Recommending a fifteen year old AI course in 2026 requires justification, so here it is. Almost every current AI course teaches you what to do. Patrick Winston taught what the problem is. That distinction has become more valuable, not less, as the tools have grown more powerful and the thinking behind them has grown shallower.
Winston, who led the MIT AI Lab and taught 6.034 for decades before his death in 2019, built the course around a claim he repeats throughout: that intelligence is inseparable from representation, and that the right representation makes a hard problem easy while the wrong one makes it impossible. Every technique in the course gets introduced as a consequence of a representational choice. Goal trees, search spaces, constraint graphs, semantic nets, frames, identification trees. Watching him work through this repeatedly builds an instinct that survives the obsolescence of any particular method.
The search and constraint material remains directly useful. Winston's treatment of heuristic search, branch and bound, and constraint propagation is clear, well motivated and applicable to work people are doing right now with agent planning and structured generation. His explanation of why a good heuristic transforms a search problem is one I have never seen bettered. The machine learning content is where age shows badly, and I will not pretend otherwise.
There is a lecture on neural nets that is essentially historical. Support vector machines and boosting get proper treatment and were state of the art when recorded, and boosting is still practically relevant, but you would not learn modern machine learning here. Anyone who watches the 2010 neural network lecture expecting to understand contemporary deep learning will be badly misled. Take this course for the first two thirds and get your learning material from CS231n or 6.S191.
The famous outlier is the lecture on how to speak, which Winston delivered annually and which has nothing to do with AI. It is about giving talks and it is genuinely excellent. Anyone whose job involves presenting technical work should watch it regardless of any interest in this course. There is also a good closing discussion of what the field was trying to do and why several of its confident predictions failed, which is worth sitting with given the current volume of confident predictions.
Practically, the OpenCourseWare version is complete: lecture video, recitation notes, assignments, quizzes and exams, all free with no registration. The assignments use Python from the era and are somewhat awkward to get running now, which is a minor irritation rather than a barrier. What I would say to someone deciding whether to spend forty hours here: if you already have a modern technical foundation and you want to develop judgement about how to frame problems, this is a genuinely good use of your time and there is nothing comparable being recorded now. If you are early in your learning and looking for skills you can apply this quarter, take CS221 instead, which covers overlapping ground with current material and a working assignment structure.
My 4.0 balances exceptional teaching against genuine obsolescence in a significant portion of the syllabus. It is a great course that is partly out of date, and both halves of that sentence are true.