Example · PhD, Computer Science and AI
Agents can replicate the baseline now. Judging whether a result holds up is still the core of a PhD.
23 of your 40 hours still need you
- 10 stay yours
- 13 faster with AI
- 17 tools could take
A typical week at the middle pace. Your plan uses your own tasks.
Is a PhD in AI still worth it?
In the work a CS and AI PhD student is heading toward, 23 of a typical 40-hour week still need a person, by our task-by-task count. Tools could already take about 17 hours. Coding agents now do a large share of the work that fills a second- or third-year PhD week: the related-work sweep, the baseline reimplementation, the training script, the sweep config, the grading. That is uncomfortable, because those are the tasks that taught the last generation of students to be researchers, and research-engineer postings are written around them.
Read more
What sets a PhD apart at the end is the part that comes next.
What holds is the part your advisor means when they ask what the numbers mean. You pick a question worth a year. You design the experiment that would change your mind. You know a run is wrong before the plot says so, and you put your name on the claim. The next 13 weeks turn one of the hardest parts of the PhD, the baseline that never matches the paper, into a public record that you can tell a sound result from a broken one. Agents make that skill more valuable, and few labs have someone who owns it yet.
The hard truth
Frontier AI models are now mostly built in industry.
If you want to train the largest models, most of that work now happens at companies with huge compute budgets. Academic labs still lead on the ideas the field cites most.
Sources (2)
- Nearly 90% of notable AI models in 2024 came from industry, up from 60% in 2023. Stanford HAI, AI Index Report 2025, 2025-04
- Academia remains the top source of highly cited AI research. Stanford HAI, AI Index Report 2025, 2025-04
What stays yours
Choosing which question is worth a year of your life
It depends on what your lab can run, what reviewers in your subfield will believe, what the field is about to make obsolete, and what you can still care about in month nine. None of that is written down.
Designing the experiment that would prove you wrong
An agent runs the experiment you ask for. Asking for the one that could falsify your idea is the research skill, and a wrong design costs months before anyone notices.
Knowing a run is broken before the curve tells you
Data leakage, a seed that flatters the method and an eval set that overlaps pretraining data are all easy mistakes to make and costly to find.
Your name on the paper
Venue policies make authors responsible for everything in a submission, including what a tool wrote. That responsibility does not move to the tool.
Mentoring an undergrad through a first project
Undergrads come to you because you sat in their seat two years ago. That trust is how labs keep running, and it is part of how search committees judge you.
What careers come after a CS PhD?
Plan A is your best move in the field you are studying for. Most sessions in the 13-week plan build toward it.
- A
Become the person in your lab who checks that results hold up
Why it fits
Why you. If you have fought with baselines that won't match, you already know more about broken replications than people who skipped them.
The gap. You do not yet have public reproducibility work, a systematic record of agent failures or an evaluation paper.
The trade-off. Reproducibility papers count for less than new methods on many faculty committees, and they can annoy the original authors. Do it carefully, contact the authors, and keep it alongside your thesis work rather than in place of it.
Other moves that fit
These are also good bets in this field. Read them if Plan A doesn't suit you.
Pick a thesis question that doesn't need huge compute budgets
Why it fits
Why you. Many students choose the second half of a thesis in the second or third year, and most academic labs cannot win scale races.
The gap. You need a conversation with your advisor and one pilot result.
The trade-off. Some of these topics read as less glamorous on the job market, and a collaborator in another field moves on their own clock.
Apply beyond the few big AI labs
Why it fits
Why you. The 61.7% of computing PhDs who go to industry (CRA Taulbee Survey, 2025) do not all go to a handful of labs.
The gap. You need one domain where you speak the language, and a network outside ML (machine learning) social media.
The trade-off. Pay at many of these employers is below frontier-lab offers, and research freedom varies by team.
What's a plan B for AI PhD students?
Plan B is a job AI is creating that these skills already fit. One session a week in the plan trains you for it, so it is ready if your job shrinks or you want a change.
BPlan B, picked for this role
LLM evaluation data scientist
Designing the experiment that would prove you wrong is what model evaluation teams hire PhDs to do every week.
Other options
Domain AI red teamer
Knowing a run is broken before the curve shows it helps you find the failure modes a model's builders missed.
AI policy analyst
Fellowships want people who can explain technical limits clearly, and defending your work when someone pushes back in a talk is practice.
In your own plan you can switch to any of the 47 roles.
The first two weeks
The first 2 of 13 weeks. Most weeks have three sessions of 45, 60 and 90 minutes, plus one 40-minute Plan B session. Weeks 6, 8 and 9 have one session plus the Plan B session, for Thanksgiving and finals. If your semester runs on a different calendar, keep the order and move the light weeks to your exams. By week 13 you'll have a reproducibility study submitted to TMLR with a public repo, a written audit of where a coding agent failed, and three internship applications sent.
- Week 1: See where your time goes
What you do
- AList a typical week of your classes, research, teaching and reading. Mark which parts AI already helps with.
- AList every replication you have tried. Next to each one, write the gap between your number and the paper's number.
- AWatch the Agentic Coding lecture from Missing Semester 2026, MIT's free course on developer tools. Pick three practices you will adopt.
- BTake Anthropic's AI Capabilities and Limitations course. Match each failure type it covers to one you have seen in your own experiments.
- Week 2: Pick a paper to replicate
What you do
- AChoose one paper in your area with released code and a reported number you can test on the compute you have. Write it up in a file called target.md.
- ADraft a request for a NAIRR Start-Up allocation from the national AI research compute program. Ask your advisor for the support letter it needs.
- ADecide whether to apply for the NVIDIA Graduate Fellowship. Check the deadline on NVIDIA's page, then start the application or write down why not.
- BTake the first half of Claude Code in Action, Anthropic's course on its coding agent. Practice on a public baseline repo and never on lab code.
Three courses from the plan
Checked 2026-10-13.
- NAIRR Pilot Start-Up projectNSF NAIRR Pilot · No fee listed
Yes, Week 2.
- TPU Research CloudGoogle Research · Free to accepted researchers
Yes, if your code runs on JAX or PyTorch/XLA.
- The Missing Semester (2026), Agentic Coding lectureMIT CSAIL · Free online; no fee listed
Yes, Week 1.
Risks beyond AI: highFunding and visa-policy risk
Federal science money and immigration rules moved more in 2025 and 2026 than in the decade before. Roughly half of CS doctorates go to nonresidents, so visa rules shape this labor market for everyone in it, citizen or not. The faculty market is tight on its own, and most new assistant professors arrive through a postdoc or another school.
Where this could be wrong
- The task blend is ours. Theory students spend far less time on experiment code; systems and robotics students spend more on hardware. Your own Week 1 log corrects it.
- The 93.0% PaperBench figure is a provider self-report in a lighter setting. If independent runs land far lower, we have overstated how exposed replication work is.
- Taulbee leaves 27.6% of 2025 graduates' outcomes unknown. Unemployment among them could be higher than the 0.7% among known outcomes.
- METR's time-horizon page is no longer actively updated, and its top estimate has an interval from about 8.5 to 55 hours.
Also matched: phd student computer science, cs phd, computer science phd. Research checked 2026-10-13. Page updated .