Sreeraj Sreenivasan

Posted on Aug 10

300 Terms Every Beginner AI Engineer Should Know

#ai #beginners #architecture #machinelearning

AI engineering has its own language. Half the battle when you're starting out isn't the math or the code — it's the vocabulary. This is a beginner-friendly glossary of 300 terms, organized into 12 categories of 25 terms each, so you can look things up fast or just read through and level up your fluency.

No fluff, just definitions you can actually use.

1. Foundations & Math

Artificial Intelligence (AI) — Any system that performs tasks normally requiring human intelligence, like reasoning or recognizing patterns.
Machine Learning (ML) — A subset of AI where systems learn patterns from data instead of following hardcoded rules.
Deep Learning — A subset of ML using multi-layered neural networks to learn complex patterns.
Supervised Learning — Training a model on labeled data (input paired with the correct output).
Unsupervised Learning — Training a model to find patterns in data that has no labels.
Reinforcement Learning — Training a model by rewarding good actions and penalizing bad ones.
Semi-Supervised Learning — Training on a mix of a small amount of labeled data and a large amount of unlabeled data.
Self-Supervised Learning — A model generates its own labels from raw data, used heavily in pretraining LLMs.
Algorithm — A step-by-step set of rules a model follows to solve a problem or learn from data.
Model — The output of a training process — a file that has learned to make predictions from data.
Dataset — A structured collection of data used to train, validate, or test a model.
Feature — An individual measurable input variable used by a model (e.g., age, pixel value, word).
Vector Space — A mathematical space where vectors can be added and scaled; the foundation for embeddings.
Matrix Multiplication — The core mathematical operation powering neural network computations.
Probability Distribution — A function describing the likelihoods of different outcomes.
Mean (Average) — The central value of a dataset.
Standard Deviation — A measure of how spread out data values are.
Normalization — Scaling data to a standard range to improve training stability.
Softmax Function — Converts raw scores into probabilities that sum to 1.
Derivative/Gradient — Measures how a function's output changes as its input changes; core to training.
Eigenvalue/Eigenvector — Mathematical concepts used in dimensionality reduction techniques like PCA.
Dimensionality Reduction — Techniques (like PCA) that reduce the number of features while preserving information.
Bayes' Theorem — A formula for updating the probability of an event based on new evidence.
Correlation vs Causation — The distinction between two variables moving together and one actually causing the other.
Outlier — A data point that differs significantly from the rest of the dataset.

2. Neural Networks & Architectures

Neural Network — A model loosely inspired by the brain, made of layers of connected "neurons."
Neuron (Node) — A single computational unit in a neural network that processes input and passes output forward.
Layer — A group of neurons that process data at the same stage of a network.
Weights — Numeric values that determine the strength of connections between neurons; these are what training adjusts.
Bias (in networks) — An extra adjustable value added to a neuron's output to help the model fit data better.
Activation Function — A function (like ReLU or sigmoid) that decides whether/how a neuron's signal passes forward.
Backpropagation — The algorithm used to calculate how to adjust weights by working backward from the error.
Convolutional Neural Network (CNN) — A network architecture specialized for images and spatial data.
Recurrent Neural Network (RNN) — An older architecture designed for sequences, like text or time series.
Long Short-Term Memory (LSTM) — An RNN variant designed to better retain information over longer sequences.
Parameters — The total count of weights and biases in a model; often used as a rough measure of model size.
Mixture of Experts (MoE) — An architecture that routes inputs to specialized sub-networks ("experts") instead of using the whole model.
Encoder-Decoder — An architecture where one part compresses input and another generates output, common in translation.
Multimodal Model — A model that processes multiple types of input, like text, images, and audio together.
Foundation Model — A large, general-purpose model trained on broad data, adaptable to many downstream tasks.
Small Language Model (SLM) — A compact language model designed for efficiency over raw scale.
Sparse Model — A model where only a subset of parameters activate for a given input, improving efficiency.
Dense Model — A model where all parameters are used for every input, unlike sparse/MoE models.
LoRA (Low-Rank Adaptation) — An efficient fine-tuning method that trains small added matrices instead of the full model.
Distillation — Training a smaller "student" model to mimic a larger "teacher" model's behavior.
Knowledge Graph — A structured network of entities and their relationships, used to ground AI reasoning in facts.
Residual Connection (ResNet) — A shortcut that lets a layer's input skip ahead, helping train very deep networks.
Layer Normalization — A technique that stabilizes training by normalizing values within each layer.
Batch Normalization — A technique that stabilizes training by normalizing values across a mini-batch.
Positional Encoding — Information added to input tokens so a transformer knows their order in a sequence.

3. NLP Fundamentals

Natural Language Processing (NLP) — The field focused on getting computers to understand and generate human language.
Tokenization — Breaking text into smaller units (tokens) — words, subwords, or characters — for a model to process.
Token — A single unit of text after tokenization; roughly ¾ of a word in English on average.
Corpus — A large collection of text used to train or evaluate an NLP model.
Stemming — Reducing words to their root form by chopping off endings (e.g., "running" → "run").
Lemmatization — Reducing words to their dictionary base form using grammar rules, more accurate than stemming.
Named Entity Recognition (NER) — Identifying names of people, places, organizations, etc. in text.
Sentiment Analysis — Determining whether text expresses a positive, negative, or neutral tone.
Stop Words — Common words (like "the," "is," "and") often filtered out before processing text.
N-gram — A sequence of N consecutive words or tokens, used to capture local context in text.
Word2Vec — An early technique for turning words into vectors based on the words around them.
GloVe — A word embedding technique that learns vectors from global word co-occurrence statistics.
Bag of Words — A simple text representation that counts word occurrences, ignoring order.
TF-IDF — A scoring method that weighs how important a word is to a document relative to a whole corpus.
Byte Pair Encoding (BPE) — A subword tokenization method that merges frequent character pairs, used by many LLMs.
WordPiece — A subword tokenization method similar to BPE, used in models like BERT.
Part-of-Speech (POS) Tagging — Labeling each word in a sentence with its grammatical role (noun, verb, etc.).
Dependency Parsing — Analyzing the grammatical structure of a sentence to show how words relate to each other.
Coreference Resolution — Figuring out which words refer to the same entity (e.g., matching "she" to a name).
Text Summarization — Automatically condensing a longer text into a shorter version that keeps the key meaning.
Machine Translation — Automatically translating text from one language to another.
Word Sense Disambiguation — Determining which meaning of a word applies based on context.
Language Detection — Automatically identifying which language a piece of text is written in.
Text Classification — Assigning a category label to a piece of text.
Language Model — A model trained to predict the next word (or token) in a sequence, the basis of modern NLP.

4. LLMs & Transformers

Large Language Model (LLM) — A massive neural network trained on huge amounts of text to predict and generate language.
Transformer — The neural network architecture behind modern LLMs, built around the attention mechanism.
Attention Mechanism — A technique letting a model weigh how relevant each word is to every other word in a sequence.
Self-Attention — Attention applied within a single sequence, letting each token "look at" every other token.
Multi-Head Attention — Running several attention mechanisms in parallel so a model can capture different types of relationships at once.
Context Window — The maximum amount of text (in tokens) a model can consider at once.
Pretraining — The initial phase where a model learns general language patterns from massive raw text.
Fine-Tuning — Further training a pretrained model on a smaller, specific dataset to specialize its behavior.
Instruction Tuning — Fine-tuning a model specifically to follow instructions and prompts well.
RLHF (Reinforcement Learning from Human Feedback) — A technique using human preferences to guide a model toward better responses.
Hallucination — When a model confidently generates false or made-up information.
Causal Language Model — A model that predicts the next token using only previous tokens, never future ones.
Masked Language Model — A model trained to predict missing (masked) words using surrounding context in both directions.
Decoder-Only Model — A transformer architecture using only the decoder stack; the basis for most modern LLMs like GPT.
Encoder-Only Model — A transformer architecture using only the encoder stack, good for understanding tasks like classification.
Base Model — A pretrained model before any fine-tuning or instruction tuning is applied.
Chat Model — A model fine-tuned specifically for multi-turn conversational interaction.
In-Context Learning — A model's ability to learn a task from examples given in the prompt, without updating its weights.
Catastrophic Forgetting — When fine-tuning on new data causes a model to lose previously learned knowledge.
Emergent Abilities — New capabilities that appear in large models once they cross a certain scale, without explicit training.
Model Weights — The learned numeric values inside a model, saved to a file after training.
Sliding Window Attention — An attention technique that limits focus to a nearby range of tokens to reduce compute cost.
Alignment Tax — The tradeoff where making a model safer or more aligned can slightly reduce raw capability.
Context Length — The actual number of tokens used in a specific request, up to the context window's limit.
Token Limit — The maximum number of tokens a model or API call allows for input and/or output combined.

5. Training & Optimization

Loss Function — A formula that measures how wrong a model's predictions are; training tries to minimize this.
Gradient Descent — An optimization method that nudges weights in the direction that reduces loss.
Learning Rate — A setting that controls how big each adjustment step is during training.
Epoch — One complete pass through the entire training dataset.
Batch Size — The number of training examples processed together before the model updates its weights.
Overfitting — When a model memorizes training data too closely and performs poorly on new data.
Underfitting — When a model is too simple to capture patterns in the data, performing poorly everywhere.
Regularization — Techniques (like dropout or weight decay) used to prevent overfitting.
Dropout — A regularization technique that randomly "turns off" neurons during training to improve generalization.
Transfer Learning — Reusing a model trained on one task as the starting point for a related task.
Hyperparameter — A setting chosen before training (like learning rate or batch size) rather than learned from data.
Hyperparameter Tuning — The process of searching for the best combination of hyperparameters.
Early Stopping — Halting training once performance on validation data stops improving, to avoid overfitting.
Weight Decay — A regularization technique that shrinks weights slightly during training to prevent overfitting.
Momentum — An optimization technique that smooths out gradient updates using a running average of past steps.
Agent (RL) — The decision-maker that interacts with an environment in reinforcement learning.
Environment (RL) — The world an RL agent interacts with and receives feedback from.
Reward Function — Defines what an RL agent is trying to maximize.
Policy — An RL agent's strategy for choosing actions given a state.
State — A snapshot of the environment at a given point in time.
Q-Learning — A foundational RL algorithm that learns the value of actions in given states.
Exploration vs Exploitation — The tradeoff between trying new actions and using known good ones.
Markov Decision Process (MDP) — The mathematical framework underlying most RL problems.
Reward Hacking — When an agent finds unintended shortcuts to maximize reward without achieving the real goal.
Policy Gradient — A class of RL algorithms that directly optimize the agent's policy.

6. Prompting & Interaction

Prompt — The input text you give a model to get a response.
Prompt Engineering — The practice of crafting inputs to get better, more reliable outputs from a model.
Zero-Shot Learning — Asking a model to perform a task with no examples given in the prompt.
Few-Shot Learning — Giving a model a handful of examples in the prompt to guide its response.
Chain-of-Thought (CoT) — Prompting a model to reason step-by-step before giving a final answer.
System Prompt — Instructions given to a model before user input, setting its behavior or persona.
Temperature — A setting controlling randomness in a model's output; low = predictable, high = creative.
Top-p (Nucleus Sampling) — A sampling method that picks from the smallest set of likely next tokens.
Context Engineering — Deliberately structuring the information fed to a model (not just the instruction) to improve results.
Prompt Injection — An attack where malicious input tricks a model into ignoring its original instructions.
Meta-Prompting — Using a prompt to help generate or refine other prompts.
Prompt Chaining — Breaking a task into multiple prompts, feeding one output into the next prompt.
Self-Consistency — Generating multiple reasoning paths and picking the most common final answer.
Tree of Thought — A prompting technique where a model explores multiple reasoning branches before choosing the best one.
ReAct (Reasoning + Acting) — A prompting pattern where a model alternates between reasoning steps and taking actions.
Negative Prompting — Explicitly telling a model what to avoid in its output.
Prompt Template — A reusable prompt structure with placeholders filled in for specific use cases.
Role Prompting — Asking a model to respond as if it were a specific persona or expert.
Multi-Turn Conversation — An interaction spanning several back-and-forth exchanges with a model.
Conversation History — The prior messages in a conversation, often included in the prompt for context.
Output Parsing — Extracting structured information from a model's raw text response.
Structured Output — A model response formatted to a predictable schema, like JSON, rather than free text.
JSON Mode — A model setting that constrains output to valid JSON format.
Instruction Following — A model's ability to accurately do what a prompt explicitly asks.
Top-k Sampling — A sampling method that restricts token choices to the k most likely options.

7. Data, Embeddings & Retrieval

Embedding — A numeric vector representation of data (text, images, etc.) that captures meaning or similarity.
Vector — A list of numbers representing a point in multi-dimensional space, used to represent data mathematically.
Vector Database — A database optimized for storing and searching embeddings by similarity.
Cosine Similarity — A common way to measure how similar two vectors are, based on the angle between them.
Semantic Search — Searching by meaning rather than exact keyword matches, powered by embeddings.
RAG (Retrieval-Augmented Generation) — A technique that retrieves relevant external data and feeds it to an LLM before generating a response.
Chunking — Splitting long documents into smaller pieces for embedding and retrieval.
Data Augmentation — Creating new training examples by modifying existing data (rotating images, paraphrasing text, etc.).
Data Labeling — The process of tagging raw data with the correct answers for supervised learning.
Data Drift — When the statistical properties of incoming data change over time, degrading model performance.
Structured Data — Data organized in a predictable format, like rows and columns in a database.
Unstructured Data — Data without a predefined format, like free text, images, or audio.
Indexing — Organizing data so it can be searched and retrieved quickly.
SQL — A language for querying and managing structured, relational data.
NoSQL — Database types (document, key-value, graph) designed for flexible, non-tabular data.
Hybrid Search — Combining keyword search and semantic (vector) search for better retrieval results.
Retriever — The component in a RAG system that fetches relevant documents before generation.
Reranking — Reordering search results using a more precise model after an initial fast retrieval step.
ETL (Extract, Transform, Load) — The process of moving and reshaping data from source systems into a usable format.
Data Pipeline — An automated sequence of steps that moves and processes data from source to destination.
Data Cleaning — The process of fixing or removing incorrect, incomplete, or duplicate data.
Data Versioning — Tracking changes to datasets over time, similar to code version control.
Feature Store — A centralized system for storing and serving features consistently across training and production.
Approximate Nearest Neighbor (ANN) — A fast search technique that finds "close enough" similar vectors instead of exact matches.
Sharding — Splitting a large dataset or database across multiple machines for scalability.

8. Agents & Tooling

AI Agent — A system that uses an LLM to plan, decide, and take actions (often using tools) to accomplish a goal.
Agentic Workflow — A process where an AI plans multiple steps and executes them with minimal human input.
Tool Calling (Function Calling) — A model's ability to invoke external functions or APIs to get information or perform actions.
Model Context Protocol (MCP) — A standard that lets AI models connect to external tools and data sources in a consistent way.
Orchestration — Coordinating multiple models, tools, or agents to complete a complex task.
Multi-Agent System — A setup where multiple AI agents collaborate (or compete) to solve a problem.
Autonomous Agent — An agent that can operate and make decisions with little or no human oversight.
Memory (in agents) — A mechanism that lets an agent retain information across steps or sessions.
Planning — The step where an agent breaks a goal into smaller, executable sub-tasks.
Copilot — An AI assistant embedded in a workflow (like an IDE) that suggests or completes actions alongside a human.
Hugging Face — A popular platform and library ecosystem for sharing and using ML models.
Open-Weight Model — A model whose trained parameters are publicly released for anyone to use or modify.
PyTorch — A widely used open-source deep learning framework.
TensorFlow — A widely used open-source deep learning framework developed by Google.
LangChain — A framework for building applications that chain together LLM calls, tools, and data sources.
Ollama — A tool for running open-source LLMs locally on your own machine.
Jupyter Notebook — An interactive coding environment popular for data science and ML experimentation.
Model Hub — A repository where pretrained models can be discovered, downloaded, and shared.
GGUF — A file format optimized for running large language models efficiently on local hardware.
API Key — A unique credential used to authenticate requests to an AI service or model provider.
Workflow Automation — Using software (often AI-driven) to run multi-step processes without manual intervention.
Toolchain — The set of tools and libraries used together to build, train, and deploy AI systems.
Plugin — An add-on component that extends an AI system's capabilities, like giving it access to a new tool.
Task Queue — A system that manages and processes jobs (like agent actions) in order, often asynchronously.
Webhook — An automated message sent from one system to another when a specific event happens, often used to trigger agent actions.

9. Evaluation, Metrics & Applied Systems

Accuracy — The percentage of predictions a model got exactly right.
Precision — Of all the positive predictions a model made, how many were actually correct.
Recall — Of all the actual positives, how many the model correctly identified.
F1 Score — A single metric that balances precision and recall.
Confusion Matrix — A table showing correct vs incorrect predictions across categories.
Perplexity — A metric measuring how well a language model predicts text; lower is better.
BLEU Score — A metric comparing machine-generated text to a reference translation or summary.
Benchmark — A standardized test/dataset used to compare model performance.
Ground Truth — The verified, correct answer used to evaluate model predictions.
A/B Testing — Comparing two versions of a model or system by exposing each to real users and measuring results.
Recommendation System — A system that predicts what a user might like based on data.
Collaborative Filtering — Recommending items based on similar users' behavior.
Content-Based Filtering — Recommending items similar to what a user already liked.
Cold Start Problem — The difficulty of making good recommendations for new users or items with no history.
Click-Through Rate (CTR) — The percentage of users who click a recommended item.
Matrix Factorization — A technique that decomposes user-item interaction data to find hidden patterns.
Ranking Model — A model that orders items by predicted relevance to a user.
Personalization — Tailoring content or recommendations to an individual user.
Anomaly Detection — Identifying unusual data points that deviate from expected patterns.
Time Series — Data points collected or recorded at successive time intervals.
Forecasting — Predicting future values based on historical time series data.
Seasonality — Recurring patterns in data tied to a fixed time period (daily, yearly, etc.).
Trend — The long-term direction of a time series, independent of short-term fluctuations.
Moving Average — A technique that smooths data by averaging over a sliding window.
Stationarity — A property of time series data where statistical characteristics don't change over time.

10. Deployment, MLOps & Infrastructure

Inference — The process of running a trained model on new input to get a prediction or output.
Latency — The time it takes for a model to respond after receiving input.
Throughput — The number of requests or tokens a system can process in a given time.
API (Application Programming Interface) — A defined way for software (including AI models) to communicate with other software.
MLOps — Practices for deploying, monitoring, and maintaining ML models in production reliably.
Model Serving — Making a trained model available to handle live requests, usually via an API.
Quantization — Reducing the precision of a model's numbers (e.g., 32-bit to 8-bit) to make it smaller and faster.
GPU (Graphics Processing Unit) — Specialized hardware that accelerates the parallel math AI training and inference need.
Model Versioning — Tracking and managing different versions of a trained model over time.
Model Drift — When a deployed model's performance degrades over time as real-world data shifts away from training data.
TPU (Tensor Processing Unit) — Google's custom chip optimized for ML workloads.
CPU — A general-purpose processor, slower than a GPU/TPU for parallel AI workloads.
Distributed Training — Splitting model training across multiple machines or GPUs.
Cloud Computing — On-demand computing resources delivered over the internet.
Containerization — Packaging software (like Docker) so it runs consistently across environments.
Kubernetes — A system for orchestrating containerized applications at scale.
Edge AI — Running AI models locally on devices rather than in the cloud.
On-Device Inference — Running a model directly on a phone or laptop instead of a server.
VRAM — GPU memory; a key constraint on how large a model you can run locally.
Model Compression — Techniques used to shrink a model's size for faster, cheaper deployment.
Load Balancing — Distributing incoming requests across multiple servers to avoid overload.
Auto-Scaling — Automatically adjusting compute resources up or down based on demand.
CI/CD (Continuous Integration/Continuous Deployment) — Automated pipelines for testing and shipping code (and models) reliably.
Monitoring (Observability) — Tracking a deployed system's health, performance, and behavior over time.
Rollback — Reverting a deployed model or system to a previous, known-good version.

11. Vision, Speech & Generative AI

Computer Vision (CV) — The field focused on enabling machines to interpret images and video.
Image Classification — Assigning a single label to an entire image.
Object Detection — Locating and classifying multiple objects within an image.
Image Segmentation — Labeling every pixel in an image by category.
Pooling Layer — Downsamples feature maps to reduce dimensionality in a CNN.
Feature Map — The output of a convolutional layer, highlighting detected patterns.
OCR (Optical Character Recognition) — Extracting text from images.
Bounding Box — A rectangle marking the location of a detected object in an image.
Vision Transformer (ViT) — A transformer architecture applied to image patches instead of text tokens.
Generative AI — AI that creates new content, like text, images, audio, or code.
GAN (Generative Adversarial Network) — Two networks — a generator and a discriminator — trained together to produce realistic data.
Diffusion Model — A model that generates data by learning to reverse a noise-adding process.
Latent Space — A compressed, lower-dimensional representation a model uses internally.
Variational Autoencoder (VAE) — A generative model that learns compressed representations of data.
Text-to-Image — Generating images from text descriptions.
Inpainting — Filling in missing or masked parts of an image using a model.
Style Transfer — Applying the visual style of one image to the content of another.
Denoising — The process of removing noise from data, a key step in diffusion models.
Autoregressive Model — A model that generates output one piece at a time, each conditioned on what came before.
Speech Recognition (ASR) — Converting spoken audio into text.
Text-to-Speech (TTS) — Converting text into spoken audio.
Speaker Diarization — Identifying "who spoke when" in an audio recording.
Voice Cloning — Generating synthetic speech that mimics a specific person's voice.
Spectrogram — A visual representation of audio frequencies over time, often used as model input.
Prosody — The rhythm, stress, and intonation patterns of speech, important for natural-sounding TTS.

12. Ethics, Safety, Security & Industry

AI Alignment — Ensuring an AI system's goals match human intentions and values.
Bias (in AI) — Systematic unfairness in a model's predictions, often from skewed training data.
Explainability (XAI) — The ability to understand and describe why a model made a decision.
Interpretability — How easily a human can understand a model's internal workings.
Guardrails — Rules or systems that constrain an AI's outputs to keep them safe and appropriate.
Red Teaming — Deliberately probing a model for weaknesses, harmful outputs, or vulnerabilities.
AI Governance — Policies and frameworks for overseeing responsible AI development and use.
Data Privacy — Protecting personal information used in or generated by AI systems.
Model Card — A document describing a model's intended use, limitations, and performance.
Responsible AI — An umbrella term for practices ensuring AI is developed and used ethically.
Adversarial Example — An input deliberately crafted to fool a model into a wrong prediction.
Model Poisoning — Corrupting a model by injecting malicious data during training.
Jailbreaking (LLMs) — Techniques used to bypass a model's safety restrictions.
Sandboxing — Running AI-generated code or actions in an isolated environment to limit risk.
Watermarking (AI content) — Embedding identifiable signals in AI-generated content to mark its origin.
AI Engineer — A practitioner who builds applications and systems using AI models.
Machine Learning Engineer — An engineer focused on building, training, and deploying ML models.
Data Scientist — A practitioner focused on analyzing data and building models to extract insights.
Prompt Engineer — Someone specializing in crafting effective prompts for LLMs.
MLOps Engineer — An engineer focused on the infrastructure and pipelines for deploying ML in production.
Human-in-the-Loop (HITL) — A system design where humans review or correct AI outputs as part of the process.
Artificial General Intelligence (AGI) — A hypothetical AI with human-level ability across virtually all cognitive tasks.
Artificial Narrow Intelligence (ANI) — AI specialized in a single task, describing nearly all AI in use today.
Artificial Superintelligence (ASI) — A hypothetical AI that surpasses human intelligence across all domains.
Scaling Laws — Observed patterns showing how model performance improves predictably with more data, compute, and parameters.

Wrapping Up

You don't need to memorize all 300 terms today. Bookmark this, and next time you hit an unfamiliar word in a paper, a job posting, or a tool's docs — come back and look it up. Fluency in the vocabulary is step one to actually building with AI.

What term would you add to this list? Drop it in the comments.

DEV Community