According to the primary source, an astrophysics assistant professor assigns projects to new PhD students Alice and Bob that the professor estimates would take one to two months personally but are intended to take students one year as a training vehicle. Alice reads papers with notes, gets stuck on coordinate systems and plotting, and receives standard feedback on checking units and intermediate outputs. Bob encounters similar issues with likelihood functions and sign conventions. (https://ergosphere.blog/posts/the-machines-are-fine/)

The post states that Bob used an AI agent to summarize papers, explain statistical methods, debug Python code, fix introduced bugs, and write the paper itself, yet his weekly updates, questions, and final paper matched Alice's. Both produced solid, publishable papers that passed minor revisions at a journal. By quantitative measures such as papers produced, the students are interchangeable per the source. (https://ergosphere.blog/posts/the-machines-are-fine/)

Per the source, most PhD students leave academia within a few years, so from the institution's view the question of deeper understanding is largely irrelevant as the department needs papers to justify funding. The post notes the academic evaluation system counts measurable outputs while the key element of assembled understanding cannot be counted. (https://ergosphere.blog/posts/the-machines-are-fine/)