This week's theme is what holds a university together. Jill Lepore asks her own university what it is for. Harvard's dean, David Deming, has his answer: a community that makes you do hard things, together. Dan Rockmore answers with the gym. Tsinghua, meanwhile, puts “cognitive autonomy” at the center of AI general education, and Andrew Ng says again that everyone should learn to code. A methods paper asks how to trust AI-made data. The issue closes with two researchers asking who gets credit when an AI does the synthesis. One of them, a Peking University–trained engineer, watched an AI prove, in about an hour, the conjecture he named nine years ago.
Policy
In a Harvard Crimson op-ed, historian Jill Lepore contrasts two neighbors. MIT's August committee report on AI in teaching warns that “many uses of AI deprive students of the opportunity to learn.” Harvard College Dean David J. Deming, by contrast, recommended that student AI use in writing-intensive classes be not only accepted but encouraged. Lepore borrows Daniel W. Hook's “Waymo Effect”: we experience the removal of human friction “as pure gain,” because its costs were visible and its benefits were not. Her example is instructors replacing undergraduate research assistants with AI. Harvard, she notes, lets every student and faculty member request a Claude token budget of $2,500 per month. More than a third of Harvard undergraduates polled last spring used AI in ways that violated course policies. She ends by repeating MIT's question: “What is it that we are all here together to do, again?”
Deming had already made his case at length on Derek Thompson's Plain English podcast. “Encouragement,” he explains, is one end of a “barbell strategy.” Where a ban can be enforced, ban AI. Where it can't, encourage AI and build guardrails, because an unenforced ban rewards students who break it and leaves the honest ones feeling “punished for doing the right thing.” In his own course, students will commit to first ideas in blue books, use AI for feedback and research on later drafts, then present to him “so I can verify that you actually did the work yourself.” He admits “it's costlier. It may mean classes need to be smaller.” His larger argument is historical. Harvard began in 1636 in an era of scarce information, then moved into an era of expertise, and is now returning to the medieval “collegium,” where some skills and virtues “can really only be built in community.” You cannot learn integrity, he argues, unless it is tested among other people and costs you something. College is a commitment device, a way to “do hard things, suffer together.” He even suggests everyone should take theater classes.
Lepore's sharpest point is about research mentoring, not cheating. Supervising an undergraduate researcher is slow, costly and uneven, while AI is fast, cheap and often better. That is exactly the calculation that could quietly shrink undergraduate research at DKU, where mentored research is central to Signature Work. Deming's course design gives DKU faculty a ready template: a first draft written in class, AI-assisted revision, then an oral check. His caveat that it needs smaller classes plays to DKU's strengths. Lepore also names what is missing: a policy process that is “meaningfully debated” and “answerable to evidence.” The MIT report she praises (featured in Issue #7) is a good model for how DKU might write its own.
Sources: “Artificial Education” — Jill Lepore, The Harvard Crimson, September 16, 2026 · spotted via @bronwenmcshea on X · “What's the Point of College?” — Plain English with Derek Thompson (guest: David Deming), The Ringer, September 11, 2026
Pedagogy
Dan Rockmore teaches mathematics and computer science at Dartmouth. In a New Yorker weekend essay, he notes that nobody outsources a spin class, even though machines move us more easily. A bench press is a proxy for strength, and he argues that problem sets and close readings “were always proxies, too.” The difference is feedback. In the gym you feel yourself getting stronger, while in the classroom “progress can be difficult to see.” So he points to teachers who make that progress visible. One colleague lets students retake exams “until the concepts finally click.” Poet Kyle Booten (UConn) builds tools that make writing harder, collected in a book called Gyms. Rockmore's division of labor: “Let the machines do what's mechanical,” and let teachers draw out the quiet student and build community. His question: “How could we build spin studios for the mind?”
This is the hopeful answer to item 1, and Deming reaches for the same image, borrowing a colleague's line that education “should be like a mental gym.” If students treat homework as a transaction, the fix is not better detection. It is making effort feel worthwhile and progress easy to see. Two ideas travel well to DKU's seven-week courses: exams students can retake until the concepts click, and exercises with added resistance, like Booten's. Rockmore's gym has one more lesson. It works partly because people sweat together, and a small residential campus is well built to be that kind of room.
Source: “Should the Classroom Be More Like the Gym?” — Dan Rockmore, The New Yorker, September 19, 2026 · spotted via @newyorker on X
Curriculum
On September 9, Tsinghua's Department of Computer Science and Technology released a red book, Artificial Intelligence General Education: Cultivating Cognitive Autonomy in an Intelligent Society (《人工智能通识教育：智能社会中的认知自治能力培养》), first-authored by Wang Dong with Ma Shaoping. The central claim is that the goal of AI general education is not tool operation but “cognitive autonomy.” That means actively regulating how you form beliefs, use evidence, allocate trust, and take responsibility when AI is part of the process. It is organized around four capabilities: induction, generalization, judgment and self-awareness. The book also lists seven ways AI education goes wrong. Two are reducing it to tool training, and treating ethics as a tacked-on warning. It comes with 16 graded readers for primary through high school, tested in more than 300 practice schools.
The name will ring a bell for anyone who knows Harvard's 1945 “Red Book,” General Education in a Free Society, and the ambition is similar. For DKU, which sits between Chinese and American traditions of general education, this is a useful counterweight to the tools-first AI literacy courses spreading everywhere. Its six questions to ask of any AI output (where does it come from, what's the evidence, does it apply here, how much should I trust it, who is responsible, have I outsourced my thinking?) could go straight into a first-year seminar. The full book is free on Zenodo.
Sources: 《人工智能通识教育：智能社会中的认知自治能力培养》 — Wang Dong & Ma Shaoping, Zenodo (full PDF) · Launch report — 腾讯新闻, September 11, 2026 · 《清华大学发布〈人工智能通识教育〉红皮书》 — AI瞭望星球 (WeChat summary)
Skills
A video clip of Andrew Ng circulated this week, summarized in Chinese by the AI engineer 宝玉 (@dotey). Ng repeats an argument he has made since last year. Some executives tell people not to learn programming because AI will automate it, and he thinks that gets it backwards. As he wrote in March 2025, “As coding becomes easier, more people should code, not fewer!” In the new clip, by 宝玉's account, Ng says he almost never writes code by hand. Learning to code, for him, means learning to state what you want in a language the computer understands. His best marketer builds websites without waiting for engineers, and his best recruiters write programs to screen résumés. In his 2025 words: “one of the most important skills in the future will be the ability to tell a computer exactly what you want, so it can do that for you.”
Read this next to item 3. Tsinghua's red book explicitly warns against reducing AI general education to algorithms and programming (技术化). Ng is making a narrower claim, that coding is a language of precise requests, like the vocabulary of art history he credits for better image prompts. Both can be true. For DKU, the practical question is whether every student, and every staff member, gets one real chance to build something small with AI-assisted code. Ng's marketer and recruiter examples apply as much to our offices as to our classrooms.
Sources: 宝玉 (@dotey) on X, summary and video clip, September 24, 2026 · Andrew Ng on X, March 13, 2025 (The Batch, issue 292)
Methods
This review from Melissa Dell (Harvard) and Ashesh Rambachan (MIT) accompanies the 2026 NBER Methods Lecture. It starts from a shift every social scientist will recognize. AI models “convert unstructured data, such as text and images, into structured variables at low cost.” That makes measurement at scale feasible when it once was not. The bottleneck is no longer finding any scalable measure. It is “choosing among many plausible ones, which may support different empirical conclusions.” The authors describe three stages at which AI enters the measurement pipeline: discovery, construct definition, and observation. They argue that credible inference with AI-generated variables requires deliberate validation. That means “anchoring measurement to explicit criteria, rather than informal claims that a proxy is reasonable.” They also show how a validation sample supports valid inference “even when AI predictions are arbitrarily biased.”
DKU students increasingly use LLMs to code interviews, news articles or social-media posts for Signature Work and methods courses. Usually the check is a quick look to confirm the labels seem right. This paper gives instructors vocabulary and a standard: pick the construct explicitly, hand-code a random validation sample, and report how the AI measure performs against it. That is a teachable unit on its own, and a natural companion to the computational social science program in Issue #10.
Source: “The Measurement Revolution? Credible Measurement and Inference in the Age of AI” — Melissa Dell & Ashesh Rambachan, NBER Working Paper 35744 · spotted via @nberpubs on X
Research culture
Toru Lin, a roboticist at Amazon's Frontier AI & Robotics lab who did her PhD at Berkeley with Jitendra Malik, watched an AI research agent called Astra autonomously build a dexterous-manipulation result. She recognized ingredients from her own field in it. It “surveys, understands, adapts, integrates, and experiments,” she writes, but leaves “no record” of the work it built on. “Academia runs largely on one incentive: having one's name attached to what one contributes,” and she doesn't think it “can be sustained on goodwill and sacrifice alone.” Malik agreed publicly, noting he co-authored in-hand rotation papers that Astra may have used: “Should we not accuse these AI models of plagiarism if they don't cite their sources?” If incentives collapse, he warns, “it is like farmers eating their seed corn.”
Mathematics had its own version this week. Zhou Longfei did his PhD in convex optimization at Peking University and has worked in industry since 2018. For nine years he kept returning, as a hobby, to a problem he had named in a 2017 Zhihu post. L.G. Khachiyan proved in 1990 that the inscribed-ellipsoid method shrinks volume by a factor of at most 0.844, and conjectured that the best constant is √e/2 = 0.82436…. On September 18, Zhou typed a prompt into Codex asking gpt-6-astra to beat 0.844, then walked away. Just over an hour later it had written a report proving the conjecture outright. On September 23 he and two co-authors posted it to arXiv, with the main theorem checked in Lean 4. Zhou notes that the AI proved a key lemma “without citation.” Only afterward, by asking another model, did he find the related prior work on submodular spectral functions from around 2010.
See Figure 3 in Zhou's post: two upper-bound curves whose lower envelope peaks exactly at √e/2.
We spend four years teaching students that citation is how scholarship keeps track of who contributed what. Lin's phrase is “intellectual lineage.” Both stories turn the familiar “don't plagiarize” rule into a design problem: if an agent reads 10 papers and builds on them, what should it cite, and who is responsible? Zhou's case also shows how much human mastery the result still needed. He had to know the problem existed, recognize a proof when he saw one, and spend the night working through it, and his co-authors then had to verify it formally. His arXiv paper offers DKU students a model disclosure: “An AI language model discovered the proof in a human-directed research process.” Together with the Lean check, it says what the machine did, what the humans did, and how the result was verified. Malik, for his part, cites the Fields Medallists' letter from Issue #10.
Sources: “What Next for Academia?” — Toru Lin, September 18, 2026 · spotted via @JitendraMalikCV on X · 《gpt-6-astra证明了我喜欢的Khachiyan常数猜想》 — Zhou Longfei, Zhihu, September 2026 · “A Spectral Proof of Khachiyan's Ellipsoid Conjecture” — Zhou, Zou & Liu, arXiv:2609.28447 · Lean 4 formalization on GitHub
I like the Deming conversation a lot. His idea that community matters more in the age of AI, not less, resonates with me. When expertise is no longer scarce, what can college really offer? Deming himself studies meetings, but I believe there is much more to it than that. Group activities, sports and clubs are all part of it, and they are exactly what we see DKU bring to our students.
— Yisu