer, because it is also the worst one. If you are doing research for the degree — the title, the line on the CV, the door it is supposed to open — my advice is not to reflect on it, and not to manage it. It is to quit, as early as possible. I mean this as concretely as anything else in this track, and the reason has two halves.

The first half is about you: you will be miserable. Research is a sequence of long, uncertain bets, most of which fail, and the only fuel that survives that failure rate is caring about the question itself. If the degree is the point, every failed experiment is pure cost — nothing was learned that you wanted to know — and you will spend years doing the psychological equivalent of holding your breath. Nobody produces meaningful research from inside that state; misery is not just the side effect of the wrong why — it is the mechanism by which the wrong why destroys the work.

The second half is about the work, and it is the half that has changed. Degree-driven research converges, with almost physical inevitability, on one artifact: the minimum publishable unit — a known problem, a small delta, a safe venue. We are living in the age of the mass-produced paper, and that is precisely the artifact agent pipelines now generate end to end for the price of a sandwich. Spending your best years hand-crafting it is not a strategy that fails in some distant future; it is competing, today, against automation on the one axis where automation already wins. It is a waste of your own time, purchased at the highest price you will ever pay for anything.

Watch the mechanism do its work: two students, ten years, and — by construction — exactly the same luck.

What it models. A decade of research as a sequence of monthly bets, run twice with the same luck. One coin per month decides whether the experiment fails or progresses, and the identical outcome is applied to both lanes — amber works for the degree, blue for curiosity and the skills. Both tanks drain a little every month; a failed month is where the lanes split. For amber it is pure cost, extra drain — nothing was learned that it wanted to know. For blue it is mild recovery — the same failure taught it something it came to learn. Shipped papers refill both tanks, amber’s far more, because papers are the only thing amber is here for. The trap is the dotted line: below it, low fuel degrades the quality of the work, which slows progress, which pushes the next refill further away — misery is not the side effect of the wrong why, it is the mechanism by which the wrong why destroys the work. At zero, the lane quits.

Knobs. How brutal the field is sets the fraction of months that fail — the field’s true failure rate, applied to both lanes equally. Step advances a single month; Reset draws a fresh decade of coins.

Try this. Watch the first year and a half at defaults: same coins, green flags landing on the same months — the first paper ships simultaneously in both lanes, which is exactly why the wrong why goes undetected until the costs are sunk. Keep watching as amber crosses the dotted line: its next paper drifts later, the delayed refill lets it sink faster, and the spiral usually closes somewhere in years two to four — while blue, on the very same coins, wobbles and recovers all the way to the end of the decade. Then push brutality to 90% and reset a few times: amber’s quit races forward toward year two, but blue never collapses — its papers just come slower. That asymmetry is the point: brutality only sets how fast the wrong why fails; the why itself decides whether it fails at all.

Quitting early is the cheap exit

There is no shame in the decision — the shame is in refusing to make it. Leaving in year one costs a year. Leaving in year five costs several of the most important years you will ever have. Not leaving at all costs the most: a career of papers nobody needed, produced in a state you hated. If the audit says go, going is the intelligent move, and it is never cheaper than it is today.

A Bet, Priced Honestly

Now suppose the degree is not your reason. You still owe yourself an honest price sheet, because a PhD is not a job and it is not school. It is a bet: enormous fixed cost, uncertain marginal return. The stake is several of the most important years of your life — the years of maximum energy, maximum optionality, maximum compounding — pushed onto a single table in one of the highest-variance domains in existence. Placing that bet can be right. Placing it without knowing what you are buying is never right.

So know what you are not buying first. Nobody guarantees you that a PhD leads to a better job — not your advisor, not your university, and certainly not the market. The Economist ran the economics fifteen years ago and the verdict has only hardened: the earnings premium of a PhD — measured against those who could have gone to university but chose not to — is about 26%; a one-year master’s captures almost all of it at 23% (a bachelor’s alone already commands 14%), and in several fields — mathematics and computing among them — the doctorate’s edge over the master’s vanishes entirely. Meanwhile the academic career the training nominally prepares you for is structurally oversupplied: the system mints graduates far faster than it mints faculty positions, which is exactly why it can run on cheap, motivated, disposable labor. If money is what you want, the master’s holder who started working five years earlier is ahead of you and staying ahead.

What it shows. The price sheet for the credential: over the earnings of those who skipped university, a bachelor’s commands a 14% premium, a one-year master’s 23%, and the doctorate — several years dearer — 26%: a marginal edge of about three points, and in maths and computing effectively nothing. Source: The Economist, “The disposable academic” (2010), drawing on Bernard Casey’s analysis of British graduates.

Then price the health line, because the base rates are genuinely bad. Levecque and colleagues measured 3,659 PhD students and found roughly one in three at risk of a common psychiatric disorder — depression above all — significantly worse than highly educated employees or other higher-education students. Evans and colleagues, surveying 2,279 graduate students across 26 countries, found anxiety and depression at more than six times the rate of the general population. These are not anecdotes about students who weren’t tough enough; they are the best-measured rates we have for the environment you are proposing to enter. The bet is collateralized by your mind and your body, not just your CV.

What it shows. The best-measured rates we have for the environment: among 2,279 graduate students across 26 countries, 41% screened at moderate-to-severe anxiety and 39% at moderate-to-severe depression, against roughly 6% in the general population (GAD-7 / PHQ-9; Evans et al., Nature Biotechnology, 2018 — a self-selected online sample, not a population census: Evans et al. note respondents with a history of anxiety or depression may have been likelier to respond); on a separate instrument, 32% of 3,659 Flemish PhD students — one in three — crossed the GHQ4+ threshold for risk of a common psychiatric disorder, versus 14% of the highly educated general population (GHQ-12; Levecque et al., Research Policy, 2017, drawn from a full survey of Flemish PhD students, which points the same direction).

The Two Reasons That Survive

After the audit, exactly two motivations are left standing. The first: you want the research experience and skills themselves. You want to learn to take an open problem nobody can hand you the answer to and push it to a result, and you want that ability at the depth only years of practice can build. That skill is real, transferable, and — in the agent era — appreciating, because it is precisely the part of research the pipelines cannot yet do. The second: pure curiosity — love of the science itself. You would think about these questions anyway; the PhD just supplies the license, the colleagues, and the compute. This is the motive Karpathy’s survival guide assumes on every page: the case he makes for the PhD is the freedom to attack hard open problems and the transformation into someone who can — never the credential.

If neither of those is honestly yours, do not start. Not “start and see” — do not start. You will not make money doing this, and, doing it for a reason that cannot carry the weight, you will wreck your physical and mental health. Those are not warnings; given the numbers above, they are forecasts.

The Why Test

All of it compresses to a single question. Each principle in this track gets a test; this is the one that comes before all of them, and the one to rerun most often — every semester, and again every time the field lurches:

The Why Test. If the degree disappeared tomorrow, would you still choose this work?

If the answer is no, you now know it years earlier than most people ever learn it, at the lowest price the lesson will ever be offered. That is a good outcome — go build something you would choose. If the answer is yes — if the work itself is the point — then you are exactly who the rest of this track is written for, and everything downstream is craft: where to stand (Work on the Frontier), what to bet on (Bet on What Scales), how fast to move (Research Is a Race) — and, at the end, the person it takes to sustain all three (The Complete Researcher).

References

• The Economist. The disposable academic: Why doing a PhD is often a waste of time. December 16, 2010. economist.com
• Katia Levecque, Frederik Anseel, Alain De Beuckelaer, Johan Van der Heyden, Lydia Gisle. Work organization and mental health problems in PhD students. Research Policy 46(4): 868–879, 2017. doi.org/10.1016/j.respol.2017.02.008
• Teresa M. Evans, Lindsay Bira, Jazmin Beltran Gastelum, L. Todd Weiss, Nathan L. Vanderford. Evidence for a mental health crisis in graduate education. Nature Biotechnology 36(3): 282–284, 2018. nature.com/articles/nbt.4089
• Chris Lu, Cong Lu, Robert Tjarko Lange, Jakob Foerster, Jeff Clune, David Ha. The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery. arXiv:2408.06292, 2024. arxiv.org/abs/2408.06292
• Andrej Karpathy. A Survival Guide to a PhD. September 2016. karpathy.github.io/2016/09/07/phd