Stanford's New AI Course: It Doesn't Teach Models—Your Homework Is Your Real Work

Stanford's CS department launched a new course this year: CS193T: Thinking with AI. The course does not teach model principles, nor does it teach machine learning. It focuses on one thing: how to use AI smoothly in real work.

The course website is already public.

The course covers very concrete content:

• How AI actually works, first building intuition
• Advanced Prompt techniques for probing the model's capability boundaries and weaknesses
• Agentic AI and reusable complex workflows. Agentic can be understood as letting the model call tools and execute tasks step by step
• Token efficiency, budget management, and environmental costs. Tokens are the basic unit for model billing and context length
• How to evaluate and audit AI systems; in the course this corresponds to Accuracy and Accountability
• Social issues such as bias, copyright, and labor impact

The tool stack directly targets the industry frontline: Claude's Chat, Cowork, and Code, ChatGPT, Gemini Enterprise, NotebookLM, open-source models such as DeepSeek and Llama, and hands-on work with Connectors such as Google Drive and Slack.

The instructor is Cynthia Bailey.

Several design choices in this course are worth a closer look.

Zero prerequisites, 2 credits, S/NC grading. S/NC is similar to pass/fail; the focus is on 'using it,' not on chasing grades. There is an 80-minute tutorial each week: the first 20 minutes are for learning tools, and the remaining hour is for hands-on building under TA guidance. Homework is assigned every week, and the content is simply for you to deploy AI on things you really need to do in your major, career, and life. The course explicitly says it does not expect this course to significantly increase workload.

The final project is to build a customized AI system and put it directly into your job-seeking portfolio; it is presented in a poster session at the end of the term, with no final exam.

The course also specifically states that it will not do mindless AI hype and that it takes individual risks such as learning impairment and emotional manipulation seriously. This is not common in university AI courses.

Some online commenters mentioned that the syllabus separately lists budget management and cost, while they themselves rarely seriously calculate this account when using AI daily. Another reply was very direct: when looking at the API bill at the end of the month, you calculate it more clearly than anyone.

Some commenters most want to see how it teaches people to judge whether what AI gives them is wrong. This point is harder than teaching prompts. Another commenter said that 'homework as real combat' moved him most: homework is about installing AI into things you truly need to do, so after finishing the course you can use it immediately. Someone then asked: if domestic universities offered courses following this approach, what would be hardest—faculty or compute?

Prompt templates are only the surface. Real tasks will expose model boundaries, token costs, and when humans should take over.

Who is this course suitable for? University teachers can see how to implement AI teaching; researchers can use it for systematic training in AI-native workflows; people who want to scale their personal methodology with AI can also get references from the assignment design.

You can learn even if you are not at Stanford. The syllabus, tutorial pages, and assignment materials on the course website are public. All lecture videos are also posted on YouTube.

Course homepage: https://web.stanford.edu/class/archive/cs/cs193t/cs193t.1272/

Lecture videos: https://www.youtube.com/watch?v=I7zvCSGMkoc

Stanford Online credit course entry: https://online.stanford.edu/courses/cs193t-thinking-ai-strategies-effective-and-ethical-human-ai-collaboration

Source: https://x.com/Fred834567/status/2121427595485221173