Colleagues — a double-week issue, and the through-line is abundance and its digestion. AI is flooding the pipeline: proofs mathematicians haven't verified, tasks workers borrow from other professions, answers produced at whatever “effort level” the dropdown says. The scarce thing, everywhere below, is the slow human work of verifying, understanding, and absorbing. Terence Tao gave that scarcity a name this week; the rest of the issue keeps circling it.
1 · Tao at the ICM: from proof scarcity to proof abundance
Terence Tao's public lecture at the International Congress of Mathematicians (July 24), “Mathematics in the age of AI,” asks the community to reason conditionally: assume AI will “reasonably soon” handle a reasonable fraction of research-level mathematics — then what? His answer reframes the field as a pipeline: proof generation → verification → exposition → publication → canonicalization into textbooks. AI is accelerating the first stages; the last and slowest, canonicalization, is the one he calls “the most valuable part of the entire process.” The evidence he permits himself: the independent First Proof challenge, where on May 28 seven of ten novel research problems were solved at publication-level quality, at compute costs of $10–$1000 per problem. His warning: “we will transition from an era of proof scarcity to an era of proof abundance” — and indigestion is already visible in dozens of unvouched-for AI proofs on the Erdős problems site.
Why it matters for us: Tao says the quiet part about teaching directly: “in many areas (particularly in education and training), it will be crucial to emphasize the human aspect of our work, and tightly restrict the use of AI tools” — even as researchers adopt them on their own terms. And his publication rule of thumb — if you can't give a clear, expert-level talk on your result, it shouldn't be published — is an oral-defense principle a small college can apply tomorrow, from capstones to signature work.
Read: the full slides · spotted via @AlexKontorovich and @lmthang on X
2 · HBR: design AI so it strengthens reasoning, not replaces it
Evgeniou, Fayard and Tamisier-Fayard, writing in Harvard Business Review (HBR China ran the Chinese translation this Tuesday), collect the growing evidence that AI use erodes critical thinking — citing an MIT Media Lab study (2025) and a Wharton study (2026) finding that AI users stop questioning and verifying outputs. The organizational risks: expertise detached from context, and increasingly homogeneous thinking, since models tend to “average” away divergent views. Their three countermoves are wonderfully concrete. Make AI the questioner: in a 24-hour hackathon run with ETH Zurich and NTU researchers (~2,000 participants), teams whose AI feedback pointed out overlap with existing projects produced more novel ideas than teams whose AI helped them optimize. Carve out AI-free times and “entry-fee” interfaces that unlock AI only after you type your own contextual thinking — one Australian telecom requires AI-free strategy meetings first. And rethink the chat window itself.
Why it matters for us: this reads as syllabus-design language in disguise. The seminar room is the original AI-free space — worth defending on purpose, not by nostalgia. The “entry fee” translates directly to assignment design: students submit their own framing before AI access is allowed. And “AI as questioner” is a ready-made exercise — have students prompt a model to attack their draft's assumptions, then grade the rebuttal.
Read: HBR (English) · HBR China translation (中文)
3 · OpenAI: nearly half of job-specific AI use crosses job boundaries
OpenAI Economic Research launched its “Work at the Frontier” series (July 27) with a study of more than 800,000 messages from U.S. ChatGPT business users. The headline pattern, “task crossover”: 16.8% of work-related messages — and 43.5% of occupation-specific ones — concern tasks historically associated with a different occupation. The extremes: 77% of occupation-specific messages from customer-experience workers, 75% from designers, 69% from HR. Direction matters too: designers import heavily (35.2% of their messages are other occupations' tasks) but export almost nothing, while engineering tasks travel everywhere. And among typical users, crossover runs higher in small workspaces (18.9% at 2–5 seats vs. 16.3% above 100), suggesting AI is most useful “as a generalist tool where specialist resources are scarce.”
Why it matters for us: a liberal-arts college's core wager — breadth plus judgment beats narrow training — just got a data point from the labor market itself. The advising translation for our students: the message isn't “AI does the other person's job,” it's “you, plus AI, will be expected to cover more ground” — most of all in the small organizations where many of our graduates (and we ourselves) work.
Read: OpenAI summary · full report (PDF) · spotted via @RonnieChatterji on X
4 · What the “reasoning effort” dropdown actually does
Sebastian Raschka's explainer (July 18) demystifies a control most of us click without understanding: the low/medium/high “reasoning effort” setting now standard on flagship models — GPT-5.6 ships with roughly five or six per model size. The mechanism is humbler than it sounds: the chosen effort is inserted as a label in the system prompt, and the model has been trained to obey it — either by reinforcement learning that charges different per-token penalties depending on the requested effort, by supervised fine-tuning on traces of matching lengths, or both. Open-weight models make the consequences visible: more effort means longer reasoning traces and generally higher accuracy, with diminishing returns and rising cost at the top settings. Turning the dial down, as one reader put it, “can quietly cost accuracy.”
◆ See Figures 18–19 in the article for how effort level trades off against response length, accuracy, and cost — Ahead of AI.
Why it matters for us: genuine AI literacy for a non-specialist campus looks exactly like this: one honest explanation of one knob students touch daily. If a course permits AI, this is a strong week-one reading — it teaches that a model's “thinking” is a purchased budget, set by whoever configured the dropdown, and that the cheap answer and the good answer are often different products.
Read: Controlling Reasoning Effort in LLMs · spotted via @rasbt on X
5 · On Zhihu: intelligence as “learnable novelty”
Physicist Yanbo Zhang (章彦博) shared his team's new paper, “Intelligence from Learnable Novelty,” in a Zhihu post that usefully compresses the idea. Earlier this year the notion of epiplexity — the information a computationally bounded learner can actually extract — stirred debate; Zhang asked what happens if you optimize it rather than measure it. The reframing: pure noise is surprising but unlearnable; blank order is learnable but unsurprising; intelligence lives at the boundary, in “learnable novelty.” Maximizing a differentiable approximation of it, his team reports, drives cellular automata toward Rule-110-like complex soliton interactions, gets an image encoder to cluster handwritten digits with no labels, and improves PPO on sparse-reward reinforcement-learning tasks — with connections drawn to the free energy principle and novelty search.
Why it matters for us: this is the closest formal cousin I've seen to what educators call the zone of proximal development — learning happens where material is novel and learnable, not maximally novel. A useful lens for course design in the AI era: if AI removes all friction, it removes the learnable novelty too. Also worth noting for our bilingual campus: this conversation unfolded in Chinese, in public, on Zhihu.
Read: the paper on arXiv · 章彦博's post on Zhihu (Chinese)
If you try an “entry-fee” assignment this fall, or an oral defense in the spirit of Tao's rule of thumb, tell me how it goes. We are collecting examples for AI experimentation at DKU.
— Yisu