A course at Nanjing University is treating AI tokens as a practical part of learning software engineering rather than an optional add-on. QbitAI reported that associate professor Jiang Yanyan has introduced a Fall 2026 class called Generative Software Engineering, built around the premise that students need to learn how to direct, evaluate, and pay for AI-assisted programming work.
The course has drawn attention because of Jiang’s blunt language about access to paid tokens. QbitAI reported that he said computer-science students who lack paid tokens should leave the field, then described the remark as rhetorical emphasis on the importance of tokens in an AI-enabled software environment. The statement is Jiang’s personal view and course framing, not an administrative policy of Nanjing University.
The more substantive development is the curriculum itself. QbitAI says students are expected to learn task decomposition, model selection, cost management, and verification rather than rely on what Jiang calls old-style programming. The approach turns a software course into a test of how students use generative tools as part of a broader engineering process.
The Course Treats Tokens as a Software Input
AI tokens are the units that model providers use to measure input and output when pricing many APIs. In a consumer chatbot, the cost may be hidden in a subscription. In a software-engineering workflow, token use can become visible and variable as a developer asks models to generate code, inspect a repository, write tests, summarize documentation, or work through a multi-step task.
QbitAI reports that Jiang’s course adopted a token self-funding policy. The instructor also said he would help students with genuine financial hardship. That combination makes the policy more nuanced than the widely repeated remark suggests: the course asks students to recognize that model use has an economic cost, but it does not present inability to pay as a formal exclusion rule.
Jiang’s own course wiki is part of the supporting record for the class. The available reporting describes a course that asks students to manage inputs, verify outputs, and decide when a particular model is appropriate. Those choices resemble the decisions that software teams increasingly make in workplaces, where an AI tool can be useful but an uncontrolled call pattern can also create costs, security concerns, or unreliable code.
The course’s focus on cost connects to a broader education question raised in EastFrontier’s report on Hong Kong’s AI training workforce-policy test. A student who learns to manage token use efficiently is learning a constraint that also applies when enterprises decide which AI tasks to automate and how much model access will cost.
Task Decomposition and Verification Become Core Skills
The reported curriculum does not say that students should accept AI-generated code without review. Instead, it requires them to break work into smaller tasks, choose tools, and verify results. Those requirements are central because a generative model can produce plausible but incorrect code, misunderstand a specification, or create an inconsistency that only appears when a full system is tested.
That makes the course different from a simple introduction to code completion. A code-completion tool can help with a local function, while an AI-assisted engineering process requires a student to define the problem, supply relevant context, judge the output, and test the result. The human role shifts from typing every instruction to specifying constraints and evaluating whether the model’s work fits a larger system.
QbitAI says Jiang recommended a low-cost DeepSeek model for cost-effectiveness. That should be read as an instructor’s tool preference, not a university procurement decision or a claim that one model is best for every task. The educational point is broader: students should understand the trade-off between model capability, latency, price, and the amount of verification a task will need.
China’s higher-education system is already adding AI-related programs and courses. In August, EastFrontier reported that Chinese universities were adding more AI and embodied-intelligence majors. Jiang’s course represents a more specific shift within that trend. It is not a new degree program. It is a change in what an existing software-engineering class treats as fundamental practice.
The course could also create a useful tension between speed and understanding. Students may be able to move faster when a model proposes code, but they still need to understand the architecture, identify errors, and explain their decisions. A grading system that rewards only output could encourage uncritical use. The reported emphasis on verification is therefore one of the most important elements of the design.
One Instructor’s Experiment Is Not a National Rule
The publicity around Jiang’s statement could easily obscure the limits of the case. This is one course taught by one associate professor at Nanjing University. The policy does not mean that Chinese universities have adopted a general requirement that all computer-science students pay for model tokens. Nor does it establish that conventional programming skills have ceased to matter.
The course does reflect a real question facing software education. If AI tools can produce portions of code, what should a student be able to do independently? Jiang’s answer, as reported by QbitAI, is that students should be able to manage an AI-assisted workflow. That includes dividing a task, selecting a model, watching costs, and validating output. It does not eliminate the need to understand logic, data structures, debugging, security, or software design.
There are also equity questions. A course that relies on paid tokens may give some students easier access to higher-capability models or more experimentation. Jiang’s stated willingness to help students with financial difficulty addresses that concern in principle, but the article does not describe the mechanism, budget, or scale of such assistance.
The model-selection issue is similarly unsettled. Tool availability and pricing can change quickly, and students may need exposure to more than one provider. The lasting lesson may be less about a particular model than about the discipline of treating model calls as an engineering resource. That is a useful habit whether a student uses a domestic API, an open-weight model, or a university-provided system.
Nanjing University’s course is therefore best understood as a visible experiment in curriculum design. It converts token use from an invisible feature of an AI product into a subject students must plan for. If other programs adopt similar approaches, the important question will be whether they pair AI fluency with careful verification, equitable access, and the foundational skills needed to judge a model’s work rather than simply accept it.