Two PhD students walk into a bar – both are working hard. One is waiting for his agentic system to clean up the latest SNP analysis and translate point estimates into a Bayesian statistical framework. The other is currently reviewing the review article that her agentic system compiled across the last 5 years of research in topical delivery systems.
Most of my work, many years ago, in my PhD was around: cleaning data, writing code / scripts to do descriptive summaries of data, and reading / compiling primary research. Once all of those things had been done I had time to think. What does this mean, what would be the next logical question, how do I design an experiment around this, is my model correct?
Now all of this grunt work is more or less automated.
I have written the data cleaning agent. Two million rows takes nine minutes, delivers a full anomaly report, and lets me stay in the loop on the judgment calls that matter. The code writing is largely gone too — not the thinking about what to build, but the syntax wrestling, the refactoring loops, the three hours lost because you changed direction. My natural language analyst is now coupled to custom storage and re-run mechanisms. Same insights, multiple datasets, no ceremony.
Literature is the one that surprises people. You still need to read. But scanning a hundred papers to find the ten worth going deep on? That is just pattern matching at scale. My research agent is coupled directly into PubMed. Let it skim abstracts. Let it summarize the landscape. Then you show up to the interesting part.
And the writing. I remember those first methods sections. The blank page in a second language is its own particular hell — not because you lack the ideas, but because formulating them precisely in English costs more energy than it should. Now I get a first draft in minutes. Reacting, refining, re-writing is easier than starting from nothing. I have fed the writing agent enough from my field that it knows the style. I can even dial it to the layout of different journals. No more reformatting grief every time a rejection means starting over somewhere new.
So what is left? Figuring out what are the interesting research questions, formulating what needs to be investigated. Looking into the non-obvious patterns, asking the left-field questions that occur when you see enough literature and data pointing in different directions. All the things that were limited because you were dealing with structuring gigabytes of genomic data, or figuring out where you had a bug.
Sure there is value in grunt work. It teaches you grit, structured approaches to code writing, gives a sense of accomplishment, and delays gratification. All things I am happy to have learned. Just as my granddad was very happy that he was fast at doing multiplications in his head, and taught me how to do it during summer vacations. But I rarely do it anymore. Likewise, the stuff that formed me as a PhD will soon be things that future PhDs rarely do. The field is moving forward. Maybe they will have more time to test 10x the number of hypotheses I did, maybe they will be able to cover much more research and draw many more disparate connections to accelerate their fields. Biology students might not need to stand in line to talk to a bio-informatician to look at their lab work.