Stanford's deep learning course: homework doesn't count, and exams make you guess experimental results

Stanford has a CS 312 course in fall 2026 whose name translates literally to Deep Learning Alchemy. A post on X mentioned that the poster saw researchers from several leading labs taking it, including DeepMind, Pluralis, and Sakana AI Labs. Original post link: https://x.com/alacheng/status/2121534854521029060. Course website: https://deep-learning-alchemy.github.io/.

The most unusual thing about this course is the grading. Homework accounts for 0%; it is only used to cite in quizzes for partial credit. 15% comes from tutorial participation, and 85% comes from 7 quizzes, of which you can drop 2. The website states upfront that this is a 5-unit course, requires running a large number of experiments, the 7 quizzes are not easy, and there will be no curve. Better think carefully before enrolling.

The instructors are Tatsunori Hashimoto and Suhas Kotha, and the CAs are Tristan Thrush and Kaiyue Wen. Classes are Tuesday and Thursday from 1:30 to 2:50 p.m. in CoDa B90. Lecture recordings and course materials will be made public.

The course page makes two beliefs very clear: Alchemy is learned by hand and Prediction is understanding. The gist of the official wording is that deep learning contains many mysteries that even top researchers have not fully figured out; the mental models taught in many courses do not match experimental experience—for example, the classic bias-variance tradeoff is inconsistent with the current scaling era. So this course has students design clean experiments themselves and first build intuition about phenomena. Exams reverse this: they give you an experiment you have not run and ask you to predict the result.

Quizzes are 40-minute experimental predictions. The problems give an experiment description, including a code diff, and students must predict what result the experiment will produce. You cannot consult external materials, not even your own reports, but you can cite reports in your answer, and sound reasoning can earn partial credit. In the sample questions, one asks you to use a depth 8 transformer for leave-one-out ablation and compare the loss ranking after removing prenorm, residual, MLP, attention, and positional embeddings. Another is harder: change the residual so that it takes 0.5 from the previous layer and 0.5 from the layer before that, and predict the loss difference. The website itself warns that both are difficult, and it is normal to lack intuition if you have not run the experiments.

The homework topics are also specific. The first 5 assignments center on small-scale language model pretraining, from hyperparameters and scaling, optimizers, and loss landscapes to depth stability and model capacity. The 6th assignment switches to DNA. Homework is not graded, but it forces you to run experiments, make plots, and write reports. There is also a 30-minute tutorial each week; each student meets with three other students plus one course staff member to discuss experiments currently being run and do quiz practice problems. Tutorial participation earns the full 15%.

This course allows AI tools for homework—implementation, visualization, and simulating quizzes are all fine. But quizzes cannot use external resources. The design is a bit ironic: most courses worry about students using AI to do homework; this course sets homework weight to 0 and then uses quizzes to confirm how much intuition is left in students' heads.

For AI practitioners, this course addresses a daily pain point: tuning hyperparameters and changing architectures often rely on a feel for things, and that feel is hard to articulate. The course breaks intuition down into predictable experimental questions. For general readers, alchemy here refers to experimental intuition, not an insult. The course also says you do not need to have thoroughly learned deep learning first; the homework comes with scaffolding, but the beginning will be more tiring. You need calculus, linear algebra, and basic probability and statistics for the math, and Python for the code.

Following the course at home requires a GPU. The course page says d8 training commonly uses a single NVIDIA H100, run on Modal. The public price as of October 2, 2026 is $3.95 per hour for an H100 SXM5, billed by the second; the Starter plan includes $30 of credit per month. Modal sponsors computing for this course. Course materials are public, so those who want to follow along can run things themselves.

Who it suits: AI practitioners who want to build experimental intuition, people willing to turn their tuning feel into testable questions. Who it does not suit: people who want to learn the implementation details of backpropagation, who want easy credits, or who only want to read a theory review. The quizzes test prediction, not memorization.