Universities have an interface problem
My academic career began pre-COVID with a lecture on engineering dynamics to ~300 first-year undergraduates. Two weeks later, the number fell to ~40 students. The dwindling numbers were a consequence of a delivery style mostly unchanged over a thousand years — the monologue lecture — and my inability to provide personalised feedback to each of the 300. Lectures, textbooks and exams are narrow interfaces — between an academic’s judgment and 300 minds — that need widening.
In 2021, I partially addressed the feedback problem with interactive online textbooks built on Jupyter Notebooks. This allowed personalised feedback, though only on code. Some resolved coding errors using Google or Stack Overflow whilst others came to my office hours for personal guidance, both valid approaches to develop understanding and judgment. But the textbook lacked technology to monitor student engagement and track learning progress. Even if this were possible, not all 300 students could receive my personal attention in office hours.
In the agentic era, I’m building better learning interfaces — for another class on orbital mechanics — that increase feedback quality and frequency to each student via a more personalised online textbook and, more importantly, during large lectures. The textbook now incorporates written technical reasoning questions alongside multiple-choice questions that appear at key junctures; students’ responses to them are fed to agents that identify reasoning errors rather than giving the answer. They evaluate responses against my benchmark answer instead of creating their own. Further, user accounts for the textbook allow students to continuously track their progress over repeated attempts of these questions, and also flag for me concepts they struggle with.
The same stack allows me to also TikTok-ify my lectures: I now break the two-hour monologue lectures into shorter segments interspersed with polling-style questions for students to respond to. Again, agents provide personalised feedback to each response while summarised performance data is projected in lecture halls. In my second lecture this term, 81% of a room of 66 chose the right option on a question but only 6% could reason why; each student got feedback in about six seconds, and I immediately retaught misconceptions before progressing to the next topic. AI implements such ideas cheaply while enabling me to more surgically address gaps in understanding from student data.
But the learning experience can be further improved. For example, students can currently highlight-and-comment portions of the textbook, à la GDocs. In the next work-phase, LLMs will ingest these notes to generate personalised in-book quizzes to allow targeted strengthening on concepts. In-class and in-textbook quiz responses could also be analysed together for each student by LLMs to present a more holistic image of learning progress before, during, and after lectures.
This approach scales to other courses and entire technical programmes culminating in a research project. Organisationally, it inverts the lecture’s hierarchy: agents handle routine marking and escalate what they can’t resolve (collectively misunderstood concepts) to me. Students exercise judgment because answers are never handed over, and the academic’s judgment isn’t relinquished to agents.
While this system makes lectures more interactive and materials more personal, it doesn’t guarantee attendance: 90 of 150 enrolled students attend my orbital mechanics course. This is an incentives problem: students don’t miss exams. If we believe better judgment is reached by robust engagement, then we must reward it in the room and textbook. Universities could grade continuous engagement, assessing students’ reasoning and responses to feedback through the same learning interface.
Such engagement-based assessment could even replace end-of-term invigilated handwritten exams, which exist because no human can continuously assess 300 students. That LLMs pass such exams easily is well known, so we should ask whether such assessments prepare students for life after university. Already in the workplace, LLMs are not merely an accelerant technology, like a pen, but partner humans in technical and non-technical work; the share of junior roles in Anthropic’s technical hiring fell from 27% to 9% in a year while non-technical staff began performing technical work with AI assistance. Universities must heed this warning and prepare students for a changing world that rewards agency and enterprise over an ability to take handwritten exams.
If invigilated assessments must remain, India’s IIT entrance exams show another form: over a million candidates a year answer multiple-choice questions that demand technical reasoning without sacrificing rigour, marked by machine. This works at hundreds-to-one student-staff ratios with nothing more than bubble sheets and a scanner, but one could administer these via computerised testing environments. This removes clerical friction with hand-marking hundreds of scripts, and frees academics’ time for more meaningful research tasks and student development.
A written component should still exist, but as a student project which builds something to potentially outlast their degree. How might this work in practice? Consider the orbital mechanics textbook; at the end of the course, it could respond to a student’s interest in space-station R&D to help them develop a project with instructor supervision. After the student works on this for a while, they might identify a venture-shaped idea. The in-textbook agents then help extend this work into a business pitch to raise funding for continued work after graduation, while the academic continues researching non-commercialisable aspects. This links to a brief mention in the Anthropic report of single-person startups with human founders steering agent swarms “to rake in billions”. Universities could follow this hint to incubate technical talent via projects, but incorporating agents across institutions will be essential to make headway; those that don’t might lose students to the places that do.
However, not all students will wish to pursue technical work. It is not hard to imagine entirely new customised programmes; for example, agents might help students with literary aspirations call on the university’s technical and humanities staff to develop hard sci-fi stories while also finding publishers. But to deliver here, we must institutionally commit to transforming our cathedrals of learning with agentic interfaces that are wider and more personal, broadening what graduates can become.