Jack Weis has a doctorate in chemical engineering from Princeton, a published paper that the Nobel committee’s background material for its 2024 physics prize cited, and a spreadsheet listing every job he has applied for since he finished his degree. This year the spreadsheet passed 2,300 applications. The frozen crab cake company on the list did not call back either.
The applications went to pharmaceutical research groups, to postdoctoral positions at national laboratories, to software companies and to consultancies, according to a report in The Wall Street Journal. Almost none produced an interview. The ones that did produced no offer.
By January he was earning $70 an hour grading answers produced by artificial intelligence systems, work that paid the rent until it ended. In March he was let go, and he moved into his parents’ basement.
Jinyan Su came to the same market with a résumé that should have been easier to sell. She is finishing a doctorate in artificial intelligence at Cornell and has been a visiting scholar at Stanford, with internships at Adobe, at Meta’s Fundamental AI Research group and on Microsoft’s Copilot team. Her work is in reinforcement learning and post-training, the part of model development that companies say they cannot hire for fast enough.
In interviews she was asked whether she had cheated, because she wrote code faster than the interviewer expected a candidate to write it. She describes the search as reinforcement learning with an unusual amount of noise, a process in which ability accounts for perhaps half of the outcome and the rest is drawn from a distribution nobody can see.
The explanation offered by people doing the hiring is structural rather than personal. A core engineering team that once ran with two senior architects and ten junior and mid-level engineers now keeps the architects and gives them clusters of agents to supervise. The junior requisitions are the ones that disappear, because the work those engineers did, writing the first version of a component and fixing what breaks, is the work that automated tools now do adequately.
That change lands on the people least able to absorb it. Engineers laid off at senior levels and new doctorates are now competing for the same smaller set of roles, and the overlap is uncomfortable for both. A candidate with fifteen years of experience and a candidate with a freshly printed thesis apply to the same posting, and the hiring manager picks the one who can describe a product that made money.
Publication count has moved down the list of what gets discussed in an interview, according to recruiters and to candidates who have sat through them. The question that decides the outcome is whether the applicant can take a capability and turn it into something a customer will pay for, which is a different skill from the one a doctoral program rewards and one that nobody teaches in a laboratory.
The spillover is visible in where the applications go. A chemist applies to software teams, a linguist applies to model evaluation, an economist applies to pricing roles at companies that did not exist when the degree was chosen. Each of those moves is rational for the individual and diluting for the market, because a hiring manager sorting a hundred applications cannot tell which of the crossovers is a genuine fit without spending time that the posting does not budget for.
Universities have responded slowly. A graduate program measures its output in papers, conference acceptances and placements into academic posts, and the labor market its students now enter values none of those as highly as it did five years ago. The advice available to a candidate is mostly about networking, which is another way of saying that the formal channel has stopped working.
The counterargument from inside the industry is that the shortage is real and the mismatch is on the other side. Companies say they cannot find people who can deploy models into an existing business, work with a customer’s data, and stay long enough to see a system survive contact with production. That is a job description, not a research agenda, and the people who can do it are the ones who already have.
None of this is confined to artificial intelligence. The entry-level role has been the training ground for every professional field, and the tasks that fill it, drafting, summarizing, checking, writing the first pass, are the tasks that current tools do best. A firm that stops hiring at the bottom saves money in the current quarter and removes the source of its senior staff in the one after that.
For Weis, the arithmetic of the search has become its own subject. Applications cost minutes and each one is cheap to send, which is why the number grew past four figures, and cheap applications are cheap to ignore. The system rewards volume on one side and filters on the other, and the equilibrium is a pile of documents nobody reads.
He is still in the basement, and he is still applying. The question he keeps running into is not whether the market is broken but which parts of it are repricing and which parts have simply stopped buying what a doctorate used to signal.
