Stanford's new CS 312 'Deep Learning Alchemy' trains intuition via prediction quizzes
stanfordnlp · x · 2026-10-15
Stanford's new fall course CS 312 "Deep Learning Alchemy," taught by Tatsunori Hashimoto and Suhas Kotha, confronts an awkward truth: deep learning's core knowledge is empirical, and neither theory nor paper-reading replaces hands-on work.
Two tenets
- Alchemy is learned by hand: run clean experiments to build intuition
- Prediction is understanding: you've mastered a phenomenon only if you can predict experiment outcomes before running them
Course arc (7 units)
Hyperparameter tuning & scaling laws → optimizers and hyperparameter invariants (NQM, μP, Muon, weight decay) → loss landscapes → training stability → architecture capacity → data-efficient generalization (switching to DNA modality) → an unrevealed unit. The reading list is a map of empirical deep learning, with authors of the μP/Muon/weight-decay lines on staff.
Format & openness
- Learning loop: run small-LM experiments → discuss assumptions → predict held-out outcomes from unseen code diffs → revise mental models
- 7 hard prediction quizzes; 85% of the grade comes from quizzes, deliberately exam-heavy with no grade inflation
- All assignments, handouts and quizzes are public; lectures recorded via Stanford Online; self-learners can reproduce experiments on Modal GPUs with a depth-8 Transformer, 614M-token baseline
- Positioned as the sequel to CS 336: that course builds LMs, this one teaches understanding model behavior
Related event: Stanford CS312 'Deep Learning Alchemy' Course Sparks Teaching Debate(8 posts)→
More from Companies & People
- SpaceXSI hiring a solo Tableau admin to power data decisions across X and Grok — Kyrannio · 2026-10-15
- NVIDIA's open-sourced CAD kernel explained: code dating back to the 1970s — rms80 · 2026-10-15
- Modal culture lesson: grow people 4-5x a year instead of hiring to scale — charles_irl · 2026-10-15
- a16z gutted its 34-person talent team, signaling tougher times ahead — SumitGup · 2026-10-15
- Northwood Space CEO Bridgit Mendler wants 1,000 SpaceXs in five years, not 25 — a16z · 2026-10-15
- McKinsey makes entry-level job candidates spar with a fake AI client in interviews — The Verge AI · 2026-10-15