The Wiki
Core6 min read

What DL left out

The honest map of gaps in this middle shelf — state-space models, multimodal fusion, and every place the story continues elsewhere.

Twenty-four lessons and a shelf of artifacts is a solid spine, but deep learning is more than one shelf. The gaps worth naming, so that nothing here sounds complete when it isn't:

Still moving without this book

  1. State-space models / Mamba lineage. The RNN memory bottleneck died under attention's parallelism, but selective linear state spaces revive a parallel-trainable, linear-cost recurrence — the current serious rival to attention on long streams.
  2. Self-supervised pretraining at large (masked image models, contrastive learning): the autoencoder instinct generalized far past reconstruction — predict-part-of-the-input as the loss family at scale.
  3. Multimodal fusion architecture. This book served vision and sequence models separately; the multimodal lesson in the LLM book is where they actually merge (patch tokens, shared attention).
  4. Reinforcement beyond the bandit frame. The ML book closed its RL door honestly into RLHF; if you want the full control-theoretic RL spine (planning, exploration outside bandits), a dedicated RL shelf is open future work here.

What this shelf deliberately claimed

  • Transformers from their attention core — chapter 4 gave the bridge; the architecture, multi-head mechanics, position systems, and everything after belongs to the LLM book, because the enormous-scale consequences live there.
  • Speech, waveform modeling, and conventional video stacks — cousins of what's covered (conv + sequence + diffusion covers the ideas), with domain-specific structure this shelf doesn't unpack.
  • Hardware arithmetic. We counted conv MACs by hand; where memory bandwidth, quantization, and kernel scheduling decide what you can actually afford — that efficiency is the LLM book's making-it-efficient chapter.

Where to go next, by want

If you want... → read next if the story of how language rivals learned everything → the LLM book's foundations to place everything above in the working-engineer's ledger → the ML practice chapters to see the same discipline applied around design → Systems Design (coming soon)

Where you are, and where each want leads.

neurontrainingvisionsequencesgenerative→ next shelf
The LLM book: tokens → attention → the whole modern stack“I understand the ideas — show me where language models scale to brains-in-a-box”The ML book: learning as a discipline“I want the working-engineer's ledger: metrics, splits, imbalance, regularization”Systems Design: coming soon“I want to place sequence models in the bigger machine-design picture”A dedicated RL shelf: future work, noted honestly“Give me the other RL paradigm properly, beyond the bandit frame”

Not a computation — a junction. The DL shelf’s last artifact is, deliberately, a door.

Illustrative vs real

The artifact below is a map of advice, not of computation — the first one on the wiki without a dial, and that's deliberate: the honest endpoint of this shelf is knowing where the other shelves are.

This lesson has exercises attached — matching stated goals to the right continuation lesson — launching once the exercises layer ships.