A brutal, opinionated syllabus for LLM alignment research

What it does This repository is a nine-chapter curriculum that maps out the papers, concepts, and coding exercises needed to move from basic deep learning familiarity to the frontiers of large language model research, with a heavy emphasis on alignment, interpretability, and scaling laws. Each chapter bundles a tight list of recommended reading, optional references, and a suggested implementation exercise. The author estimates each chapter requires between half a week and two weeks of full-time study.

The interesting bit Rather than pretending to be a neutral survey, the curriculum openly admits it is “heavily biased” toward the author’s own research interests and explicitly warns that it may be “extremely challenging to take on alone.” It also ranks the chapters by importance—transformers and reinforcement learning come first, while training at scale and modeling objectives are treated as lower priority—so a student can triage their time.

Key highlights

• Nine chapters covering transformers, scaling laws, optimization, RL, alignment, interpretability, and adversarial training
• Every chapter includes curated reading lists and a suggested implementation exercise, with community solutions collected in a shared page
• The author recommends a non-sequential order of attack (1, 6, 2, 7, 4, 8, 9, 3, 5) rather than reading front-to-back
• Explicitly targeted at people with a strong quantitative background who already know basic deep learning and PyTorch
• Offers alternative textbooks and courses for those who lack a mentor or study partner

Caveats

• Frozen as of July 2022; the author states they do not intend to keep it up-to-date, though selective PRs are welcome
• Extremely difficult without mentorship—the README includes a prominent warning that solo students should expect to struggle
• The narrow focus on LLM alignment means it skips or deprioritizes large swaths of mainstream deep learning

Verdict Worth bookmarking if you are a quantitatively strong developer aiming to break into technical AI alignment research and can find a study partner or mentor. Look elsewhere if you need a gentle, maintained, or broadly general introduction to deep learning.