Toward AI-Native Education: Lessons from LLM-Wiki

2026-08-02

AIeducationLLM-Wikiknowledge-managementTerrateaching

If a university is serious about educating AI-native students, it should pay for every student to have serious access to frontier AI. Give each of them the $100 ChatGPT Pro tier, or an institutional equivalent, and then let them spend enough time with it to get past occasional prompting. It sounds almost irresponsibly easy. Students need that time to learn what these tools make possible, what they make imaginable, and what should never be delegated to them. AI literacy comes from sustained contact: trying ambitious things, getting impressive results, encountering confident nonsense, revising a question, and slowly developing judgment. I teach coding and data analysis every year because most problems students meet after graduation will have no answer key. They will have to decide what the agenda is, discuss it with other people, negotiate a shared position, and support their claims with the best data available at that moment. That is what doing science means. AI has become part of that environment. Leaving it as an optional accessory lets family income determine who gets to practice.

The new subscription gap

Korean universities already know what an access gap looks like. More than a decade ago, English-medium lectures spread across campuses, often because international rankings rewarded them. In classes where more than 90 percent of the students were Korean, every discussion and activity still had to happen in English. The result was predictable. Students who had grown up with stronger English education could write and participate with relative ease. Others had to learn the subject and cross a language barrier at the same time. AI subscriptions are beginning to create a similar divide. Some students use the strongest models every day: they ask for explanations, debug code, compare arguments, reorganize notes, prototype agents, and iterate on writing. Other students ration free messages or use whatever tool happens to be available. The difference compounds into what students attempt, how quickly they recover from failure, and how large a project they can imagine doing on their own. A university that requires students to become AI-literate while leaving access to family income has built a new inequality into the curriculum.

University is where students learn to choose

When universities discuss AI, they often begin by deciding what students are allowed to do. The language is full of limits, detection, disclosure, and control. Some of that is necessary, but we also have a strange habit of defining all the possible uses in advance and calling the result AI education. It resembles giving a child an iPhone with an hourly schedule already attached: one hour for this approved activity, one hour for that one, and no wandering. The device is present, while the freedom through which a person learns what the device is has disappeared. Students need unassigned time with powerful tools. They might connect AI to a concept from basic science, build a small agent around a private curiosity, analyze a paper that was not on the syllabus, or make something apparently useless. Some of that time will be wasted, as it is in a library, a laboratory, or a conversation. Exploration rarely looks productive before it becomes productive. What if students play games, make unrelated videos, or use an expensive subscription for something that has nothing to do with class? Universities should be able to tolerate a certain amount of freedom without demanding a dashboard of approved clicks.

The educational value of AI begins where the approved use case ends.

Teaching a field that does not yet have a textbook

This semester I have been teaching virtual cells and single-cell foundation models. Students need mathematical foundations before they can understand the models, yet the applied field itself has not settled into a standard workflow. When I teach conventional omics analysis, there are familiar stages. The tools change, but the broad pipeline is recognizable. AI applications in omics foundation models work differently. A research question comes first; then we ask whether and how a model can predict something relevant to that question. The field sits awkwardly between mathematical foundations, model architecture, biological interpretation, and opportunistic application. Preparing a lecture therefore becomes a form of research. I choose a topic because I think, "This would be useful to show," begin organizing the literature, and sometimes discover that I am already outlining a paper. A fixed lecture cannot keep up with a field in which the syllabus is still being invented. In a stable field, a textbook can compress a mature consensus. In this field, students need tools for handling uncertainty, tracing claims back to evidence, and building a provisional structure that can change next week.

LLM-Wiki as a personal curriculum

I used LLM-Wiki with ChatGPT in my graduate course last semester, and I am now adapting it for undergraduates. LLM-Wiki gives each student a place to design knowledge around questions that are genuinely their own. A conventional course imposes one sequence: week one, week two, midterm, final. A personal wiki can begin from the same shared course materials and then branch. One student may become interested in perturbation prediction. Another may keep returning to noncoding variants. Another may care about the history of how a biological concept changed. Their source collections, questions, concept pages, and links gradually diverge because their attention differs. The student decides how the knowledge is organized. Each student selects papers and other source materials worth keeping, asks questions that expose gaps or contradictions, turns useful answers into durable concept pages, and links new material to what is already understood. When a synthesis becomes too smooth, the student returns to the original evidence. As the wiki grows, the student can see which parts reflect a narrow reading history and decide where to widen it. Every step requires thought and, in practice, more reading. AI can propose a structure, but the student decides whether that structure is faithful, useful, and worth preserving. The wiki leaves those decisions open to inspection, criticism, and revision. Thought is outsourced when a student asks for an answer, submits it, and leaves no intellectual trace. In LLM-Wiki, each answer becomes material for another question, and every question changes a knowledge system the student remains responsible for. Students who learn this way can make their own paths.

Why Terra fits this work so well

Today I ran a small but revealing LLM-Wiki benchmark. I compared an existing GPT-5.6 Sol ingest with a fresh GPT-5.6 Terra Extra High ingest of the same 28-page genome-inversion benchmarking paper. Terra received the original paper and normalized metadata, but it did not see Sol's existing source note, wiki page, or synthesis page. Both models got the paper's identity and central conclusions right. Sol was more concise. Terra did better on the parts that matter when a chat response will become durable knowledge: evidence traceability, methods specificity, numerical coverage, and careful handling of limitations. It produced a six-claim ledger tied to pages and figures, distinguished the manuscript text from the final publication metadata, and gave a more qualified account of where the benchmark could and could not support a conclusion. This was one paper, so it cannot support a general model ranking. I also could not compare latency or cost fairly because the original Sol run did not record the same settings. OpenAI describes Terra as the balanced model for everyday work. In my LLM-Wiki, Terra was unusually good at the unglamorous work of building trustworthy notes. A student needs to see which figure supports a claim, what a method actually compared, where a number originated, and which caveat limits the conclusion. Terra preserved those boundaries while building the wiki. In this task, Terra was a very good teaching assistant: it showed its receipts.

Teaching students to take AI writing apart

Next semester I want to use AI heavily in a writing course. Students will generate text with AI and then dismantle it. Why does the model write this way? How is its paragraph logic different from scientific writing before LLMs? What disappears when every introduction converges on the same smooth funnel? For the past two months, I have been downloading and examining papers by major scientists from the 1990s, before AI-mediated prose. How did they build an introduction, in three paragraphs or four? Did they use the inverted-pyramid structure I prefer? Did Results and Discussion remain strictly separate, or did interpretation appear alongside observation? How did one sentence create the logical need for the next? These papers make AI's default style easier to see. Students can use AI while recognizing its habits and recovering forms of scientific argument that the model tends to flatten. The goal is to give students enough historical and rhetorical awareness to choose. They should be able to say: the model wrote it this way, the scientific tradition offers these alternatives, and I am choosing this structure for a reason.

What universities should provide

Universities can fund serious AI access for every student and negotiate institutional agreements with frontier-model providers such as OpenAI or Anthropic. They can build university-level cloud computing so that students move from prompting to actual experimentation. Students should learn evidence, source tracking, privacy, and accountability through repeated practice. They also need time to explore without having to justify the destination in advance. Substituting a weaker model simply because it can be labeled "sovereign AI" is a poor educational strategy. Students learn fastest when the tool is strong enough to reveal both capability and failure. A limited model teaches a limited imagination and can create false confidence about what the technology already does. AI belongs in university education because it changes what a student can attempt, how quickly a private curiosity can become a project, and how visibly a personal knowledge structure can grow. Universities should give students powerful tools, shared standards for evidence, and enough freedom to discover a direction worth choosing.