Stable Learning

This is a collection of learning materials maintained by the lab. Each topic is structured as a progressive reading path with interactive visualizations. Pick a topic below and start from the beginning, or jump to any section.

Research 101

How to choose problems and how to work — the advice we give every new lab member, before their first project.

Prerequisites: none. If you are new to the lab, read this track first.

1. How to Do Research Start with why (and whether) to do a PhD at all; then three principles with live simulators — work on the frontier, bet on what scales, treat research as a race; and the whole-person game of taste and stamina.

Reinforcement Learning

From MDPs to modern policy optimization for LLM alignment.

Prerequisites: calculus, basic probability, familiarity with neural networks.

1. Action Chain & Rewards Watch a live agent teach itself a gridworld, then build the MDP vocabulary to describe what it's doing.

2. Value-Based Methods Bellman equation, Q-learning, and DQN — learning the policy through value functions.

3. Monte Carlo Tree Search UCT, the four-phase MCTS loop, a worked tic-tac-toe example, and AlphaZero.

4. Policy Gradient REINFORCE, the score function estimator, and variance reduction.

5. PPO Clipped surrogate objectives and GAE — run the update loop live and toggle the clip off to watch training collapse.

6. GRPO Group Relative Policy Optimization — watch groups of rollouts teach the policy, then stop teaching as prompts get solved.

Speculative Decoding

How modern LLM serving generates 3–6× faster without changing a single output token — from the lossless acceptance rule to EAGLE, DFlash, and DSpark.

Prerequisites: transformer basics (attention, KV cache, logits). Pairs well with the RL track’s LLM context but does not require it.

1. Fundamentals Watch a draft model race the target model live, prove the two-line rule that makes it lossless, and price the speedup.

2. The EAGLE Line Draft from the target's own features, grow confidence-priced draft trees, and the training-time test that unlocked data scaling.

3. Parallel Drafting DFlash drafts whole blocks with a diffusion model; DSpark repairs the block's joint (the mean-field trap) and meters verification per request under real load; MTP as the one-axis frame for every drafter.

Cache Coherence & Consistency

How multicore CPUs keep caches correct, from hardware protocols to memory fences.

Prerequisites: basic computer architecture (registers, memory, assembly helps but not required).

1. Cache Fundamentals Memory hierarchy, cache organization (lines, sets, ways), and read/write policies.

2. Coherence Protocols MSI, MESI, MOESI — the problem of stale data and how snooping solves it.

3. Consistency Models Sequential consistency, TSO (x86), relaxed models (ARM, RISC-V), and why ordering matters.

4. Modern Systems Directory-based protocols, NUMA, memory fences, and acquire-release semantics.