This week's theme is what is it for? Becca Rothfeld finds that question on the lips of nearly every AI critic she meets, from a mother fighting screens in schools to a UChicago professor defending the seminar. A long Zhihu answer turns it into a practical one: which layer of thinking can a student hand over, and which must stay? Brandeis answers it for the liberal arts with a second transcript. Two mathematicians, Wei Zhewei in Beijing and Christian Szegedy in the AI industry, ask what a mathematician is for once machines prove theorems. And an Anthropic engineering lead offers a tutorial on asking AI for the goal beyond the task.
Ideas
Becca Rothfeld went to the Pro-Human Assembly in Washington, an AI-critical gathering headlined by Bernie Sanders and Steve Bannon, and found that attendees talked less about jobs and regulation than about “what they thought a human life was for.” Commentators call this “A.I. populism.” Rothfeld calls it “anti-A.I. humanism,” a coalition of teachers, doctors, artists, religious communities and the young. Education runs through the piece. Lila Byock of Schools Beyond Screens says AI “is designed to make tasks easier and to reduce friction, and friction is how learning happens.” Jenny Trinitapoli, an architect of UChicago's Social Sciences Core policy (Issue #7), says the rules let faculty ask “What are we doing in these small seminar courses? What are they for?” And a UChicago senior, Owen Yingling, explains why the young are most skeptical. His parents have already become who they are. “But for us, because we haven't done anything or become anyone, it's a sort of existential threat.”
Our students are Yingling's age, and some of them share his view. A campus AI policy that treats every student as an eager adopter misreads the room, and so does one that treats every skeptic as a technophobe. Rothfeld's refrain, “what is this activity for?”, is also the right first question for any course redesign. Trinitapoli's answer, “the habits of mind that our students are going to have for the rest of their lives,” is the kind a seminar at DKU can give honestly, and a prompt-engineering unit cannot. Note that this is reportage from one conference and a round of phone calls, and most of the voices are American.
Source: “The Humanism Driving the Anti-A.I. Movement” — Becca Rothfeld, The New Yorker, September 30, 2026 · spotted via @afrazhaowang on X
Evidence
Does leaning on AI weaken independent thinking? A long Zhihu answer by the writer LostAbaddon, who has worked on an AI-and-learning project in Chengdu since last year, reframes the question: which layer, handed over at what moment, weakens which layer (把哪一层，在什么时机交出去，会削弱哪一层). Using Bloom's taxonomy, he argues that outsourcing recall costs little, outsourcing reasoning starts to hurt, and outsourcing problem definition and judgment hurts most. The calculator analogy fails because a calculator replaces one sub-task while generative AI hands you a finished, confident argument. His anchor is a working paper by Strömberg, Lei and Wu that followed 26,811 secondary students in a Chinese county for 30 months. Homework scores rose 18%, completion time fell 30%, and closed-book monthly exam scores fell 20% within six months. About 81% of users were “outsourcing homework” (作业外包); the roughly 19% who used AI as a tutor lost nothing. His rule: what is already internalised, AI amplifies; what is still forming, AI cuts off (已经内化的部分，AI 能放大；正在形成的部分，AI 会截断).
This is a synthesis, not a study, and the headline numbers come from a difference-in-differences paper on exam-oriented schooling, which the author himself says is not a randomised trial. It is here because the layered framing is immediately usable. For each assignment, ask which layer it is meant to train and design the AI rule for that layer alone, rather than one policy for the course. The author's design principle, “let the friction happen before the AI output,” echoes last issue's Bocconi and Harvard trials. His closing warning is the one to sit with: AI will not weaken everyone's thinking evenly, it will widen the gap between active thinkers and passive consumers.
Source: 《过度依赖AI会削弱人的独立思考吗？》 — LostAbaddon's answer on Zhihu (date not shown to automated readers) · the student study as reported in 《他们追踪2万多中学生发现，AI会让考试成绩大幅下滑》, 上海科技报
Institutions
Jon Marcus of the Hechinger Report profiles Brandeis University's answer to the question families now put to liberal arts colleges. Under president Arthur Levine, a long-time scholar of higher education, Brandeis is reorganising into four schools that pair professional programs with liberal arts disciplines, assigning every arriving student a career counsellor alongside an academic adviser, and issuing a “career-competency” transcript listing the skills employers value, “from communicating and writing to proficiency with AI,” plus internships and microcredentials. To fill it, the university is piloting an ETS tool that “uses AI to measure what students know and can do.” Levine's diagnosis: “we have tended to teach the liberal arts at many institutions as if they were something apart from the real world.” The faculty voted overwhelmingly for the plan. Its senate chair, Jeffrey Lenowitz, explains why: “If employers are saying, ‘You need critical thinking and communication skills,’ well what do you think a seminar is?” Bachelor's degrees in the liberal arts fell to a 30-year low in 2024.
DKU makes the same promise Brandeis is trying to make visible, and the question is whether we want to make it visible the same way. Lenowitz's point is the useful one: the skills were always taught, they were never named. A lighter version of the second transcript is within reach here, for instance asking each course to state in its syllabus which durable skills it trains and how students show them. The caution is in Marcus's own reporting. Marcus notes the skepticism that these skills can be measured at all, and the measuring is being handed to an AI product sold by a testing company. Richard Detweiler's line applies to us too: “I agree with the argument, but they haven't made the case.”
Mathematics
Wei Zhewei is a professor and vice dean at Renmin University's Gaoling School of Artificial Intelligence. His essay for 知识分子 opens with his own case. Tusnády's problem in discrepancy theory had a gap between its bounds that stood for more than forty years. This year, in continuous exchanges with GPT and Claude from July, he closed it with an elementary proof, now on arXiv with a Lean-verified lower bound. He posed the problem, judged directions and checked the argument; the core idea formed in the exchange. It is already hard to call this “assistance” (辅助), he writes. His thesis is that proof is search (证明即搜索), and a mathematician's intuition is a prior over where to search, learnable from the literature. Society supports mathematicians to get theorems, not to watch humans prove them, so he expects mathematics to gain while mathematicians lose, as translators did. He separates results that are usable (可用) from those that are understood (可理解), and lists what humans still add: posing problems, and checking, selecting and digesting results.
See Figure 4 in the essay: the two translation steps between a theorem in a paper and a Lean proof, and why passing Lean does not settle the original claim.
Wei's list of what remains human is a syllabus. Posing a good question, verifying a machine's claim, choosing which results matter and digesting them into understanding are teachable, and they are what a mathematics or statistics seminar at DKU should now spend its time on. He adds a warning for the rest of us: a readable explanation is not understanding, because the hard part is what the summary omits. He also names teaching and keeping human judgment intact as reasons to support mathematicians that must be argued on their own. His closing question is the week's theme in one sentence: if every problem has an answer one day, will we still have the eye to tell which answers are worth treasuring? The OpenAI claims he discusses are reported as he describes them.
Source: 《数学的黄金时代与数学家的失落时代》 — 魏哲巍, 知识分子 (WeChat), October 12, 2026 · “Tight Bounds for Tusnády's Problem in the Plane” — Zhewei Wei, arXiv:2610.08130, October 6, 2026
Mathematics
The AI researcher Christian Szegedy wrote his reply to recent essays by Terence Tao, Kevin Buzzard (Issue #12) and Timothy Gowers as a personal story. He grew up in a mathematics-specialised class in Hungary, with eight hours of mathematics a week and camps devoted to Olympiad problems. When the final exams ended, so did the competitions, and he “didn't know who I was without it.” He expects many mathematicians to feel that now. His picture of pure mathematics is an unmapped jungle whose explorers are revered for climbing rocks nobody could climb. AI is the airplane, still primitive and “obscenely expensive,” but improving fast. “Do not bet against the airplane,” he writes. “Anyone who stakes their identity on ‘AI can't do task X’ is building on sand.” Universities and tenure committees will look the same for years, but what they reward will shift toward “choosing which terrain is worth exploring, steering fleets of airplanes, and explaining what they find to the rest of the world.” “That grief is real,” he concludes. “But it will not last.”
Read beside Wei, this is the optimist's version of the same forecast, and the two agree on the skills that survive: choosing problems, directing tools, explaining results. Szegedy's own story is the part for us. A student who has built an identity on being the one who can solve the problem will feel the loss he describes, and sooner than tenured faculty will. Advising conversations should expect that. The essay is a personal view from inside the AI industry, with no evidence beyond the author's reading of recent results, and his claim that today's skeptics are “describing last spring's models” is a prediction, not a finding.
Source: “Is Mathematics Over, or Just Graduating?” — Christian Szegedy, published Google Doc, October 3, 2026 · spotted via @ChrSzegedy on X
Practice
Felix Rieseberg leads engineering for several Claude products at Anthropic. His tutorial admits that “be more ambitious” is tricky advice to follow, and breaks it into four kinds. Speed is the same work in less time. Altitude asks “What's the goal beyond the goal?”: he used to ask Claude to fix individual crashes, then asked it to keep the apps crash-free, which meant first giving it crash reports, version history and test machines. Depth means using the methods of the best people in a field, not their vocabulary: “models often switch vocabulary without actually switching methods.” Range means bringing in fields you could not previously afford, found by replacing nouns with verbs: a quarterly review is not a deck but getting busy people to decide under uncertainty in 30 minutes. One rule recurs: “Do not ask the model to judge itself.” The Chinese post that brought it to our inbox adds a gloss: many AI workflows just bolt a faster engine onto the old way of working (给过去的工作方式装上了一台更快的发动机).
This is advice for professionals who already know their field, and item 2 says why that matters: what is internalised, AI amplifies. For faculty's own work, course design and research, it is a good checklist, and “altitude” applied to a syllabus is the week's theme again: what is this course for? Two of his rules transfer straight to assignments. Require an outside check, a dataset, a test or a reader, rather than the model's own verdict. And ask students to say what a top professional would do that they are not asking for. The “depth” prompt he shares could open a Signature Work meeting. For first-year students still forming the skill, the Zhihu answer's caution comes first.
Source: “How to be more ambitious: A tutorial for humans with models” — Felix Rieseberg, October 13, 2026 · spotted via @sleepy0x13 on X
It surprises me how quickly the world of mathematics has been turned upside down in the past few months, even faster and more furiously than the economic pressure AI is putting on universities. From all I can read, no one has an answer yet. The ground is shifting beneath our feet. I suggest we keep our eyes and our minds open, because the old ways of doing research and teaching are bound to change.
If machines will one day have the answers, our job may be to keep the eye for which answers are worth treasuring, and to pass that eye on to our students.
— Yisu