Foundation Models · Roadmap¶
Short on time?
Do the two-week Get Started first: it compresses stages 0, 2, 3 and F2, F5 into one end-to-end project; then choose a direction below.
Two backbone courses, one implementation track¶
Lee's Chinese overview → Karpathy's Zero to Hero. Build the system picture before modifying models. Estimates assume 8–10 hours per week; use experimental results to set the pace.
| Stage | Main resource | Scope for your background | Practice and completion |
|---|---|---|---|
| 0 · Overview, 1–2 weeks | Lee's introduction | Lectures 1, 6–10 and 12: systems and limitations | F0: explain training, inference, retrieval and tools |
| 1 · Diagnostic, 1–3 sessions | PyTorch Quickstart | Explain training, validation and save/load; study Basics only for gaps | F1 as needed; skip already-mastered Python and mathematics |
| 2 · Implementation, 2–4 weeks | Zero to Hero | Micrograd if backprop is unclear; first two makemore lectures, then GPT; revisit other lectures as needed | F4 first half: small GPT, curves, tensor shapes and one controlled change |
| 3 · Pretrained models, about 1 week | HF LLM Course chapter 2 as interface reference | Tokens, loading, context and generation; not a second full course | Use a model and explain differences from your own implementation |
| 4 · Application exploration, 1–2 weeks | One official tutorial in F2 or F3 | Choose document Q&A or image Q&A | Fixed examples with successes and failures |
Read the embodied overview alongside stage 0. F1 is shared preparation and only needs completing once.
Choose depth after practice¶
| Direction | Next step | Resource |
|---|---|---|
| Model algorithms | Finish F4, then adapt to one task | HF fine-tuning; CS336 later for training research |
| Applications and agents | Extend F2 with one verifiable read-only tool | One project entry in Applications |
| Inference systems | F5 quality/latency/memory comparison | llama.cpp in Inference |
| Multimodality and robots | E3, then select E4 | LeRobot in the embodied roadmap |
Text-based alternative: chapters 2–5 of LLMs-from-scratch can replace Karpathy's implementation track. Choose a medium and stay with its codebase.
MoE, full preference optimization, large-scale parallelism, world models and multi-agent systems remain electives. Topic pages are not additional required courses.
Learning and career decisions¶
Complete both overviews and one implementation project, then compare 3–5 target roles. Classify skills as explain, reproduce or independently improve; use project evidence to select the next step.
Run the original example, change one factor and evaluate on examples not used for tuning. Keep versions, results and engineering decisions.