Resources

How to prepare

The harder topics like RLHF and reasoning models sit on top of a lot of fundamentals. Post-training only makes sense once you know what pretraining does, pretraining makes more sense once you know why Transformers replaced RNNs, and so on back to n-grams. So start at the beginning of the reading order and work forward.

Andrej Karpathy has talked about the 10,000-hour rule, the idea that you can get good at almost anything by putting in enough focused hours. We think it applies here too. Most of the understanding comes from repetition, going back over the same ideas and getting questions wrong until they stick.

That's why we built the question bank. You do a few questions a day, and the ones you miss come back.

We have curated a list of completely free resources to help with your prep!

Courses

Videos

  • Neural Networks: Zero to HeroAndrej Karpathy

    You write backprop, a tokenizer and a GPT yourself, which makes most fundamentals questions easy.

  • Neural networks3Blue1Brown

    The quickest way to a picture of backprop and attention you can redraw at a whiteboard.

  • GPU MODE lecturesGPU MODE

    Practitioners teaching CUDA, Triton and kernel writing.

Books

Blogs and guides

Repos

  • LLMs from scratch (code)Sebastian Raschka

    Runnable notebooks for each piece of an LLM, from tokenizer to finetuning.

  • nanochatAndrej Karpathy

    A full ChatGPT-style pipeline small enough to read end to end.

Reading order

Post-training, RLHF and reasoning models build on everything that came before them, and each step in that history fixes a problem with the one before. N-gram models count word sequences but can't share anything between similar words, so neural language models learn word vectors. A feedforward model only sees a fixed window, so recurrent networks carry a state through the whole sequence, and LSTMs keep that state from fading. Squeezing a sentence into one vector loses detail, so attention lets the decoder look back at every word, and the Transformer keeps the attention and drops the recurrence.

Pretraining those Transformers at scale gives GPT, BERT and T5, and only then do post-training and RL make sense, because they adjust what pretraining already put in the model. Read in order, and skip a stage only if you could explain it to someone else.

Counting words

  1. Naive Bayes and Text ClassificationJurafsky & Martin, appendix B, 2026

    Documents as bags of words, the simplest representation and the baseline everything else beats.

  2. N-gram Language ModelsJurafsky & Martin, chapter 3, 2026

    The Markov assumption, smoothing and perplexity, which is still how language models are scored.

  3. Hidden Markov ModelsJurafsky & Martin, appendix A, 2026

    Hidden states behind the words, and the Viterbi and forward algorithms for them.

  4. EmbeddingsJurafsky & Martin, chapter 5, 2026

    Term-document and co-occurrence matrices, tf-idf, then word2vec.

Neural language models

  1. Neural NetworksJurafsky & Martin, chapter 6, 2026

    Feedforward networks, backprop and a feedforward language model over a fixed window.

  2. A Neural Probabilistic Language ModelBengio et al., 2003

    Learned word vectors let similar words share statistics, which n-grams can't.

  3. Efficient Estimation of Word Representations in Vector SpaceMikolov et al., 2013

    word2vec: skip-gram and CBOW, cheap enough to train on billions of words.

  4. GloVe: Global Vectors for Word RepresentationPennington et al., 2014

    Word vectors fitted to global co-occurrence counts, linking the matrix view to word2vec.

Recurrent networks

  1. A Primer on Neural Network Models for NLPGoldberg, 2015

    Feedforward and recurrent models in one notation, with RNNs as acceptors, encoders and transducers.

  2. RNNs and LSTMsJurafsky & Martin, chapter 14, 2026

    Backprop through time, stacked and bidirectional RNNs, LSTMs and the encoder-decoder.

  3. Understanding LSTM NetworksChristopher Olah, 2015

    Each LSTM gate drawn out, and why gating keeps gradients from vanishing.

  4. Speech Recognition with Deep Recurrent Neural NetworksGraves et al., 2013

    Stacked bidirectional LSTMs, and what depth buys a recurrent model.

Sequence to sequence and attention

  1. Learning Phrase Representations using RNN Encoder-DecoderCho et al., 2014

    The encoder-decoder for translation, and the GRU.

  2. Sequence to Sequence Learning with Neural NetworksSutskever et al., 2014

    Deep LSTMs that map a sentence to one vector and back.

  3. Neural Machine Translation by Jointly Learning to Align and TranslateBahdanau et al., 2014

    Attention: the decoder looks back at every source word instead of one fixed vector.

Pretrained Transformers

  1. Deep Contextualized Word RepresentationsPeters et al., 2018

    ELMo: a word's vector depends on its sentence, taken from a bidirectional LSTM language model.

  2. Attention Is All You NeedVaswani et al., 2017

    Recurrence dropped for self-attention, so a whole sequence trains in parallel.

  3. Improving Language Understanding by Generative Pre-TrainingRadford et al., 2018

    GPT: a decoder-only Transformer pretrained left to right, then finetuned per task.

  4. BERTDevlin et al., 2018

    An encoder pretrained with masked language modelling, reading both directions at once.

  5. Exploring the Limits of Transfer Learning (T5)Raffel et al., 2019

    Every task as text to text with an encoder-decoder, and a careful comparison of objectives.

  6. Language Models are Unsupervised Multitask LearnersRadford et al., 2019

    GPT-2: a bigger decoder doing tasks zero-shot, with no finetuning.

Scaling

  1. Language Models are Few-Shot LearnersBrown et al., 2020

    GPT-3, and in-context learning appearing with scale.

  2. Scaling Laws for Neural Language ModelsKaplan et al., 2020

    Loss falls as a power law in parameters, data and compute.

  3. Training Compute-Optimal Large Language ModelsHoffmann et al., 2022

    Chinchilla: scale data with parameters, about 20 tokens each.

Modern LLM architecture

  1. RoFormer: Rotary Position EmbeddingSu et al., 2021

    RoPE, the default way to encode position.

  2. FlashAttentionDao et al., 2022

    Exact attention made fast by tiling around GPU memory.

  3. GQA: Grouped-Query AttentionAinslie et al., 2023

    Shared key and value heads for a smaller KV cache.

  4. Tensor Programs V: µPYang et al., 2022

    Tune hyperparameters on a small model and transfer them.

  5. Switch TransformersFedus et al., 2021

    Mixture of experts with one expert per token.

  6. LLaMATouvron et al., 2023

    A compact recipe for a strong open model.

  7. The Llama 3 Herd of ModelsGrattafiori et al., 2024

    A full frontier pipeline written down, data to post-training.

  8. DeepSeek-V3 Technical ReportDeepSeek-AI, 2024

    Latent attention, fine-grained experts and FP8 training.

Data and midtraining

  1. The FineWeb DatasetsPenedo et al., 2024

    Filtering and deduplicating web text, with every ablation.

  2. Efficient Training of Language Models to Fill in the MiddleBavarian et al., 2022

    Infilling for code at almost no cost to left-to-right quality.

Post-training

  1. High-Dimensional Continuous Control Using GAESchulman et al., 2015

    Generalized advantage estimation and the λ trade-off.

  2. Proximal Policy Optimization AlgorithmsSchulman et al., 2017

    The clipped objective behind classic RLHF.

  3. Training Language Models to Follow Instructions with Human FeedbackOuyang et al., 2022

    InstructGPT: SFT, a reward model, then PPO.

  4. Constitutional AIBai et al., 2022

    AI feedback in place of human harmlessness labels.

  5. Direct Preference OptimizationRafailov et al., 2023

    Preference tuning with no reward model and no RL loop.

  6. LoRAHu et al., 2021

    Finetuning through low-rank weight updates.

Reasoning and RL

  1. Chain-of-Thought PromptingWei et al., 2022

    Writing out steps before the answer.

  2. Let's Verify Step by StepLightman et al., 2023

    Process rewards against outcome rewards on maths.

  3. DeepSeekMathShao et al., 2024

    Introduces GRPO, PPO with a group baseline instead of a critic.

  4. DeepSeek-R1DeepSeek-AI, 2025

    Long reasoning from RL on verifiable rewards.

  5. DAPOYu et al., 2025

    What breaks in GRPO at scale, and the fixes.

Inference

  1. PagedAttentionKwon et al., 2023

    The KV cache managed like virtual memory, from vLLM.

  2. Fast Inference via Speculative DecodingLeviathan et al., 2022

    A small model drafts, the big one verifies in one pass.

  3. GPTQFrantar et al., 2022

    4-bit weights with second-order error correction.

  4. AWQLin et al., 2023

    Protect the weights that see large activations.