Clawdemy Lessons

Planning brief for the WGAN-GP lesson: learning outcomes, prerequisites, and scope for the Wasserstein distance, 1-Lipschitz critic, and gradient penalty.

Brief on training a variational autoencoder: the reparameterization trick, the closed-form Gaussian KL term, and the per-example reconstruction-plus-KL loss.

Orientation for the score matching lesson: the objectives, why the explicit form does not scale, denoising score matching, and the bridge to diffusion.

Overview of the normalizing flows lesson: learning outcomes, prerequisites, where it fits among generative paradigms, and the time and difficulty to expect.

What the maximum likelihood and forward-KL lesson covers: KL divergence, the derivation, prerequisites, learning outcomes, and time estimates.

Overview of the ELBO lesson: what you'll learn, prerequisites, the math tools involved, and how the latent-variable paradigm fits the generative track.

Overview of the GANs lesson: the minimax objective, deriving the optimal discriminator, the Jensen-Shannon divergence, and paradigm-level failure modes.

What the lesson on evaluating generative models covers: likelihood limits, FID and Inception Score, precision and recall, prerequisites, and time to budget.

Overview of the energy-based model lesson: what you will learn, where it fits, prerequisites, and the math, from the EBM density to the score-function escape.

A study guide to the DDPM diffusion lesson: prerequisites, learning outcomes, where it fits among score-based and SDE lessons, and read and practice time.

A brief on factoring distributions with the chain rule, training by negative log-likelihood, and enforcing causality in autoregressive models.

Editorial brief for the variational inference lesson: scope, capability gained, source, and how the ELBO and reparameterization set up control as inference.

Editorial brief for the value-based RL lesson: the Q-branch, Q-learning from Bellman optimality, dual-path Q-iteration, and the deadly triad.

Editorial brief for the RLHF lesson: scope, learning outcomes, source papers, and how the InstructGPT pipeline maps to a variational framework.

A guide to the RL fundamentals lesson: the MDP tuple, value functions, the Bellman equation, a worked 2-state solve, and the deep-RL dispatch table.

Editorial brief for the PPO lesson: how to derive the clipped surrogate from the on-policy stability problem and frame PPO as the RLHF workhorse.

Overview of the REINFORCE lesson: the log-derivative trick, why environment dynamics drop out, a worked sigmoid bandit, and the variance refinements.

Editorial brief for the lesson on planning with a learned model: capability gained, sources, and an artifact-by-artifact outline of the lesson set.

Editorial brief for the model-based RL lesson: learning outcomes, the worked least-squares dynamics fit, compounding-error rollout, and model-class choices.

Orientation to deep reinforcement learning: the ML regimes, the agent-environment loop, returns, discounting, and what makes the deep variant hard.

Overview of behavioral cloning: the algorithm, why it appeals, how distribution shift makes errors compound, DAgger, and where copying an expert suffices.

Editorial brief mapping each DQN trick to a deadly-triad leg, with the closed-form max-overestimation bias worked on a small example.

Editorial brief for the control-as-inference lesson: the optimality-conditioned graphical model, soft Bellman backup, and how SAC, RLHF, and DPO unify.

A roadmap to actor-critic methods: the two-network split, advantage estimators (MC, TD, n-step, GAE), the bias-variance tradeoff, and the modern RL family.

What to expect from the world modeling lesson: the reactive-vs-predictive split, the pixel-vs-latent trade-off, landmark architectures, and prerequisites.