AI/LM curriculum: demonstrate your understanding
For readers with undergraduate mathematics and programming experience. Follow dependencies, pass the checkpoints, then specialize in research or engineering.
Entry diagnostic
Before starting, explain matrix multiplication shapes, conditional probability, a derivative and a train/test split. If you cannot implement a stable softmax or describe data leakage, begin at stage 1. Familiarity with Python or JavaScript is sufficient; GPU access is not required for local teaching labs.
Diagnostic: why not split adjacent video frames at random?
Nearby frames share scene, actor and temporal information, making the test set too similar to training. Split by episode or scene before preprocessing; keep a distinct distribution-shift test.
- 1
Math & numerical foundations
Derive softmax cross-entropy gradients; verify central differences and log-sum-exp stability.
Study & experiment →Evidence checkpointCompare analytical and numerical gradients over several step sizes; explain cancellation at tiny steps.
- 2
Architecture & representation
Trace tensor shapes through attention, masks, residuals and projections. Distinguish representation quality from pretraining loss.
Study & experiment →Evidence checkpointWrite a causal-mask test, then explain why padding and causal masking solve different problems.
- 3
Training & adaptation
Connect optimizer state, gradients, activations and parameter precision to a memory ledger.
Study & experiment →Evidence checkpointCompare full tuning and LoRA with identical data splits; report quality, peak memory and trainable parameters.
- 4
Retrieval & evaluation
Separate retrieval recall, answer grounding and evaluator reliability. Prevent document leakage.
Study & experiment →Evidence checkpointHold out whole documents; compare chunk size and overlap independently; label retrieval and generation errors.
- 5
RL & decision making
Derive Bellman targets, policy gradients and clipped objectives before studying preference optimization.
Study & experiment →Evidence checkpointConstruct one reward-hacking example and one control experiment that distinguishes learning from evaluator exploitation.
- 6
Serving & systems
Account for KV memory, queueing, prefill/decode and tail latency under a declared workload.
Study & experiment →Evidence checkpointDeliver a latency-versus-load curve with p50/p95/p99, warmup policy and failed-request counts.
- 7
Embodied intelligence
Connect perception uncertainty, coordinate frames, action contracts and closed-loop evaluation.
Study & experiment →Evidence checkpointImplement transform round trips and a delayed-feedback toy controller before proposing hardware tests.
- 8
Research synthesis
Compare competing explanations, fair tuning budgets and reproducibility constraints.
Study & experiment →Evidence checkpointSubmit a negative or positive result with raw traces, versioned configuration and one falsifiable next hypothesis.
Three capstone projects
Engineering: build a retrieval pipeline and show which errors come from search, generation and judging. Research: reproduce one narrow claim with a baseline, two ablations and uncertainty estimates. Physical AI: build a simulated perception-to-control loop and test calibration error, sensor delay and action saturation independently.
Assessment: score derivation, implementation, evaluation and explanation separately from 0–2. A stage is complete only when every dimension is demonstrated independently (2), with reproducible evidence. Reading completion alone is not mastery.
Coverage and remaining depth
Available: local teaching labs, bilingual paper explanations, RL chapters and an introductory robotics sequence. Still limited: distributed-training implementations, full-scale model benchmarks, hardware robotics validation and complete translations beyond English/Chinese. The systems track below adds analytical engineering exercises; it is not a GPU benchmark service.