4.0 KiB
Glossary — artificial intelligence engineer
Definitions are written in our own words; each entry links its source for verification. (Paraphrased from Wikipedia — no text copied.)
Machine learning (ML)
Algorithms that improve at a task through data instead of explicit programming: a model is trained on examples, validated on held-out data and then applied to unseen inputs. Most production AI work is applied ML, not research. Source: https://en.wikipedia.org/wiki/Machine_learning
Deep learning
Machine learning with multi-layer neural networks that learn hierarchical representations directly from raw data (pixels, audio, text). It powers modern vision, speech and language systems, at the price of large data and compute needs. Source: https://en.wikipedia.org/wiki/Deep_learning
Large language model (LLM)
A neural network — typically transformer-based — trained on very large text corpora to predict tokens, which makes it usable for generation, reasoning, extraction and dialogue. Engineers rarely train them from scratch; they select, prompt, fine-tune and orchestrate them. Source: https://en.wikipedia.org/wiki/Large_language_model
Transformer
The dominant neural architecture for sequence data. Its attention mechanism lets every token weigh every other token, which parallelizes training and scales far better than recurrent networks. Basis of GPT-, BERT- and Llama-style models. Source: https://en.wikipedia.org/wiki/Transformer_(deep_learning_architecture)
Fine-tuning
Continuing the training of a pre-trained model on a smaller task- or domain-specific dataset. Cheaper than training from scratch; the standard lever when prompting alone cannot reach the required quality or style. Source: https://en.wikipedia.org/wiki/Fine-tuning_(deep_learning)
Retrieval-augmented generation (RAG)
An architecture that answers with an LLM but grounds it first: relevant documents are retrieved (usually via embeddings from a vector store) and injected into the prompt. The default pattern when answers must reflect private or current knowledge without retraining. Source: https://en.wikipedia.org/wiki/Retrieval-augmented_generation
Vector database
A datastore optimized for similarity search over embeddings — high-dimensional vectors representing text, images or other content. The retrieval backbone of RAG systems and semantic search. Source: https://en.wikipedia.org/wiki/Vector_database
MLOps
The discipline of running ML in production: versioning data and models, automated training and deployment pipelines, monitoring for drift and regressions. It transfers DevOps thinking to systems whose behaviour depends on data, not only code. Source: https://en.wikipedia.org/wiki/MLOps
Model drift
The degradation of model quality over time because production data no longer matches the training distribution (data drift) or because the relationship between input and target changed (concept drift). The reason deployed models need monitoring, not just launch metrics. Source: https://en.wikipedia.org/wiki/Concept_drift
Prompt engineering
Designing the instructions, context and examples given to a generative model so its output is reliable for a task — including system prompts, few-shot examples and structured output constraints. Often the cheapest quality lever, always the first one to try. Source: https://en.wikipedia.org/wiki/Prompt_engineering
Model evaluation
Measuring model quality against defined metrics on data the model has not seen — offline (test sets, benchmarks, human review) and online (A/B tests, guardrail monitors). In generative systems, evaluation design is frequently harder than model selection. Source: https://en.wikipedia.org/wiki/Evaluation_of_machine_translation
Responsible AI / AI safety
Practices that keep AI systems fair, transparent, privacy-preserving and controllable: bias testing, explainability, human oversight, incident processes. In the EU increasingly a legal duty (AI Act), not just an ethic. Source: https://en.wikipedia.org/wiki/Regulation_of_artificial_intelligence