Turn a paper pile into a reading system.

A code-first explanation of the transformer that makes the architecture concrete before heavier papers.

Use it to connect equations, modules, and implementation details.

Builds intuition for why complexity can rise and fall rather than only accumulate.

Read as a conceptual lens for later compression and representation readings.

Shows how simple character-level recurrent models learn surprising structure.

Pay attention to the qualitative behavior of learned sequence models.

Explains the gating mechanism that made recurrent memory easier to reason about.

Understand the cell state and gates before reading recurrent regularization work.

Studies dropout placement for recurrent networks without destroying temporal memory.

Notice how a small regularization detail changes sequence-model training behavior.

Connects neural-network regularization to a compression-style preference for simpler explanations.

Use it as the bridge from practical regularization to MDL-style thinking.

Uses attention to choose positions in the input sequence as outputs.

Read for attention as selection rather than only summarization.

Marks the practical arrival of large convolutional networks on ImageNet-scale vision.

Separate the architecture from the compute, data, and training choices that made it work.

Explores how sequence models behave when the target object is naturally unordered.

Read for the mismatch between model order and problem structure.

Shows a practical route for splitting large model training across accelerator partitions.

Read for the systems bottleneck: scale needs scheduling as much as architecture.

Makes very deep convolutional networks easier to optimize through residual connections.

Understand why identity-style paths help gradients and representation reuse.

Uses dilation to expand receptive fields while preserving resolution.

Read for the tradeoff between local detail and broad context in vision models.

Frames molecule modeling as message passing over graph structure.

Use it to understand graph computation as repeated local communication.

Replaces recurrence with stacked attention and feed-forward blocks.

Spend time on multi-head attention, positional information, and parallelism.

Introduces alignment-style attention for translation models.

Read before or after transformers to see what changed and what carried forward.

Clarifies the role of identity paths in residual network optimization.

Read after ResNet to understand why the skip path details matter.

Introduces a simple module for pairwise relational reasoning over objects.

Track how object representations become inputs to explicit relation computation.

Separates useful latent structure from nuisance detail in a variational setting.

Read for what a representation should keep and discard.

Combines recurrent state with relational computation over memory slots.

Compare this with plain LSTMs and later attention-based memory.

Turns the complexodynamics intuition into a small formal model.

Use it as a theory-side complement to representation-learning papers.

Augments neural networks with differentiable memory access.

Read for the memory interface and addressing mechanism.

Demonstrates end-to-end speech recognition at large data and compute scale.

Read for the scaling and systems recipe, not only the acoustic model.

Quantifies how loss varies with model size, data, and compute.

Focus on the compute allocation mindset and its assumptions.

Introduces model selection as compression in a clear tutorial form.

Read before making broad claims about compression and intelligence.

A dissertation-length treatment of machine intelligence and capability framing.

Skim for definitions and assumptions rather than implementation technique.

Provides the mathematical background for complexity and randomness claims.

Use selectively; the list points readers to later sections, not necessarily cover-to-cover reading.

A practical course bridge for convolutional networks, optimization, and visual recognition.

Use it to ground the vision papers in implementation and training practice.