Stanford to Offer New CS 312 'Deep Learning Alchemy' Course in Fall

Stanford University will offer an autumn new course CS 312 "Deep Learning Alchemy," taught by two instructors, Tatsunori Hashimoto and Suhas Kotha. The course addresses a common problem: most of deep learning's core knowledge is empirical, theory cannot explain it, and reading papers cannot replace hands-on work.

The course takes two creeds as its fulcrum: alchemy must be learned by hand; prediction is understanding, meaning only being able to accurately predict results before running experiments counts as mastery. Put together, these two sentences constitute a testable teaching standard—not "do you understand backpropagation," but "can you say what it will do before the next experiment."

Hashimoto's own formulation is: deep learning is often taught as a sanitized set of facts and mathematics, but work at the AI frontier is messy and empirical, and the new course aims to teach the art and science of deep learning in the hands-on spirit of "Language Modeling From Scratch." What is noteworthy here is the wording—"art and science" placed side by side, rather than treating the empirical as a temporary substitute when theory falls short.

The course is arranged into seven units: hyperparameter tuning and scaling laws, optimizers and hyperparameter invariances (NQM, μP, Muon, weight decay), loss landscapes, training stability, architectural capacity, data-efficient generalization (switching to the DNA modality), and one undisclosed unit.

This main line is itself informative. It does not start from architecture or mathematical foundations, but cuts in from hyperparameter tuning and scaling laws, then enters optimizers and hyperparameter invariances—names like μP, Muon, and weight decay are all objects that have repeatedly appeared in recent empirical research.

Judging from the unit order, the course's center of gravity falls on "what happens during training" rather than "what a model is." The three sections on loss landscapes, training stability, and architectural capacity are linked together, pointing to the same class of question: why the same recipe sometimes works and sometimes does not.

The data-efficient generalization section explicitly switches to the DNA modality, the only jump outside the language modality among the seven units. As for the last undisclosed unit, there is currently no content explaining it.

It should be noted that the above unit arrangement comes from a single secondhand relay and has not received first-party confirmation; the specific content of the "undisclosed unit" is likewise unclear.

The relay mentions that the reading list is regarded as a map of empirical deep learning, and that multiple authors (of the μP, Muon, and weight decay series) are on the team.

If this description holds, it explains why a new course can directly use these topics as its skeleton: the teaching materials are not a secondhand review compiled after the fact, but organized by participants in the relevant work. However, the specific referents and scope of "multiple authors are on the team" are not clear, and this point can only be regarded as the relayer's statement.

The course address is deep-learning-alchemy.github.io.

Assignment code, quizzes, and lecture notes are all public, and course recordings are released by Stanford Online. For a course centered on hands-on work, public materials mean the course's core assets—those assignments requiring repeated trial and error—can be fully obtained from outside.

The relay also states that self-learners can use Modal to rent GPUs and fully reproduce the course content. The significance of this path is that it turns compute from "whether you have a lab" into "whether you are willing to pay," but its feasibility and cost have not been verified and can only be regarded as the relayer's statement.

Whether the course recordings are all public and when they will be released are also currently uncertain.

The course is regarded by the relayer as a sequel to CS 336 (Building LMs From Scratch): that course teaches building models, this one teaches understanding model behavior.

If this positioning holds, the two courses form a complementary pair: the former solves "can it be built," the latter solves "why does it behave this way after being built." But what needs to be distinguished is whether this connection is the relayer's judgment or an official positioning, which currently cannot be confirmed.

Judging from the course units and creeds, the course emphasizes experimental predictive ability rather than theoretical derivation as the standard for mastery—this is an inference based on existing materials, not the course provider's own words.

The relay states that the course assessment is 7 exams accounting for 85% of the grade, plus 15% for offline discussion assessment.

If this ratio is accurate, it means the course uses exams almost entirely as its main evaluation method, and the number of exams is as many as 7—roughly corresponding to the seven units. Offline discussion accounts for only 15%, indicating that the course does not intend to use projects or open-ended assignments to carry the main evaluation.

But this assessment structure comes from a single secondhand relay and has not received first-party confirmation; whether the specific ratio is accurate still needs to be held in reserve.

What can currently be confirmed is: Stanford University will offer an autumn new course CS 312 "Deep Learning Alchemy," taught by two instructors, Tatsunori Hashimoto and Suhas Kotha.

The remaining details—the course motivation and two creeds, the specific arrangement of the seven units, all materials being public, recordings released by Stanford Online, Modal GPU rental reproduction, the sequel relationship with CS 336, 7 exams accounting for 85% and 15% offline discussion assessment, and reading list authors being on the team—all come from a single secondhand relay and belong to source claims rather than first-party confirmation.

In addition, whether the course has been officially approved for offering by Stanford currently rests only on a social media announcement; the post was published on October 14, 2026, and the course is an autumn new course, but the specific start date is not stated. Enrollment numbers, prerequisite requirements, and credit information are all not provided, and there is no benchmark testing or quantitative effectiveness data of any kind.