Reading order

Papers

36 landmark papers in three tracks, sequenced so each one prepares you for the next. Read in order, write a summary in your own words, and mark what you finish.

Sign in to track reading

AI research

From the first deep convolutional net to preference optimisation, in the order the ideas were needed.

  1. 01

    ImageNet Classification with Deep Convolutional Neural Networks

    Krizhevsky, Sutskever, Hinton · 2012 · NeurIPS · 35 min · Approachable

    This is the paper that made scale plus GPUs the default strategy. Read it to see how few ingredients the modern deep learning stack started with.

    visioncnngpu
  2. 02

    Adam: A Method for Stochastic Optimization

    Kingma, Ba · 2014 · ICLR 2015 · 30 min · Moderate

    Almost every model you will read about after this was trained with Adam or a close variant. Understanding the two moment estimates explains most optimiser folklore.

    optimisationtraining
  3. 03

    Deep Residual Learning for Image Recognition

    He, Zhang, Ren, Sun · 2015 · CVPR 2016 · 35 min · Moderate

    The skip connection is the single most reused architectural idea in the field. Transformers, diffusion U-Nets, and modern vision models all depend on it.

    visionarchitecturedepth
  4. 04

    Attention Is All You Need

    Vaswani et al. · 2017 · NeurIPS · 60 min · Dense

    This is the architecture the rest of the track builds on. Read it slowly; the scaled dot-product attention equation is the one formula worth memorising.

    transformersattentionnlp
  5. 05

    BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding

    Devlin, Chang, Lee, Toutanova · 2018 · NAACL 2019 · 40 min · Moderate

    It established the pre-train then fine-tune workflow that dominated applied NLP for years and still shapes how embeddings are built.

    pretrainingnlptransformers
  6. 06

    Language Models are Few-Shot Learners

    Brown et al. · 2020 · NeurIPS · 60 min · Moderate

    This is where prompting became an interface. Read it to see what few-shot behaviour looked like before instruction tuning made it feel ordinary.

    llmscalingprompting
  7. 07

    Scaling Laws for Neural Language Models

    Kaplan et al. · 2020 · arXiv · 45 min · Moderate

    Every later decision about model size and training length refers back to these curves, including the correction Chinchilla made to them.

    scalingllmcompute
  8. 08

    Training Compute-Optimal Large Language Models

    Hoffmann et al. · 2022 · NeurIPS · 40 min · Moderate

    It corrected the field's most-cited scaling result and explains why later models are smaller and trained far longer.

    scalingllmdata
  9. 09

    Training language models to follow instructions with human feedback

    Ouyang et al. · 2022 · NeurIPS · 50 min · Dense

    This three-stage recipe is what turned raw language models into assistants, and it is the reference point for every later alignment method.

    alignmentrlhfllm
  10. 10

    LoRA: Low-Rank Adaptation of Large Language Models

    Hu et al. · 2021 · ICLR 2022 · 30 min · Moderate

    It made fine-tuning large models practical on ordinary hardware and is the basis of most adapter ecosystems today.

    fine-tuningefficiencyllm
  11. 11

    FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness

    Dao, Fu, Ermon, Rudra, Ré · 2022 · NeurIPS · 50 min · Dense

    It is the clearest example of treating a model as a systems problem. Longer contexts and cheaper training both trace back to this kernel.

    attentiongpusystems
  12. 12

    Direct Preference Optimization: Your Language Model is Secretly a Reward Model

    Rafailov et al. · 2023 · NeurIPS · 45 min · Dense

    It replaced a fragile RL pipeline with one loss function and is now the default starting point for preference tuning.

    alignmentllmtraining

Systems

The papers behind the storage, coordination, and streaming systems this course designs.

  1. 01

    Time, Clocks, and the Ordering of Events in a Distributed System

    Lamport · 1978 · Communications of the ACM · 40 min · Dense

    Every discussion of consistency, causality, and versioning in this course stands on this paper. It is short and every paragraph earns its place.

    orderingclockstheory
  2. 02

    The Log-Structured Merge-Tree (LSM-Tree)

    O'Neil, Cheng, Gawlick, O'Neil · 1996 · Acta Informatica · 50 min · Dense

    RocksDB, Cassandra, and most modern write-heavy stores are LSM-trees. The course's storage-engine module builds one from scratch.

    storageindexingwrite-path
  3. 03

    The Google File System

    Ghemawat, Gobioff, Leung · 2003 · SOSP · 60 min · Moderate

    It is the template for object and blob storage: a metadata service, dumb data nodes, and replication chosen for the workload rather than for generality.

    storagereplicationlarge-files
  4. 04

    MapReduce: Simplified Data Processing on Large Clusters

    Dean, Ghemawat · 2004 · OSDI · 40 min · Approachable

    Counting, aggregation, and analytics pipelines in this course are all descendants of this model, whether they run on Hadoop, Spark, or a warehouse.

    batchanalyticsfault-tolerance
  5. 05

    Bigtable: A Distributed Storage System for Structured Data

    Chang et al. · 2006 · OSDI · 55 min · Moderate

    It defined the wide-column model that HBase and Cassandra copied, and shows how range partitioning and sorted storage fit together.

    storagewide-columnpartitioning
  6. 06

    Dynamo: Amazon's Highly Available Key-value Store

    DeCandia et al. · 2007 · SOSP · 60 min · Dense

    It is the origin of most techniques in the NoSQL toolbox and the clearest worked example of a deliberate consistency trade-off.

    availabilityreplicationconsistent-hashing
  7. 07

    Paxos Made Simple

    Lamport · 2001 · ACM SIGACT News · 45 min · Dense

    Consensus is the foundation of every coordination service and replicated log. Read this before Raft so you can see what Raft was simplifying.

    consensustheory
  8. 08

    The Chubby Lock Service for Loosely-Coupled Distributed Systems

    Burrows · 2006 · OSDI · 50 min · Moderate

    It explains why coordination is offered as a service rather than a library, and it shaped ZooKeeper and etcd.

    coordinationlocksconsensus
  9. 09

    ZooKeeper: Wait-free Coordination for Internet-scale Systems

    Hunt, Konar, Junqueira, Reed · 2010 · USENIX ATC · 45 min · Moderate

    Kafka, HBase, and many in-house systems used it for membership and leader election, and its recipes are still how coordination problems are decomposed.

    coordinationprimitives
  10. 10

    In Search of an Understandable Consensus Algorithm

    Ongaro, Ousterhout · 2014 · USENIX ATC · 60 min · Moderate

    etcd, CockroachDB, and most new replicated logs implement Raft. This is the version of consensus you will actually read code for.

    consensusreplicationleader-election
  11. 11

    Spanner: Google's Globally-Distributed Database

    Corbett et al. · 2012 · OSDI · 60 min · Dense

    It shows what it costs to get strong consistency at global scale, and why bounded clock error is a systems property worth paying for.

    transactionsclocksconsistency
  12. 12

    Kafka: a Distributed Messaging System for Log Processing

    Kreps, Narkhede, Rao · 2011 · NetDB · 35 min · Approachable

    The partitioned log is the backbone of event-driven architectures. The course's queue and stream modules design one from this paper's ideas.

    streaminglogsmessaging

ML math

The results that explain why training works, from information theory to double descent.

  1. 01

    A Mathematical Theory of Communication

    Shannon · 1948 · Bell System Technical Journal · 90 min · Dense

    Every loss function in the track is an information quantity. Reading the source makes perplexity and KL divergence feel obvious rather than memorised.

    information-theoryentropy
  2. 02

    Learning representations by back-propagating errors

    Rumelhart, Hinton, Williams · 1986 · Nature · 25 min · Moderate

    Automatic differentiation frameworks are this algorithm generalised. Understanding the chain of local derivatives explains vanishing gradients and why depth was hard.

    backpropagationgradients
  3. 03

    Gradient-Based Learning Applied to Document Recognition

    LeCun, Bottou, Bengio, Haffner · 1998 · Proceedings of the IEEE · 70 min · Moderate

    Convolution as weight sharing is the mathematical idea behind every vision backbone. Reading the original makes the parameter counting concrete.

    cnnvisionconvolution
  4. 04

    Long Short-Term Memory

    Hochreiter, Schmidhuber · 1997 · Neural Computation · 60 min · Dense

    Gating is the idea to extract: multiplicative control over what is kept and what is forgotten. It reappears in GRUs, highway networks, and attention itself.

    recurrentgradientssequence
  5. 05

    Dropout: A Simple Way to Prevent Neural Networks from Overfitting

    Srivastava, Hinton, Krizhevsky, Sutskever, Salakhutdinov · 2014 · JMLR · 40 min · Approachable

    It is the cleanest example of a regulariser with both an intuitive story and an ensemble interpretation, and it is still in most Transformer blocks.

    regularisationtraining
  6. 06

    Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

    Ioffe, Szegedy · 2015 · ICML · 35 min · Moderate

    Normalisation layers are everywhere; this paper shows the mechanics, including the train-time versus test-time statistics gap that still causes bugs.

    normalisationtrainingoptimisation
  7. 07

    Auto-Encoding Variational Bayes

    Kingma, Welling · 2013 · ICLR 2014 · 60 min · Dense

    The evidence lower bound and the reparameterisation trick are the two tools that connect probability to backpropagation; diffusion models are their descendants.

    generativevariational-inferencelatent-variables
  8. 08

    Generative Adversarial Nets

    Goodfellow et al. · 2014 · NeurIPS · 40 min · Moderate

    Adversarial training reframed generation as a game and introduced instabilities that a decade of papers then tried to tame. It is short and worth reading for the proof alone.

    generativegame-theory
  9. 09

    Understanding Deep Learning Requires Rethinking Generalization

    Zhang, Bengio, Hardt, Recht, Vinyals · 2017 · ICLR · 40 min · Moderate

    It set the agenda for generalisation research and is the fastest way to unlearn the textbook bias-variance picture.

    generalisationtheory
  10. 10

    The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks

    Frankle, Carbin · 2019 · ICLR · 35 min · Moderate

    It links initialisation, pruning, and why big models train more easily than small ones, which matters for both theory and deployment.

    pruninginitialisationsparsity
  11. 11

    Deep Double Descent: Where Bigger Models and More Data Hurt

    Nakkiran et al. · 2019 · ICLR 2020 · 40 min · Moderate

    It explains behaviour that contradicts the classical U-shaped curve and informs how to read scaling results honestly.

    generalisationscaling
  12. 12

    Denoising Diffusion Probabilistic Models

    Ho, Jain, Abbeel · 2020 · NeurIPS · 70 min · Dense

    Diffusion is the dominant generative approach for images and audio, and its derivation ties together the VAE bound, score matching, and Langevin dynamics.

    generativediffusionvariational-inference
124 online