What Should Students Learn to Avoid AI Obsolescence? CNCC Roundtable Insights

The article stems from a CNCC 2026 roundtable where Professor Bao Yungang presented a striking experiment: using Codex and GPT-5.5 with 150 hours of effort and 450,000 words of lecture notes, an AI virtual student completed the full chip design flow to tape-out standard. This raises a sharp question: if AI can achieve our talent cultivation goals, do we still need to teach students that capability?

The experiment also reveals a critical control: when lecture notes were reduced to 4,500 words or 450-word prompts, the AI failed. The 450,000-word knowledge base was created by humans who defined the path, set standards, and provided deep guidance. The core conclusion: AI's apparent breadth is bounded by the cognitive depth of the human who defines the task and evaluates the output.

Knowledge Has Not Lost Value, But Its Role Has Changed

Forum consensus: knowledge is no longer the end product; it becomes an "internal ruler" for humans to judge, question, and define goals. AI excels at execution, retrieval, and answer generation; humans decide what to do, what is right or wrong, and where the bottom lines lie.

Traditional education treats knowledge as output — deriving formulas, reciting algorithms, restating concepts. In the AI era, AI outperforms humans in speed and accuracy on these tasks. If education still anchors on "computing the answer yourself," it cultivates capabilities exactly where AI is strongest.

Knowledge has not depreciated; it has shifted roles. Students still need foundational principles because without them there is no ruler to detect AI errors. Learning basics moves from "producing the answer" to "having a benchmark to judge correctness." A student who understands computer architecture can spot hidden timing risks in AI-generated RTL code; one who does not can only check if it runs.

Knowledge's new role: it is not the deliverable, it is the qualification to audit AI deliverables.

A vivid teaching design was shared: after students complete a chip design, three random bugs are injected and they must locate and fix them within one hour. This demands deep familiarity with low-level knowledge, the ability to locate anomalies, and professional judgment. Only with such foundation can one drive AI without being misled by its output. Injecting bugs and time-boxed debugging closes the loop between foundational knowledge, critical judgment, and AI verification.

Eight Capabilities, One Core

The forum identified eight capabilities for AI-era education:

Problem definition

Verification, critical thinking, and traceability

Complex value judgment and ethical decision-making

Systemic creativity and architectural thinking

Human-AI collaboration and steering

Deep empathy and communication

Intrinsic drive and metacognition

Aesthetic and meaning perception

These share a common core: they all point to the human as definer, auditor, and decision-maker, not as executor.

Problem definition extracts true problems from chaotic reality and sets constraints — AI does not know what problem to solve. Verification checks evidence, traces sources, spots logical gaps — AI fluently produces plausible but fabricated content. Value judgment makes trade-offs amid conflict — AI optimizes given objectives but cannot form its own values. Architectural thinking assembles modules into coherent, stable systems — AI excels at local generation but top-level architecture requires humans. Human-AI collaboration allocates roles and sets safety guardrails — not banning AI but not outsourcing independent thinking. Empathy understands human emotion, needs, cultural context — AI can write scripts but struggles to truly understand people. Intrinsic drive and metacognition involve knowing what one knows, what one doesn't, and when stuck — AI supplies endless answers but cannot build internal purpose. Aesthetic and meaning perception judges what is good, beautiful, worth pursuing — AI can generate pretty images but cannot autonomously sense value.

These are "meta-capabilities" — not about "how to do" but about "deciding what to do, judging if it's done right, and why." This directly mirrors the experiment's conclusion: AI's ceiling depends on human cognitive depth, which manifests in definition, judgment, and choice, not execution.

MIT 48-Hour Case: Insights and Boundaries

A widely circulated case: an MIT graduate used Google NotebookLM to cram an unfamiliar course in 48 hours and pass a qualifying oral exam. His method was not asking AI for highlights but asking three core questions: (1) What are the five core mental models experts agree on? (2) What are the three most heated expert disagreements? (3) Produce 10 deep discriminative questions that separate true understanding from rote recitation.

This flips the learning sequence: map first, details later; disagreements first, conclusions later; use high-level questions to drive understanding rather than linear knowledge accumulation. It echoes "problem definition" and "architectural thinking" — build the disciplinary landscape first, then dive into details.

However, boundaries exist. This suits qualifying exams and macro framework assessment, not hardcore foundational courses requiring deep derivation, rigorous proof, and engineering stability training. Algorithms and architecture involve deep logic, boundaries, and hardware constraints that 48 hours can only survey, not internalize into reliable intuition. Crucially, the student already possessed strong liberal arts and critical thinking foundations; a student lacking basics would just rapidly acquire plausible AI hallucinations.

The real value is not "learning a course in 48 hours" but demonstrating a new AI-era learning possibility: AI as "literature consultant, adversarial examiner," while the core human capabilities are designing high-quality prompts, verifying AI output, and identifying blind spots — precisely "problem definition" and "verification."

Pedagogy Must Co-Evolve: The Deep Value of Task-Driven Learning

Capability shifts demand pedagogical shifts. "Task-driven teaching" was repeatedly emphasized: use real, complete engineering tasks with clear deliverables as the backbone, embedding knowledge points into the task completion flow so students actively learn principles, research, design, debug, and verify to finish the task.

Traditional mode: teach knowledge points → exercises → final integrated application. Task-driven: throw target task → encounter gaps → actively learn knowledge → iterate implement, debug, accept.

In the AI era, task-driven value upgrades. It is no longer just "produce an artifact" but: throughout completing a complex task, force students to build complete systems thinking, understand interdependencies among knowledge points; in high-pressure debugging scenarios, train low-level principle intuition, professional judgment, traceability — the very bedrock for auditing, verifying, and steering AI.

In one sentence: task-driven learning is a highly effective vehicle for training "humans as auditors and decision-makers."

Risks exist. The biggest is "heavy output, light principle" — students rely on AI to copy solutions, skip derivations, only chase task completion. Countermeasures: independent debugging checkpoints, principle defenses, fault injection, disallowing AI from doing deep debugging. Task difficulty calibration is hard: too simple yields low training value; too hard causes frustration. Faculty demands rise: teachers become task designers, process guides, assessment gatekeepers, cognitive blind-spot detectors.

These risks show AI-era education reform is not "put AI in the classroom" — it requires redesigning the entire teaching process, assessment mechanisms, and teacher-student roles. Tools alone cannot solve motivation and classroom interaction; reform must address these depths.

Employment Challenge: Not "No Jobs" But "Capability Mismatch"

A panelist noted AI-driven roles and required skills iterate rapidly; we are in early stages, impact will amplify. "Unemployed upon graduation" will shock society.

Deeper analysis: the challenge is not AI stealing jobs but a mismatch between educated capabilities and AI-era needs. If campus skills expire at graduation, the problem is not vanishing jobs but capability structures lagging role demands. AI replaces standardized, repeatable, rule-based tasks; AI amplifies demand for problem definition, value judgment, system architecture, human collaboration. If education anchors on the former, graduates feel "what I learned is useless"; if it pivots to the latter, they face larger opportunities.

Two notable judgments: platform employment will rise as a share of total employment, traditional fixed employment organized around conventional factors will decline. The service sector will be the primary arena for AI deployment and the main source of new jobs — as AI replaces standardized work, it amplifies service-oriented, human-facing, personalized demand.

Both point to the same conclusion: future employment will be more flexible, dispersed, reliant on individual capabilities rather than job labels. Education must not just produce "people who can fill a role" but "people who can define their own value in continuously changing environments." This returns to the eight meta-capabilities — they are not role-specific skills but foundational literacies for any role.

A Deeper Insight: Education's Dual Mission

A panelist's remark, easily drowned in technical discussion, may be the most profound: education has not only a production-facing, functional service side but also a "consumption" value. Like the ancient arts of music, chess, calligraphy, painting — essentially spiritual "consumption." In the AI era, how education enhances students' spiritual consumption capacity, satisfying higher-level spiritual needs, is equally important.

This warns against utilitarian anxiety when asking "what to learn to avoid AI replacement" — what secures jobs, what avoids substitution. Education's other face: how humans settle their spiritual lives, perceive meaning, appreciate beauty, gain spiritual richness.

AI can endlessly generate content, but appreciation, feeling, spiritual richness depend on human general knowledge, aesthetic and humanistic cultivation. The ceiling of one's ability to unlock AI potential is set by oneself — this "self" includes not just technical ability but liberal arts accumulation and comprehensive literacy.

Thus, what should university teach? A dual-layer answer: facing production, learn capabilities that let you define problems, audit AI, architect systems; facing life, learn capabilities that let you appreciate, feel, settle your spirit. The former keeps you from being replaced by AI; the latter keeps you from being swallowed by utilitarianism.

Conclusion: Redefining University's Meaning

Professor Wen warned of alarming trends: students questioning university's meaning, feeling years are spent hearing decade-old knowledge; some PhD candidates drop out, unwilling to invest time in what they see as low-value learning.

This is a severe challenge for educators. But flipped: student skepticism points the reform direction. If university defines itself as a "knowledge transmission institution," it will be replaced by AI; if it redefines as a "capability training ground, judgment incubator, spiritual growth soil," its value becomes more prominent in the AI era.

AI can give answers but cannot ask questions for you. AI can generate solutions but cannot judge right from wrong for you. AI can write scripts but cannot understand people for you. AI can produce works but cannot perceive meaning for you.

What university should teach is exactly what AI cannot replace. The solution to employment challenges lies not in chasing every new tool or skill but in guarding the foundational capabilities that make us human — defining problems, auditing judgments, architecting systems, understanding others, perceiving meaning. These capabilities never expire because they are not a skill list for a specific era but meta-capabilities for any era. The faster AI iterates, the more valuable they become.