58 articles worth reading, picked by hand. Every one is free and was opened to check it loads. Each pairs with a lesson, so you can read it right after.
Showing 58 of 58
Step-by-step diagrams of self-attention and the encoder-decoder stack, the clearest visual route into how transformers work.
How LLMs workStarterPairs with Transformers and attention
Long, patient explanation of next-token prediction, probabilities and neural nets, starting from a blank page.
How LLMs workCorePairs with Next-token prediction and sampling
Plain-English tour of word vectors, transformer layers and training that assumes no maths background.
How LLMs workStarterPairs with Pretraining, fine-tuning and RL
Survey of attention variants with formulas, tracing how the mechanism grew from translation models to transformers.
How LLMs workStretchPairs with Attention in numbers
Character-level language models generating Shakespeare and code, a gentle preview of why predicting the next token goes far.
How LLMs workCorePairs with Machine learning and the deep learning turn
Explains the main ways reasoning models are built and trained, including inference-time scaling and reinforcement learning.
How LLMs workStretchPairs with Reasoning and thinking models
Walks through how BPE builds a vocabulary and tokenizes text, with runnable code.
How LLMs workCorePairs with Tokens and the context window
Explains the three RLHF stages that turn a raw language model into a chat assistant.
How LLMs workCorePairs with Reinforcement learning for reasoning
Compares greedy, beam, top-k and top-p sampling with examples, showing what temperature-style settings really change.
How LLMs workCorePairs with Sampling parameters
Defines hallucination, covers why it happens and the detection and mitigation methods researchers use.
How LLMs workStretchPairs with Limits and hallucinations
Official starting point for writing prompts, with pointers to techniques like examples, structure and reasoning.
PromptingStarterPairs with Prompting basics
OpenAI's guide to message roles, few-shot examples and prompt structure for reliable outputs.
PromptingStarterPairs with System prompts, roles and constraints
Survey of few-shot, chain-of-thought and other prompting methods with the research behind each.
PromptingCorePairs with Zero-shot, few-shot and chain-of-thought
Practical rules for prompts that work in production: clear instructions, examples, structured output and careful iteration.
PromptingCorePairs with Zero-shot, few-shot and chain-of-thought
Frames context as a finite budget and covers compaction, notes and sub-agents for keeping it useful.
PromptingCorePairs with Compaction, isolation and long context
Names four ways long contexts go wrong (poisoning, distraction, confusion, clash) and how to avoid each.
PromptingCorePairs with How context fails
Short argument for why the term context engineering describes the real work better than prompt engineering.
PromptingStarterPairs with Context engineering
The standard reference on workflows versus agents, with five composable patterns and advice to start simple.
AgentsStarterPairs with Workflow patterns
Breaks an agent into planning, memory and tool use, with the main techniques for each.
AgentsCorePairs with Memory and state
How to design, test and refine tools so an agent can use them well, with concrete principles.
AgentsCorePairs with Designing tools for agents
Engineering account of an orchestrator-and-subagents design, including what worked, costs and failures.
AgentsCorePairs with Multi-agent systems
How to keep an agent productive across many context windows using progress files, feature lists and clean handoffs.
AgentsStretchPairs with Long-running agents
Settles on agent meaning an LLM running tools in a loop to achieve a goal, with examples of other definitions.
AgentsStarterPairs with Workflows vs agents
Argues that parallel agents lose shared context and make conflicting decisions, and why a single thread is often better.
AgentsStretchPairs with Multi-agent systems
Official introduction to the protocol that connects AI apps to tools and data sources.
AgentsStarterPairs with Model Context Protocol
Deep guide to agent tools, planning, reflection and failure modes from a working ML engineer.
AgentsStretchPairs with The agent loop
Shows how adding chunk context before embedding, plus hybrid search and reranking, cuts retrieval failures.
RetrievalCorePairs with Advanced RAG techniques
Accessible talk-turned-article on embeddings with practical demos like semantic search and clustering.
RetrievalStarterPairs with What are embeddings?
Free book-length overview of embeddings from early word vectors to modern models, for engineers.
RetrievalStretchPairs with Choosing an embedding model
Compares fixed-size, sentence, recursive and semantic chunking, with guidance on picking one.
RetrievalCorePairs with Chunking documents
Explains why a fast first retrieval plus a slower reranker beats vector search alone, with code.
RetrievalCorePairs with Reranking search results
The practical case for evals: unit tests, human review and a loop that lets you improve with confidence.
EvalsStarterPairs with Why evals come first
Step-by-step method for building an LLM judge you can trust, anchored on human labels.
EvalsCorePairs with LLM as judge
How to design graders, tasks and metrics for evaluating agents that take many steps.
EvalsCorePairs with Trajectory evals for agents
Shows how to look at your data and traces, find failure patterns and fix them before building tooling.
EvalsCorePairs with Error analysis
A simple test for when an agent can be tricked into leaking your data, with real incidents.
SafetyStarterPairs with MCP: capabilities, resources vs tools, and trust
How a lab approaches defending a browsing agent against injected instructions, with measured results.
SafetyCorePairs with Prompt injection and the lethal trifecta
Summarizes six architectural patterns that limit what an injected prompt can do.
SafetyStretchPairs with Prompt injection and the lethal trifecta
Community-maintained list of the main security risks in LLM apps, with descriptions and mitigations.
SafetyCorePairs with Misuse, privacy and moderation
Investigation of a risk-scoring algorithm used in courts, a concrete case of bias in automated decisions.
SafetyStarterPairs with Bias, fairness and honesty
Tactical, operational and strategic lessons from practitioners who shipped LLM products.
BuildingCorePairs with 12-factor agents
Essay defining the AI engineer role as distinct from ML researcher, and why it is growing.
BuildingStarterPairs with What is an AI engineer?
Seven patterns (evals, RAG, fine-tuning, caching, guardrails, defensive UX, user feedback) with trade-offs.
BuildingCorePairs with Deploying an AI app
Walks through the components of a production AI platform, including guardrails, caching, routing and observability.
BuildingStretchPairs with Tracing and observability
Official guide to getting good results from a coding agent: context, verification, planning and parallel sessions.
Coding with AIStarterPairs with The agentic coding workflow
A veteran developer's concrete habits for coding with LLMs, including limits and tricks.
Coding with AIStarterPairs with Chat apps and coding agents
A spec-first workflow: brainstorm, write a plan, then let the model implement step by step.
Coding with AICorePairs with Spec-driven development
Explains why AI gets you most of the way fast but the last stretch still needs engineering skill.
Coding with AICorePairs with The agentic coding workflow
Builds a working coding agent in a few hundred lines, showing it is a model, a loop and tools.
Coding with AICorePairs with Tool use and function calling
Short essay arguing that general methods that scale with compute beat hand-built knowledge.
HistoryStarterPairs with Why bigger worked: scaling laws
Explains the Chinchilla scaling result and what it implied about data being the bottleneck.
HistoryStretchPairs with Why bigger worked: scaling laws
Builds entropy, cross-entropy and KL divergence from pictures, the maths behind training losses.
HistoryCorePairs with Entropy and KL divergence
Lab write-up of a system that mastered Go by self-play with no human games, a landmark in reinforcement learning.
HistoryStarterPairs with Games and reinforcement learning
Argues that with general recipes in place, progress shifts from building methods to defining better evaluations.
HistoryStretchPairs with What is an AI engineer?
Essay arguing AI will diffuse slowly like past technologies rather than bring sudden transformation.
Critical viewsStretchPairs with A field guide to AI buzzwords
A prominent sceptic's case that scaling alone will not deliver reliable reasoning; useful to read against the Bitter Lesson.
Critical viewsCorePairs with Why bigger worked: scaling laws
Randomized trial finding experienced developers were slower with AI tools than they expected, a counterweight to hype.
Critical viewsCorePairs with Chat apps and coding agents
Research summary showing reasoning models degrade on harder puzzles, a sceptical look at claimed reasoning.
Critical viewsStretchPairs with Reasoning and thinking models