Reference
Glossary
A localized term, its English original, and a one-sentence definition. English names remain because documentation, papers, and error messages usually use them.
Reference
Sources
Official documentation, peer-reviewed papers, open benchmarks, standards, and risk catalogs. Everything in the original language. Footnotes in the lessons lead here.
After courses
What to read next
Documentation, tools, standards, and channels through which the area is actually updated. Without news noise and lists of trendy frameworks.
Understand the basics
The Illustrated Transformer: analysis with pictures ↗Attention Is All You Need: original ↗Deep Learning Book: free textbook ↗3Blue1Brown: neural networks visually ↗Platform documentation
Capabilities and model prices ↗Structured output according to the schema ↗Tool call ↗Prompt caching ↗Batch processing ↗Token count ↗How agents and pipelines are built
Building effective agents ↗Introducing Contextual Retrieval ↗Anthropic Cookbook: working examples ↗Model Context Protocol ↗LlamaIndex ↗LangChain ↗Data, search and storage
pgvector: vectors in PostgreSQL ↗Qdrant ↗Faiss ↗Elasticsearch: BM25 and vector search ↗Unstructured: document preparation ↗Airbyte: ready connectors to systems ↗dbt: transformations and data quality ↗Evaluation and Observability
Ragas: metrics RAG ↗Promptfoo: tests of prompts and models ↗Langfuse: tracing and evaluation ↗OpenTelemetry: conventions for AI ↗Risks, standards, regulation
OWASP Top 10 for applications with LLM ↗NIST AI Risk Management Framework ↗EU AI Act: text navigator ↗GDPR: text of the regulations ↗ISO/IEC 42001: AI management system ↗Engineering conferences and standards
MLSys: machine learning and systems ↗OSDI: systems ↗Hot Chips ↗Open Compute Project ↗MLCommons Benchmarks ↗Streams of new works
arXiv cs.CL: language processing ↗arXiv cs.IR: information retrieval ↗arXiv cs.LG: machine learning ↗arXiv cs.DC: distributed systems ↗Axes along which everything changes
Not a list of commercial rates, but technical directions along which practice shifts. Each is related to a lesson.
Contract instead of agreementsThe response schema, tool invocation, and verifiable fields displace the parsing of free text.
Data is more important than the modelThe corpus, its relevance, and search within it define the system's ceiling more strictly than the class of the model.
Evaluation as part of the productThe evaluation set and the judge model turn edits from guessing into measurable work.
Economics through cache and packetsCaching and batch processing change the count multiplicatively, unlike choosing a neighboring model.
Trust boundary for dataEverything the model read is data, not a command. Hence all protection against injections.
Autonomy based on error costThe level of autonomy is chosen not by the beauty of the demonstration, but by the rollback cost.
Memory has become a bottleneckBandwidth and memory capacity grow more slowly than arithmetic, and hence almost all modern engineering.
Rack-scale co-designChip, network, power, and cooling have stopped being designed independently from each other.