AI · Roadmap¶
Short on time?
Do the two-week Get Started first, then complete the three rounds below. It covers round two's shared foundations plus one deployment project.
The map provides coverage; this route sets priorities. Reuse your mathematics, signals, geometry and embedded experience, diagnose gaps, implement one model, then explore one or two directions.
Three passes¶
| Pass | Content | Resource scope | Completion |
|---|---|---|---|
| 1 · Overview | AI map, topic tables and frontier index | Purpose, interfaces, relationships and limitations; roughly 1–2 weeks of selected reading | Classify terms as architectures, objectives, tasks or systems methods |
| 2 · Foundation and implementation | P0 data, optimization, evaluation and PyTorch gaps | Existing Karpathy/F4 track; skip mastered material; roughly 2–4 weeks | Run training/validation, change one factor and explain outcomes |
| 3 · One or two explorations | Vision, generation, graphs, RL or scientific AI | One official example per chosen direction, not parallel full courses | Use evidence and interest to choose specialization |
Durations are estimates. Full graphical models, topology, SDEs, SLAM and large-scale pretraining are not entry requirements.
Project pool: choose rather than collect checkmarks¶
| Project | Entry | Prerequisites/resources | Experience |
|---|---|---|---|
| AI-1 baselines | Foundations | Python/splits; CPU | Overfitting, metrics and leakage |
| AI-2 transfer | Representation | F1-equivalent training; GPU convenient | Frozen/full fine-tuning, labels and shift |
| G1–G3 geometry/generation | Generative topics | G1 CPU; G2/G3 probability and PyTorch | Geometry, denoising, fields and sampling |
| AI-3 small GCN | Graphs | Graphs/PyTorch; small CPU | Relations, graph splits and baselines |
| AI-4 CartPole | Decision learning | MDPs/training; CPU entry | Interaction, exploration and repeated evaluation |
| AI-5 signals/physics | Scientific AI | Temporal splits; operators/PINNs later | Research transfer, physical checks and generalization |
| AI-6/AI-7 systems/reliability | Systems / evaluation | Reuse any existing model | Performance, reliability and limitations |
Default suggestion: shared-gap checks and F4 first; then one of AI-2 or G2/G3, plus one of AI-5 or robot E3. Substitute graphs/RL if preferred. AI-6/AI-7 audit existing work rather than add major projects.
Core resources & learning budget¶
-
Backbone · Intuition & Scratch Implementation
Lee's lectures establish the global taxonomy; Karpathy Zero to Hero builds a GPT from tensors.
-
Vision & Representations · CS231n 2026
Read lectures on convolutions, residual connections and self-supervision using public slides.
-
Geometry & Generation · MIT 6.S184
A unified lecture and lab track covering diffusion, flow matching and latent vector fields without notation conflicts.
-
Specializations · Graphs / Decision / Science AI
Enter primary tracks (CS224W / David Silver / NeuralOperator) only after choosing a specialization focus.
Connect to jobs¶
After orientation and an implementation, compare 3–5 actual roles using interest, transferable skills, resources and demonstrable results. Improve one main project; vocabulary familiarity is distinct from independent capability.
Use the foundation roadmap for LLM post-training, RAG, agents and inference, and the embodied roadmap for policies/control/transfer.