A reading list by level: the papers, posts and books worth your time

The field moves fast, but its foundations do not. Most of what practitioners argue about in 2026 traces back to a few dozen papers, posts and books. This list gathers twenty-eight of them and sorts them by the level at which they pay off most. You do not need to read them in order, and you do not need to finish a paper to benefit from it: read the abstract, look at the figures, and come back when a problem at work makes it relevant.

• 3Blue1Brown, “Neural networks” video series. Visual explanations of how a neural network learns and, in later episodes, how a Transformer uses Attention. The best first hour you can spend.
• Anthropic, “Building effective agents” (2024). A plain-language guide to what an AI agent is, and why simple, well-structured workflows often beat elaborate ones.
• Simon Willison, “The lethal trifecta for AI agents” (2025). A short post explaining why an assistant that reads private data, sees untrusted content and can send messages is risky. Clear enough for anyone; important for everyone.
• Chip Huyen, AI Engineering (O'Reilly; released December 2024, dated 2025). The most complete single book on building applications on top of large language models: evaluation, retrieval, agents, cost.

• Jurafsky and Martin, Speech and Language Processing (3rd ed. draft, free online). The standard textbook for language technology; its chapters on embeddings and transformers are a gentle, rigorous foundation.
• Vaswani et al., “Attention Is All You Need” (2017). The paper that introduced the transformer, the architecture behind every modern language model.
• Brown et al., “Language Models are Few-Shot Learners” (2020). The GPT-3 paper; showed that a big enough model can learn a task from a few examples in the prompt, which is in-context learning.
• Wei et al., “Chain-of-Thought Prompting” (2022). Asking a model to show intermediate steps improves its reasoning; the root of today's reasoning models.
• Ouyang et al., “Training language models to follow instructions with human feedback” (2022). The InstructGPT paper: how human feedback turns a text predictor into an assistant.
• Lewis et al., “Retrieval-Augmented Generation” (2020). Introduced RAG: look things up first, then answer, so the model is not limited to what it memorised.

• Robertson and Zaragoza, “The Probabilistic Relevance Framework: BM25 and Beyond” (2009). Why BM25, a keyword scorer from before deep learning, is still in every serious search stack.
• Karpukhin et al., “Dense Passage Retrieval” (2020). Showed that learned dense retrieval can beat keyword search for question answering; the template for embedding-based search.
• Malkov and Yashunin, HNSW (2016). The graph index behind most vector databases; explains why nearest-neighbour search is fast and only approximately exact.
• Cormack, Clarke and Büttcher, “Reciprocal Rank Fusion” (2009). A three-line method for merging ranked lists that still powers most hybrid search.
• Liu et al., “Lost in the Middle” (2023). Models use information at the start and end of a long prompt better than in the middle; a key reason context order matters.
• Yao et al., “ReAct” (2022). Interleaving reasoning with actions: the pattern behind most agent loops.
• Zheng et al., “Judging LLM-as-a-Judge” (2023). When one model can grade another, and its biases; the basis of LLM-as-judge evaluation.
• Model Context Protocol specification. The open standard for connecting models to tools and data; read it to understand what MCP servers actually exchange.

• Kwon et al., “PagedAttention” (2023). The vLLM paper; explains the KV cache memory problem and why serving throughput depends on it.
• Leviathan et al., “Fast Inference via Speculative Decoding” (2022). How a small draft model speeds up a large one without changing its output; see Speculative decoding.
• Hu et al., “LoRA” (2021). Fine-tune a large model by training a tiny set of extra weights; the default way to adapt Open-weights vs closed models models.
• Greshake et al., “Not what you've signed up for” (2023). The paper that defined indirect prompt injection: instructions hidden in content a model reads.
• Beyer et al., Site Reliability Engineering (O'Reilly, 2016; free online). Google's book on running production systems: SLOs, on-call and blameless postmortems. Every AI service is still a service.
• Kleppmann, Designing Data-Intensive Applications (O'Reilly, 2017). The book on Replication, Sharding & partitioning, consistency and streams that every retrieval and agent backend depends on.

• Kaplan et al., “Scaling Laws for Neural Language Models” (2020). Model quality improves predictably with size, data and compute; the starting point for scaling-law reasoning.
• Hoffmann et al., “Training Compute-Optimal Large Language Models” (2022). The Chinchilla paper: for a fixed budget, train smaller models on more data. It changed how labs spend compute.
• Sculley et al., “Hidden Technical Debt in Machine Learning Systems” (2015). The model is a small box in a large system of data, configuration and monitoring; still the best argument for platform investment.
• NIST, AI Risk Management Framework 1.0 (2023). A shared vocabulary for governing AI risk across an organisation; useful when the conversation moves from engineering to policy.