George Hotz to teach CMU course where students build an ML compiler from scratch up to training LLMs on GPUs
QuixiAI · x · 2026-11-05
George Hotz (tinygrad) is teaching 10-411: Compilers for Machine Learning at Carnegie Mellon in Spring 2027. The course website is live; assignments will be posted online but lectures won't be recorded.
The hands-on, one-semester course has students build a compiler from scratch, starting with elementwise programs and ending with training modern LLMs on GPUs. Topics include term rewriting, code generation, movement operators, kernel fusion, memory hierarchies, GPU architecture, automatic differentiation, and flash attention.
There is no starter code — the spec uses a simple interchange format for the compiler IR — and students work individually or in pairs in a language of their choice. No formal prerequisites, but the course assumes deep familiarity with 15-213-level material, C, some GPU programming, and the stamina for a semester-long project.
More from Companies & People
- Bemlo founder joins GetEnergy after bringing AI to Sweden's healthcare regions — gabriel1 · 2026-11-05
- Anthropic Residency opens applications: paid program, no ML background required — JessicaHullman · 2026-11-05
- The Upskilling Company launches AI-native L&D service for team training — Aiden_Tech_Ai · 2026-11-05
- Podcast: ex-OpenAI preparedness lead Aleksander Mądry on the AI arms race — aleks_madry · 2026-11-05
- Stanford study: median unicorn founder age is 34 across 1,540 US unicorns — arthurcolle · 2026-11-05
- AI startup Energy is 4 people with $10M+ deals, says AI does 80% of all tasks — gabriel1 · 2026-11-05