Skip to main content

Shareable map · Bookmark this page

ML Theory Roadmap

The whole curriculum on one page, from measure-theoretic foundations through modern deep learning and the research frontier. Tier-1 landmarks are the 198 core pages worth reading first.

Layer 0A · Axioms

60 topics

Sets, functions, logic, linear algebra, real analysis, measure-theoretic basics.

Foundations

Numerical Optimization

Layer 0B · Infrastructure

26 topics

Measure theory, functional analysis, convex duality, numerical foundations.

Layer 1 · Core Tools

75 topics

Concentration, estimation, information theory, optimization primitives, CLT.

Statistical Foundations

Sampling MCMC

Layer 2 · Learning Theory

163 topics

ERM, VC, Rademacher, PAC, stability, kernels, uniform convergence.

ML Methods

Applied Statistics

Predictive Uncertainty

Layer 3 · ML Methods

152 topics

Regression, SVMs, neural nets, optimization, regularization, NTK.

Layer 4 · Deep Learning

113 topics

Transformers, attention, training dynamics, double descent, scaling.

Statistical Foundations

Applied ML

Layer 5 · Frontier

66 topics

RLHF, alignment, interpretability, reasoning, agents, scaling laws.

Modern Generalization

LLM Construction

How to use this map

  • Amber dots are tier-1 landmarks. Read these first.
  • Each page links down to its prerequisites and up to what builds on it. No concept floats without grounding.
  • Use the gap finder to pick a destination and get a BFS-ordered reading list.
  • The interactive graph gives you the same graph with click-to-explore and path tracing.

Planned additions

Topics in progress, primarily AI safety and alignment.

  • Scalable oversight. Bowman et al. 2022, debate and market-based precedents, sandwiching experiments. Scope conditions matter: what the setup can and cannot tell us.
  • Deceptive alignment. Hubinger et al. 2019/2021 mesa-optimizer framing. Separate the empirical evidence from the philosophical argument.
  • Alignment faking. Greenblatt et al. 2024 (Anthropic). Include the limitations section explicitly.
  • DPO. Currently folded into dpo-vs-grpo. Deserves its own page: Rafailov et al. 2023, the implicit-reward view, and the overoptimization story. Follow-up on IPO, KTO, SimPO and the broader DPO family.
  • Verifiable-reward RL (RLVR). Reasoning training with programmatically checkable rewards: math graders, code executors, proof verifiers. Scope what verifiers can and cannot certify, and the reward-hacking surface when the verifier is imperfect. Needs careful separation from general RLHF.
  • Inference-time scaling beyond CoT. Budgeted search, verifier-guided decoding, reward-model reranking, parallel sampling with aggregation. Current inference-time-scaling-laws page covers the scaling story; deserves a systems-level companion on how the compute is actually spent.
  • Agent systems as systems. Long-horizon tool use, failure recovery, memory design, evaluation under distribution shift, benchmark contamination. Current agent pages cover the components; a systems-view page on how they compose and fail in production is missing.
  • Weak-to-strong generalization. Burns et al. 2023 (OpenAI). What the setup can and cannot tell us about alignment at scale.
  • Instrumental convergence. Omohundro, Bostrom framings. Flag explicitly where the philosophical argument outruns the empirical support.
  • Jailbreaks. Attack taxonomy, measurement difficulties, why robust alignment is not a solved problem. Needs honest threat-model scoping, not incident anecdotes.
  • Superposition. Elhage et al. 2022 toy-models paper, the interference vs capacity trade-off, and the connection to sparse autoencoders.