Under the Generative AI Wave, AI Education Is Being Rewritten: Courses, Tools, and Deployment Have All Changed ​

We · September 2026

AI technology evolves quickly. The courses, tools, and models mentioned in this article reflect the current stage and are for reference only.

Why Can't Online AI Services Meet Teaching Needs? ​

Some might ask: Can't students just use online AI services like ChatGPT and Midjourney?

We believe online services have clear limitations in teaching scenarios—uncontrollable costs (API fees for a class of 40 students over a semester could reach thousands of USD), data privacy and compliance issues (teaching and research data can't simply be sent to third-party servers), students only learning "how to use" rather than "how to build" (APIs are black boxes—you can't see inside the model), and services that may shut down or change at any time. We'll discuss these issues in depth in the second article of this series, Why Universities Need to Build Local AI Lab Environments.

Here's the conclusion up front: Local deployment is not an "optional add-on" but a "must-have" for generative AI education. From Tsinghua University's executive workshop to Beijing University of Posts and Telecommunications' full-stack AI course, local deployment is treated as a core component. With that in mind, let's look at just how much change generative AI is bringing to education.

The Entire Technology Paradigm Is Shifting ​

If you wound the clock back to 2022 and asked a university computer science department, "What does an AI major study?", the answer would probably be: Python, machine learning, deep learning, neural networks, computer vision, natural language processing...

But after ChatGPT took off in 2023, everything changed.

Generative AI brings not just new models, but a whole new technology stack, new development practices, and new engineering workflows. This means AI education has to change with it—and the magnitude of the change is likely much larger than many people assume.

More importantly, this is no longer just a computer science department matter. Design schools, media schools, architecture departments, finance programs... almost every discipline is being penetrated by generative AI. We believe this is the most easily overlooked side of this shift: the audience for AI education is expanding from a few hundred CS students to tens of thousands of students across every major on campus.

This article examines what this change has actually rewritten, from two levels: theoretical teaching and hands-on practice. We'll then look at how course systems are changing, how software tools are changing, how deployment is changing, and finally pull together a global progress comparison.

What's Actually Different About Generative AI? ​

Before talking about education, let's clarify one question: what's the fundamental difference between generative AI and previous AI?

Traditional AI/machine learning is centered on "discrimination"—classifying data and making predictions. For example, deciding whether an image is a cat or a dog, predicting housing prices, recommending products you might like. Its output is a label, a number, a ranking.

Generative AI is centered on "generation"—it produces new content. A passage of text, an image, a voice clip, a video, a piece of code. It isn't answering "what is this?"—it's creating "what could this be?"

This isn't a simple feature upgrade; it's a shift of the entire technology paradigm.

At the Theory Level: The Body of Knowledge Is Expanding ​

From the perspective of theoretical teaching, generative AI introduces a whole set of new concepts that didn't exist in traditional courses:

The knowledge system of traditional AI courses is relatively stable—supervised learning, unsupervised learning, neural networks, optimization algorithms—these have changed little over the past decade. But generative AI introduces entirely new concepts: attention mechanisms, the Transformer architecture, diffusion processes, prompt engineering, context engineering, chain-of-thought, retrieval-augmented generation (RAG), agents, model quantization, parameter-efficient fine-tuning...

This means teachers have to learn these things first, then teach them to students. For many teachers, this isn't a matter of "preparing one lecture"—it's a matter of "learning it all over again."

At the Practice Level: From "Training Small Models" to "Building a System" ​

The shift in hands-on practice is even larger than in theory.

A traditional AI practicum follows roughly this flow: take a dataset (Iris, MNIST, CIFAR), write code to train a small model, tune parameters, look at evaluation metrics. The whole process can be done inside a Jupyter Notebook, and Colab's free tier is enough to run it. A student can go from idea to validation in maybe half an hour.

Generative AI practice is on a completely different scale. Say you want students to build a "Q&A system based on a local knowledge base." The "idea" itself is simple, but getting it working requires:

• First setting up a large model inference environment (Ollama or similar)
• Downloading a suitable model weights file (several GB to tens of GB)
• Building the RAG pipeline: document parsing → vectorization → store in a vector database → retrieval → generation
• Possibly adding an Agent layer so the system can call external tools
• Debugging prompts, tuning retrieval quality, handling hallucinations...

Each step is its own engineering task. A student may need to spend a whole day just setting up the environment before even starting the project.

We have observed that this is in fact the most common bottleneck schools hit when rolling out generative AI teaching: it's not that students can't learn the theory, but that the cost of standing up a practice environment is so high that a large amount of teaching time gets consumed just by "setting up the environment."

Researchers at Oklahoma Christian University discussed this shift in a paper. The author designed the course in two stages—the first stage covers traditional machine learning fundamentals, and the second stage covers large language model applications, including prompt engineering, RAG, model deployment, quantization, LoRA, and the Model Context Protocol. This structure itself makes the point: generative AI is not replacing traditional AI; it's adding an entirely new layer on top of the traditional foundation—and the practical threshold of this new layer is much higher than the old one.

Not Just a CS Matter: Non-AI Majors Are Being Penetrated ​

Everything above is about changes to "AI majors" and "AI courses." But We believe the most underestimated impact of this shift is actually happening in non-AI majors.

Generative AI lowers the barrier to using AI—you don't need to understand algorithms or write training code; as long as you can use the tools, you can integrate AI into your own field. This means design, media, architecture, finance, medicine... almost every discipline is rapidly adopting AI tools.

But here's the problem: the students and teachers in these majors mostly have no programming background. Getting them to use AI tools is a completely different matter from getting CS students to use AI tools.

Design Majors: ComfyUI Has Become a Required Course ​

Design-related majors (visual communication, product design, industrial design, fashion design) are among the fields where generative AI has penetrated fastest. The core tool is ComfyUI—a node-based workflow platform that can do text-to-image, image-to-image, ControlNet control, LoRA fine-tuning, animation generation, and more.

An AIGC course at Peking University explicitly requires students to "master the deployment, configuration, and debugging of a ComfyUI environment based on Stable Diffusion," covering workflows such as text-to-image, image-to-image, LoRA fine-tuning, ControlNet, IPAdapter, InstantID, AnimateDiff animation, and SVD video generation.

The School of Art and Design at Dalian Polytechnic University runs a 24-class-hour "AI Generative Design" course centered on ComfyUI, structured as "foundational cognition—environment setup—core functions—advanced practice—project application."

In March 2026, the Communication University of China ran a ComfyUI hands-on training session with nearly 400 faculty and students participating—note, not 400 CS students, but faculty and students from across the entire university.

So what do these design students run into when installing ComfyUI?

We compiled the typical difficulties:

In ComfyUI's official troubleshooting documentation, "missing models," "Python version mismatch," and "insufficient VRAM" are the most common issues. For CS students, these are all minor issues; for design students, any one of them can make them give up on the spot.

Media/Film Majors: The Barrier for Video Generation Tools Is Even Higher ​

The tools used by media and film majors are heavier: AI video generation (WAN 2.2, HunyuanVideo, LTX), AI voice synthesis (GPT-SoVITS, IndexTTS, CosyVoice), AI digital humans.

The Communication University of China's "AI Super Creator Program" course covers hands-on use of tools such as Midjourney, Flux, Sora2, and ComfyUI workflows.

But the deployment barrier for video generation tools is higher than for image generation:

• The 14B version of WAN 2.2 typically needs around 24GB of VRAM under FP8 quantization; the 5B lightweight version is friendlier to VRAM but still needs 12GB+.
• LTX Desktop requires NVIDIA CUDA, 16GB+ VRAM, 16GB RAM, and 160GB+ of disk space.
• Video generation also depends on ComfyUI, custom nodes, and specific versions of PyTorch and CUDA—which means you have to first install ComfyUI, then install the video nodes, then download the video models; the whole chain is longer.

We noticed a contradiction: media students need AI video tools the most to boost their creative efficiency, yet they are precisely the group least equipped to handle these technical deployment problems.

Architecture Majors: AI Renderings Are Changing the Design Workflow ​

Architecture and interior design majors are also rapidly adopting AI. The main toolset is Stable Diffusion + ControlNet, used to generate architectural renderings, interior design concept images, landscape planning images, and more.

The Liverpool School of Architecture ran an AI-driven architecture program in its 2025 winter school, including generative AI drawing workshops and an AI drawing competition, with students using Adobe Firefly and Midjourney for concept and composition training.

One architecture education research paper noted that in 2023, students mainly used AI as an "image generation tool," but by 2024 they had started engaging more deeply with generative strategy design.

The typical AI workflow for architecture majors is: export line drawings or depth maps from SketchUp or Rhino → import into ControlNet → use Stable Diffusion to generate renderings. But this requires installing ComfyUI, downloading ControlNet weights (one file per control mode), and downloading architecture LoRA models—tens of GB of model files in total, plus complex path configuration; for architecture students, this is also a hurdle.

Finance/Business Majors: RAG and Knowledge Base Demand ​

Finance and business majors have a different AI need—they need local-knowledge-base-powered LLM Q&A systems (RAG) to analyze research reports, retrieve regulations, and assist with investment decisions.

Shanghai University of Finance and Economics – Dishui Lake Advanced School of Finance MF program has set up a financial AI track, with courses covering machine learning, large models, and Python for finance.

Academia also has dedicated finance RAG research that designs modular systems supporting different retrieval mechanisms, language models, and prompting strategies. The banking industry has already started using RAG systems to build intelligent customer service, combining LLM reasoning with the bank's internal private knowledge base.

But the barrier for finance students to deploy a RAG system locally is just as high: they need to install Ollama to run LLMs, configure LangChain for the RAG pipeline, install a vector database (FAISS or ChromaDB), and handle document parsing—each step is a challenge for business students with no programming background.

The Shared Plight of Non-Technical Users ​

We summarized the shared difficulties across these majors in one table:

In a sentence: these students aren't unable to learn AI—they get blocked by environment setup before they even start.

Course Systems Are Being Rewritten ​

Back to the course changes within AI majors themselves. Universities around the world are iterating quickly.

United States: From Foundations to Applications, Fast Course Iteration ​

Stanford's new CS 224G in winter 2026 is themed "LLMs, LMMs, and Reasoning Models," covering reasoning models (OpenAI o1/o3, DeepSeek R1), chain-of-thought prompting, RAG architecture, memory systems, context engineering, and token management. Note these keywords: context engineering, memory systems, token management—these don't exist at all in traditional AI courses.

Columbia University's fall 2024 deep learning and LLM course brings LangChain directly into the classroom.

MIT's 6.868 Deep Generative Models focuses on diffusion models, autoregressive models, GANs, VAEs, and their applications in imaging, video, robotics, biology, and materials science.

UPenn's Wharton School's Generative AI Studio has MBA students develop generative AI prototypes in a project-based format—the business school is keeping pace too.

China: Micro-Majors + Practical Training Courses, Fast Follow-On ​

Shanghai Jiao Tong University launched a "Generative Artificial Intelligence AIGC" micro-major in 2025, now in its second cohort.

Tsinghua University's School of Continuing Education runs a "Chief Intelligence Officer LLM Workshop" that's more enterprise-oriented, with content covering RAG construction and model fine-tuning practice, enterprise API calls and fine-tuning, and local-first AI agent deployment.

Beijing University of Posts and Telecommunications launched a "Full-Stack Domestic AI Specialty Course" in 2025, covering the entire chain from CANN operator programming to large language model technology (Transformer, LoRA/QLoRA fine-tuning, RAG architecture, Agent, model quantization W8A8).

There are also more grounded changes: Nanjing Institute of Technology offers Introduction to Large Language Models using Ollama for practice, Guangzhou Institute of Software Engineering offers AI-Native Application Development in Practice using LangChain as the textbook, and the Communication University of China has taken an art + AI route.

The National University of Singapore (NUS) announced that starting in August 2026, all undergraduates must take an AI course. This is no longer a CS matter, it's a university-wide matter—which echoes exactly what we said earlier: generative AI education is shifting from specialist education to general education.

What Exactly Has Changed in Course Content? ​

Software Tools: From "A Few Things" to "A Whole Ecosystem" ​

Behind the course changes is an explosive growth in the tool stack.

The Traditional ML Era: Few and Stable Tools ​

Traditional machine learning teaching has a simple tool stack: Python, Pandas, NumPy, Scikit-learn, TensorFlow or PyTorch, Jupyter Notebook, Google Colab. pip install a few packages and you can run a class. Models are a few MB to a few hundred MB, and Colab's free tier can run them. The path from idea to validation is short.

The Generative AI Era: Multiple Layers, Multiple Choices ​

The generative AI tool stack is more than an order of magnitude more complex:

Model layer — Large language models include GPT, LLaMA, Qwen, DeepSeek, GLM, Mistral and dozens more; image generation has FLUX and the Stable Diffusion family; speech has IndexTTS, GPT-SoVITS, CosyVoice; video has WAN, Hunyuan, LTX...

And model iteration speed is astonishing. We noticed that in 2026 alone, a large number of major models have emerged:

In video generation, MiniMax released H3 in July 2026—a general-purpose multimodal generation model that is no longer just "text-to-video" but can simultaneously output video and native dual-channel stereo audio, supports 11 languages, runs up to 15 seconds and 2K resolution, and can control generated content via multimodal reference inputs (up to 9 images + 3 videos + 3 audio clips). LTX released LTX-2.5 (22B parameters) in August 2026, supporting native audio-video sync generation and multi-shot sequences. Add in Wan 2.2 (27B MoE architecture) and HunyuanVideo 1.5 (8.3B) released in 2025, and video generation models have evolved from "producing pictures" to "pictures + sound + multi-shot" in one step.

In large language models, 2026 is just as busy. Alibaba released Qwen3.8-2.4T-A95B in August, a 2.4-trillion-parameter sparse MoE architecture with a 1-million-token context window. DeepSeek released V4.1-Flash in September, with 552B main parameters, native multimodal input, and also a 1-million-token context.

In image generation, Black Forest Labs released the FLUX.2 series in January 2026, with the 4B version running on just 8.4GB of VRAM under an Apache 2.0 open-source license.

We believe this is both a good thing and a higher bar for teaching. The more powerful the models and the more features they have, the more course content has to keep updating; the larger the VRAM requirements, the higher the deployment barrier; and the faster the iteration, the more a course may be outdated as soon as it's written. H3 raises the bar over previous video models—native audio, multimodal references, instruction-based editing—but this means the tools students need to learn are more complex, and the hardware threshold to run them is higher. Stronger models will keep appearing, and the pace of teaching iteration has to keep up.

Inference runtimes — Ollama for entry-level, llama.cpp for CPU inference, vLLM and SGLang for production environments.

Application framework layer — For RAG there's LangChain, LlamaIndex, and Haystack; for agents there's LangGraph and DSPy; for vector databases there's FAISS, ChromaDB, Pinecone, and Weaviate.

Image generation tools — ComfyUI is the de facto standard, alongside A1111 and SwarmUI.

Fine-tuning and quantization — LoRA/PEFT/LLaMA-Factory for fine-tuning, GGUF/AWQ/GPTQ/FP8 for quantization.

Model platforms — Download models from Hugging Face or ModelScope, each with different licenses.

Some researchers have noted that the tools and frameworks involved in generative AI teaching have gone from a single-digit number to dozens, and are still increasing rapidly.

We believe this implies a fundamental change for teaching: teachers not only have to teach "how to use," but also "how to choose"—choose a model, choose a framework, choose a quantization method, choose a deployment plan. And tools update so fast that the LangChain usage taught today may have a new version next month.

From a practical standpoint, this means the barrier for students to quickly validate an idea is much higher. In the traditional ML era, trying out an algorithm took ten lines of code to run. In the generative AI era, trying out a RAG approach may require installing Ollama, downloading models, configuring LangChain, and standing up a vector database... environment setup alone can eat up most of a day.

Deployment: From pip install to Systems Engineering ​

If the proliferation of tools is only a "learning cost" issue, then the change in deployment is a real "engineering barrier." And this is precisely the most impactful factor in the practice phase—especially for non-technical students.

Traditional ML Deployment: Relatively Simple ​

Deploying traditional machine learning is essentially just installing Python packages:

pip install scikit-learn pandas numpy pytorch jupyter

Models are small, dependencies are few, and the environment is relatively stable. From installation to running, it might take ten minutes.

Generative AI Deployment: Six Layers of Hurdles ​

Layer 1: Is the hardware enough? — A 7B model in FP16 needs 14GB of VRAM, and about 3.5GB after INT4 quantization; for models over 30B, you basically can't run them without 24GB of VRAM.

Layer 2: Which model to choose? — Parameter scale, quantization format, context length, inference speed, license... every factor affects the choice.

Layer 3: Where do the model weights come from? — Download from Hugging Face or ModelScope, tens of GB or even over a hundred GB. Is the school's network fast enough? Is the student's home network fast enough?

Layer 4: How to configure the dependency environment? — Taking ComfyUI as an example: Python version, CUDA version, PyTorch version, custom node dependencies... any incompatibility in any link and it won't run.

Layer 5: Quantization and inference optimization — FP8, INT4, GGUF, AWQ, GPTQ... students not only need to understand models, they also have to understand these engineering techniques. This essentially doesn't exist in traditional ML teaching.

Layer 6: Management and maintenance — A lab with tens or hundreds of machines: how do you distribute models uniformly? How do you manage versions? How do you recover quickly when something goes wrong?

Environment Is the Core Bottleneck of Practice Teaching ​

Researchers at Oklahoma Christian University noted in their paper that LLM technology evolves so fast that course materials need frequent updates. Other research has pointed out that adoption of generative AI in education faces multiple strategic obstacles: governance gaps, teacher trust deficits, disparities in student usage, and lagging school policies.

We believe these discussions mostly focus on "what to teach" and "how to teach," but the more fundamental question of "how to set up the environment" is actually under-discussed. Yet in actual teaching, this is precisely the decisive factor in whether students "can quickly get hands-on with practice."

Look at the reality:

• Top-tier universities have compute platforms and technical teams, so the environment isn't a problem (Shanghai Jiao Tong University's "Zhiyuan No. 1" intelligent computing platform, Beijing Institute of Technology's AI Practice LAB co-built with Huawei).
• Regular universities have bought GPUs but can't stand up or maintain the software environment.
• Vocational schools have it harder—limited faculty technical capability and uneven student foundations.
• Schools in developing countries often don't even have a stable GPU environment.
• Non-technical majors go without saying—design students have never even opened a terminal.

Some research directly points out that the resource-intensive nature of generative AI models creates deployment barriers for resource-constrained institutions.

Global Progress Comparison: Who's Ahead, Who's Catching Up ​

First Tier: United States ​

The fastest course iteration, the most complete toolset, and deep coverage from theory to application. Stanford, MIT, CMU, and Berkeley iterate fastest on generative AI courses. And it's not just CS—business schools and design schools are keeping up quickly.

Second Tier: Top Chinese Universities + UK/Switzerland ​

Top Chinese universities: Fast institutional build-out, large investment, courses lean toward application and engineering practice. In 2024, Chinese universities added 91 new AI degree programs.

UK: Research-oriented, with Cambridge, Edinburgh, and Imperial College strong in generative AI foundational research and graduate education.

Switzerland: ETH and EPFL are world-class in AI research and teaching, and are rapidly updating their courses toward generative AI.

Third Tier: Regular Chinese Universities + Singapore and Others ​

Regular Chinese universities are growing fast in the number of degree programs, but faculty, course quality, and practice environments vary widely. Singapore has mandated university-wide AI coursework, but its domestic generative AI research largely follows the US.

Fourth Tier: Southeast Asia, the Middle East, Parts of Europe ​

There is willingness and policy push, but the foundations are thin—short on faculty, teaching materials, practice environments, and methodology. UNESCO's AI competency framework for teachers is also trying to help these regions keep pace.

The Real Barrier Isn't in the Curriculum ​

After this tour, We want to emphasize one point:

Everyone is discussing "what to teach," but what actually blocks many schools is "how to get students up and running."

The content of generative AI courses is public—Stanford's syllabus, MIT's lecture slides, various open-source tutorials are all available online. The real barrier is: do you have a stable, usable, maintainable practice environment that lets students actually get hands-on and quickly validate ideas?

In the traditional ML era, this wasn't a big deal—just pip install and you could go from idea to validation in half an hour.

In the generative AI era, this problem is amplified: models are tens or hundreds of GB, dependency relationships are so complex that installation often fails, quantization and VRAM and inference runtimes are all new things, and the technology iterates so fast that a course may be outdated as soon as it's written. A student trying to validate an idea may have to spend a day setting up the environment first.

And the plight of non-technical majors is more acute—design students need to install ComfyUI, media students need to install video generation tools, architecture students need to install ControlNet workflows, finance students need to build RAG systems. Many of them have never used a command line.

Quantization: Let the Teaching Environment Return to Teaching Itself ​

We believe there's a key practical principle that many schools haven't yet realized: in the teaching process, you don't need to use the original full-precision model every time.

Original (FP16/FP32 precision) models do give the best results, but they also demand the most VRAM. A 14B original model needs 28GB of VRAM—most schools' machines can't run it. But with INT4 quantization, it only needs about 7GB—and a regular RTX 3060 can run it. The same goes for image generation models: the FLUX.2 [klein] 4B version needs only 8.4GB of VRAM, and the quality is already sufficient for teaching.

The goal of teaching is for students to understand principles, master the workflow, and be able to produce things—not to chase inference quality identical to the paper. Choosing an appropriately quantized model to deploy locally, based on actual hardware configuration and course needs, is a more pragmatic and sustainable approach. A lab equipped with RTX 3060s running a quantized 7B LLM, a FLUX 4B image model, and a Wan 2.2 5B video model can already cover most generative AI teaching needs.

We noticed there are now some tools trying to solve the environment setup problem—Ollama simplifies running large models locally, ComfyUI turns the image generation process into nodes, and there are also some all-in-one AI application deployment tools doing something like "one-click install, ready to use." The very emergence of these tools shows that the industry has recognized "environment setup" as one of the core bottlenecks for generative AI to actually land. The market already has solutions that support rapid deployment of original, quantized, and accelerated versions of common models, which is especially valuable for teaching scenarios—schools don't need to figure out how to quantize each model or configure each dependency themselves; they just need to choose the right version based on hardware conditions.

But the existence of tools is only one side. We believe what's more critical for schools is integrating these tools into the teaching workflow—so that environment setup is no longer a prerequisite obstacle to practical courses, and students can spend their time where it really matters: training models, validating ideas, and iterating on solutions.

Generative AI is rewriting every link of AI education. Courses are changing, tools are changing, and deployment is changing too. We hope this article helps you see the direction of the change—and, amid the change, the pain points that remain constant.