AI Concepts Encyclopedia 2026

A subset of AI where systems learn patterns from data instead of being explicitly programmed. The three main paradigms are supervised, unsupervised, and reinforcement learning.

Learning from labeled examples, the model sees inputs paired with correct outputs and learns to predict outputs for new inputs. Includes classification and regression.

Learning patterns from unlabeled data, the model discovers hidden structure without being told the 'right answer'. Includes clustering, dimensionality reduction, and anomaly detection.

A computing system inspired by biological brains, composed of layers of interconnected nodes (neurons). Each connection has a weight that's adjusted during training. The foundation of deep learning.

A subset of ML using neural networks with many layers (hence 'deep'). Excels at learning hierarchical representations from raw data, each layer learns increasingly abstract features.

Features are the input variables (columns) used to make predictions. Labels are the output variable (what you're trying to predict). Feature engineering, crafting good inputs, is often more important than the model choice.

Overfitting: the model memorizes training data but fails on new data (too complex). Underfitting: the model is too simple to capture patterns. The goal is the sweet spot, generalizing well to unseen data.

Bias: error from overly simplistic assumptions (underfitting). Variance: error from sensitivity to training data fluctuations (overfitting). The tradeoff is fundamental, reducing one often increases the other.

Classification predicts discrete categories (spam/not spam, cat/dog). Regression predicts continuous values (price, temperature). The choice depends on whether your output is a category or a number.

Using a model trained on one task as the starting point for a different task. Instead of training from scratch, you fine-tune a pretrained model, dramatically reducing data and compute requirements.

Training: the process of learning from data (expensive, slow, uses GPUs). Inference: using the trained model to make predictions (fast, cheap). Understanding this distinction is key to ML economics.

Techniques to prevent overfitting by adding constraints to the model. L1 (Lasso) drives weights to zero for feature selection. L2 (Ridge) penalizes large weights. Dropout randomly disables neurons during training.

Few-shot: learning from just a handful of examples. Zero-shot: performing tasks with no examples, using only a description. LLMs excel at both, GPT-3 showed that scale enables few-shot learning.

The simplest ML model, fits a straight line (or hyperplane) through data points to predict a continuous value. Despite its simplicity, it's the starting point for understanding all regression models.

A hybrid approach using a small amount of labeled data with a large amount of unlabeled data. Leverages the structure in unlabeled data to improve learning, practical when labeling is expensive.

The fundamental optimization algorithm for training neural networks. It iteratively adjusts model parameters in the direction that reduces the loss function, like walking downhill to find the valley.

The algorithm for computing gradients in neural networks by propagating errors backward through layers. Popularized by Hinton in 1986, it's the engine that makes deep learning possible.

A function that measures how wrong the model's predictions are. Training = minimizing the loss. Common losses: MSE (regression), cross-entropy (classification), contrastive (embeddings).

Non-linear functions applied to neuron outputs, they give neural networks the ability to learn complex patterns. ReLU, Sigmoid, Tanh, GELU, and Swish are the most common.

Dense vector representations that capture semantic meaning. Words, images, and even users can be embedded into continuous vector spaces where similar items are close together.

Reducing the number of features while preserving important information. PCA, t-SNE, and UMAP are the main techniques, essential for visualization and fighting the curse of dimensionality.

The art of creating, transforming, and selecting input features to improve model performance. Often more impactful than model choice, 'garbage in, garbage out' applies doubly to ML.

The simplest clustering algorithm, assigns data points to K groups by iteratively moving cluster centers to minimize within-cluster distances. Fast, intuitive, but requires choosing K upfront.

Identifying rare, unusual data points that differ significantly from the majority. Critical for fraud detection, cybersecurity, manufacturing defects, and health monitoring.

An ensemble technique that combines many weak learners (usually decision trees) sequentially, with each new tree correcting the errors of previous ones. XGBoost and LightGBM dominate Kaggle competitions.

THE architecture that changed everything (2017). Uses self-attention to process sequences in parallel, replacing RNNs. Foundation of GPT, BERT, and virtually every modern AI system.

A technique allowing models to focus on relevant parts of input when producing output. Self-attention (in Transformers) lets each position attend to all other positions in a sequence.

A special case of attention where a sequence attends to itself, each token computes relationships with every other token. This is the core mechanism powering Transformers.

Architecture designed for grid-like data (images). Uses learned filters (kernels) to detect features at different scales, edges → textures → objects. Revolutionized computer vision.

Architecture for sequential data that maintains a hidden state across time steps. Suffers from vanishing gradients for long sequences, largely replaced by Transformers but still relevant for understanding.

An RNN variant that solves the vanishing gradient problem using gates (forget, input, output) that control information flow. Was the dominant sequence architecture before Transformers.

A decoder-only Transformer trained to predict the next token. Each GPT generation, from GPT-1 through GPT-5 and beyond, showed that scaling this simple idea produces increasingly capable AI. The architecture behind ChatGPT.

An encoder-only Transformer that reads text bidirectionally (both left-to-right and right-to-left). Trained with masked language modeling. Still a workhorse for cheap, high-volume classification, extraction, and retrieval, where a small encoder beats paying for a generative model.

An alternative to Transformers that processes sequences with linear-time complexity instead of quadratic attention. Mamba and S4 are the foundational SSM architectures, and hybrid attention-SSM designs now appear in production models, offering advantages for very long sequences.

Applying the Transformer architecture to images by splitting them into patches and treating each patch as a token. Showed that attention can match or exceed CNNs for vision tasks at scale.

AI models that process multiple types of data, text, images, audio, video, simultaneously. Today's frontier models from OpenAI, Google, and Anthropic are all natively multimodal; text-only models are the exception, not the rule.

Training a smaller 'student' model to mimic a larger 'teacher' model's behavior. The student learns from the teacher's soft outputs (probabilities), capturing knowledge that's lost with hard labels alone.

Reducing model size and computational cost by using lower-precision numbers (e.g., 32-bit → 4-bit floats). Enables running LLMs on consumer hardware. GPTQ, GGUF, and AWQ are popular methods.

Using AI to design AI architectures. Instead of hand-designing networks, NAS automatically searches for optimal architectures, the meta approach to building neural networks.

Combining neural networks (learning from data) with symbolic AI (logical reasoning). Aims to get the best of both worlds, learning ability plus interpretable, compositional reasoning.

The branch of AI focused on enabling computers to understand, interpret, and generate human language. Encompasses tasks from sentiment analysis to machine translation to question answering.

Breaking text into smaller units (tokens) for processing. Modern tokenizers use subword methods (BPE, SentencePiece), 'unhappiness' → ['un', 'happiness']. Token count determines cost and context window usage.

Representing words as dense vectors where semantic relationships are captured geometrically. The famous result: king - man + woman ≈ queen. Predecessor to modern embeddings used in LLMs.

The art and science of crafting inputs to LLMs to get optimal outputs. Techniques include zero-shot, few-shot, chain-of-thought, ReAct, and system prompts. One of the most practical everyday AI skills.

Combining LLMs with external knowledge retrieval. Instead of relying solely on training data, RAG retrieves relevant documents and passes them to the LLM, reducing hallucinations and adding current information.

Adapting a pretrained LLM to a specific task or domain by training on additional data. LoRA and QLoRA enable efficient fine-tuning by only updating a small fraction of parameters.

A prompting technique where the model 'shows its work' by reasoning step-by-step before answering. Dramatically improves performance on math, logic, and complex reasoning tasks.

Specialized databases optimized for storing and searching high-dimensional vectors (embeddings). Enable fast similarity search, the backbone of RAG, recommendation systems, and semantic search.

The empirical finding that LLM performance improves predictably with more data, compute, and parameters. Chinchilla scaling laws (2022) showed that most models were undertrained relative to their size.

The technique that made ChatGPT helpful and safe. Trains a reward model on human preferences, then optimizes the LLM against it. The key ingredient between a base model and a useful assistant.

The maximum number of tokens an LLM can process at once. A window of roughly one million tokens is now standard at the frontier, with a few models advertising ten million. That is enough for entire codebases or books, but effective recall degrades well before the advertised limit, so the usable window is usually much smaller than the number on the spec sheet.

How LLMs produce text, sampling from probability distributions over tokens. Temperature controls randomness. Top-k and top-p (nucleus) sampling balance creativity and coherence.

Determining the emotional tone of text, positive, negative, or neutral. One of the most common NLP applications. Modern approaches use fine-tuned BERT or zero-shot LLMs.

Automatically translating text between languages. Evolved from rule-based → statistical → neural (seq2seq) → Transformer-based. Google Translate and DeepL use Transformer architectures.

Identifying and classifying named entities in text, persons, organizations, locations, dates, etc. A fundamental NLP task used in information extraction, search, and knowledge graphs.

Assigning a label to an entire image. The task that started the deep learning revolution, AlexNet won ImageNet 2012 by a huge margin, proving deep learning works.

Locating and classifying multiple objects within an image with bounding boxes. YOLO (You Only Look Once) made real-time detection practical. Used in autonomous driving, security, and robotics.

Classifying every pixel in an image, assigning each pixel to a category. More granular than object detection. Used in autonomous driving, medical imaging, and satellite imagery.

A real-time object detection system that processes the entire image in a single forward pass. The family is updated almost yearly (YOLOv8, YOLO11, YOLO26 and beyond), each generation faster and more accurate, keeping it the go-to for practical deployment.

Creating new images from text descriptions or other inputs. Stable Diffusion, Midjourney, FLUX, and the image models built into ChatGPT and Gemini can generate photorealistic images from text prompts.

Applying the artistic style of one image to the content of another, e.g., making your photo look like a Van Gogh painting. Uses CNN feature representations to separate content from style.

AI-generated fake videos and images that swap faces or create entirely synthetic people. Uses GANs and diffusion models. Raises serious ethical concerns about misinformation.

Predicting future values based on historical temporal data. Traditional methods (ARIMA) are being augmented by deep learning (LSTMs, Transformers). Critical for finance, weather, and demand planning.

AI is only as good as its data. Understanding data quality, bias, preprocessing, augmentation, and the data lifecycle is fundamental. 'More data beats better algorithms', but only if the data is good.

Artificially generated data that mimics real-world data. Used when real data is scarce, expensive, or privacy-sensitive. GANs and diffusion models can create realistic synthetic training data.

The architecture behind most modern image and video generators, including Stable Diffusion and the image models inside the major assistants. Works by learning to reverse a noise-adding process, start from pure noise, gradually denoise to create an image. Replaced GANs as the dominant generative approach.

Two neural networks in competition: a generator creates fake data, a discriminator tries to detect it. The adversarial game produces increasingly realistic outputs. Introduced by Goodfellow in 2014.

Generative models that learn a compressed latent space representation. Unlike GANs, VAEs can both generate AND encode data. The latent space is continuous, enabling smooth interpolation between outputs.

The open-source text-to-image model that democratized AI art. Works in a compressed latent space (not pixel space) for efficiency. SDXL and SD 3.5 produce photorealistic images from text prompts, and open-weight successors such as FLUX build on the same latent-diffusion recipe.

Generating video from text descriptions. Google's Veo, Runway's Gen models, Kling, and OpenAI's Sora pushed the field from short silent clips to video with native synchronized audio; the leaderboard changes every few months. Extends diffusion models with temporal consistency, the next frontier of generative AI.

AI systems that can replicate any voice from a short sample, or generate natural-sounding speech from text. ElevenLabs and OpenAI's TTS models can produce indistinguishable synthetic voices.

Creating original music using AI, from background scores to full songs with vocals. Suno and Udio generate music from text prompts. The copyright fight moved from open question to concrete outcome: major labels have signed licensing deals with both while other suits continue, and licensed models change what users may download and reuse.

LLMs trained on code that can write, debug, and explain programs. GitHub Copilot, Claude, and other frontier models can generate working code from natural language descriptions, and coding agents can now carry out whole tasks. Transforming software development.

Creating 3D objects and scenes from text or 2D images. Neural Radiance Fields (NeRFs) reconstruct 3D scenes from photos. Gaussian Splatting (2023) was a breakthrough for real-time 3D rendering and is now the standard approach.

Using AI to edit specific parts of an image, removing objects, changing backgrounds, adding elements. Inpainting fills in missing or masked regions. Powered by diffusion models.

Learning through trial and error, an agent takes actions in an environment, receives rewards/penalties, and learns a policy to maximize cumulative reward. Behind AlphaGo and game-playing AI.

Q-Learning estimates the value of actions in each state. Deep Q-Networks (DQN) use neural networks to approximate Q-values, DeepMind's DQN played Atari games at superhuman level (2013).

Policy gradient methods directly optimize the policy (action selection) rather than value estimates. Actor-Critic combines both: the actor selects actions, the critic evaluates them.

For years the default RL algorithm, simpler and more stable than its predecessors, and the method used in the original RLHF pipelines. Newer group-relative variants have largely displaced it for language-model post-training, though PPO remains a standard baseline in robotics and control. Used in RLHF to fine-tune ChatGPT, in robotics, and in game AI. Balances exploration with stable training.

The mathematical framework underlying RL. Defines states, actions, transition probabilities, and rewards. The Markov property: the future depends only on the current state, not history.

Designing reward functions is RL's hardest problem. Reward hacking: agents find unintended shortcuts to maximize reward without actually solving the task. A core challenge in AI alignment.

Multiple agents learning simultaneously in a shared environment, they can cooperate, compete, or both. Adds complexity: each agent's optimal strategy depends on what others do.

AI that builds internal models of how the world works, predicting what happens next without actually experiencing it. Championed by Yann LeCun, who left Meta at the end of 2025 to found AMI Labs (Advanced Machine Intelligence) specifically to pursue this, as the next paradigm: learning through observation, not just language.

Training RL agents in simulation and deploying them in the real world. Overcomes the cost and danger of real-world training. Domain randomization helps bridge the gap between sim and reality.

Learning the reward function from observing expert behavior, inferring WHAT the expert is optimizing, not just imitating their actions. Key for building AI that understands human preferences.

GPUs (NVIDIA) and TPUs (Google) are the hardware that makes deep learning possible. Their parallel processing capabilities accelerate matrix operations by 100x+ versus CPUs.

Getting trained models into production where they serve real users. Involves containerization (Docker), API frameworks (FastAPI), model servers (TensorRT, vLLM), and monitoring.

The practice of deploying, monitoring, and maintaining ML models in production. Combines ML, DevOps, and data engineering. The difference between a notebook prototype and a reliable production system.

Systematically logging hyperparameters, metrics, and artifacts for every ML experiment. Without it, ML research is unreproducible chaos. Weights & Biases and MLflow are the standards.

Automated workflows for collecting, cleaning, transforming, and loading data for ML. The unsexy but critical infrastructure that makes everything else possible.

Running AI models directly on devices (phones, IoT, cars) instead of in the cloud. Enables low-latency, privacy-preserving AI. Requires model compression and optimization.

Training ML models across multiple devices/organizations without sharing raw data, each participant trains locally and shares only model updates. Preserves privacy while enabling collaborative learning.

Training models across multiple GPUs or machines. Essential for frontier models, which are trained on clusters ranging from tens of thousands to several hundred thousand accelerators. Techniques: data parallelism, model parallelism, pipeline parallelism.

Tracking different versions of ML models with their metadata, performance metrics, and lineage. Essential for reproducibility and production deployment. MLflow Model Registry is the standard.

Watching ML models in production for degradation. Models decay over time as data distributions shift. Monitoring detects data drift, concept drift, and performance drops before they impact users.

AI systems that can plan, use tools, browse the web, write code, and complete multi-step tasks autonomously. The defining frontier of the current AI era. Frameworks such as LangGraph and CrewAI, plus vendor SDKs (OpenAI Agents SDK, Claude Agent SDK, Microsoft Agent Framework), are common starting points.

LLMs that can call external tools, APIs, databases, calculators, web browsers. Transforms LLMs from text generators into action-taking agents. Function calling is now supported by every major vendor, and the Model Context Protocol (MCP) has become the industry-standard way to connect models to tools.

Self-driving cars and intelligent robots that perceive, plan, and act in the physical world. Combines computer vision, RL, sensor fusion, and path planning. Tesla, Waymo, and Figure lead the field.

AI that suggests content, products, or connections based on user behavior and preferences. Powers Netflix, YouTube, Amazon, and TikTok. Uses collaborative filtering, content-based, and hybrid approaches.

Using AI to identify, design, and optimize drug candidates, dramatically accelerating the traditional 10-15 year drug development cycle. AlphaFold made protein structure prediction reliable enough to use as a tool. Insilico Medicine's rentosertib is the first candidate whose biological target and molecule were both found with generative AI, and it cleared a Phase IIa readout in 2025. It is still investigational, not approved, which is the honest state of the field.

Techniques to make AI decisions interpretable by humans. SHAP and LIME explain individual predictions. Critical for healthcare, finance, and legal applications where decisions must be justified.

Using ML to improve search results and ranking. Moved from keyword matching to semantic understanding. Encoder models moved search from keyword matching to semantic understanding, and vector search enables meaning-based retrieval. The frontier has moved again: major engines now generate an answer as the primary result and cite sources into it, so visibility increasingly means being cited rather than being ranked.

AI systems for natural dialogue, from simple rule-based chatbots to LLM-powered assistants. Key challenges: maintaining context, handling ambiguity, knowing when to escalate to humans.

AI for diagnosis, treatment planning, drug discovery, and medical imaging. CNNs match or exceed radiologists for certain diagnoses. Faces unique challenges: regulation, liability, bias in clinical data.

AI tools transforming art, music, writing, film, and design. From AI-assisted writing to AI-generated music and video. Raises questions about authorship, copyright, and the future of creative work.

The challenge of ensuring AI systems pursue goals beneficial to humans. The most important unsolved problem in AI. Includes technical approaches (RLHF, Constitutional AI) and governance frameworks.

AI systems can perpetuate and amplify societal biases present in training data. Understanding, measuring, and mitigating bias is critical for responsible AI deployment.

When AI generates confident but factually incorrect or fabricated information. A fundamental limitation of current LLMs, they predict plausible text, not truthful text. RAG and grounding techniques help reduce it.

Systematically testing AI systems by trying to make them fail, produce harmful outputs, or bypass safety measures. Essential for finding vulnerabilities before deployment. Both manual and automated approaches.

The emerging legal and policy landscape for AI. The EU AI Act is the first comprehensive AI law. Its obligations phase in over several years, and a 2026 amendment pushed the high-risk rules for standalone systems to December 2027 and for AI embedded in regulated products to August 2028. Treat the phase-in dates as current as of reading, not fixed. Understanding regulation is now a career advantage, every AI team needs someone who understands compliance.

An open standard, introduced by Anthropic in late 2024, for connecting AI models to external tools, data sources, and applications through one common interface. Developers expose a capability once as an MCP server and any compatible model can use it, replacing per-model custom integrations. By 2026 it is supported by every major AI vendor.

LLMs trained to produce a long internal chain of reasoning before answering, spending extra compute at inference time to get better answers on math, code, and logic. OpenAI's o-series and DeepSeek-R1 established this as a second scaling axis alongside model size. Most frontier assistants now offer a thinking mode built on this idea.

A Transformer variant where each layer holds many parallel expert subnetworks and a router activates only a few per token. This decouples total parameter count from per-token compute, so models can grow far larger without proportional cost. Many frontier and open-weight LLMs use MoE.

Agents that operate software the way a person does: looking at the screen, moving the cursor, clicking, and typing, or driving a web browser directly. Introduced commercially in 2024-2025 and now built into AI-enabled browsers. It turns any existing application into something an AI can operate, with no API required.

Coding assistance that goes beyond autocomplete: an agent reads the codebase, plans, edits multiple files, runs tests and commands, and iterates until the task is done. Tools built on this pattern made it the dominant developer AI workflow.

A reinforcement learning algorithm that simplifies PPO by dropping the value network: it samples a group of responses per prompt and scores each one against the group average. Introduced in DeepSeekMath and used to train DeepSeek-R1, it became a standard recipe for training reasoning models on verifiable rewards like math and unit tests.

Measuring what language models can actually do, via static benchmarks, head-to-head human preference arenas, and LLM-as-judge scoring. Benchmark saturation and training-data contamination make honest evaluation one of the field's hardest open problems: leaderboard gains do not always translate to real-world capability.

The practice of deciding what occupies a model's context window on each call: system instructions, tool definitions, retrieved documents, and prior turns. Once an agent runs for many steps the binding constraint stops being prompt wording and becomes the budget of useful tokens, so compaction, summarising older turns, keeping notes outside the window, and retrieving just in time matter more than phrasing.

An attack where instructions hidden inside content the model reads are followed as if the user had issued them. Direct injection comes from whoever is typing; indirect injection arrives through a fetched web page, document, or tool result, which makes it the central security problem for tool-using agents. No complete fix exists, so defences layer restricted permissions, separation of instructions from data, and human confirmation before consequential actions.

Reverse engineering a network's internal computation into human-readable parts: features, the circuits that combine them, and how information moves between layers. Sparse autoencoders are the main tool for pulling interpretable features out of layers whose individual neurons each represent many unrelated things. It differs from feature-attribution methods, which say which inputs mattered without describing the mechanism that used them.

The engineering that makes serving a model affordable. Generation splits into a compute-bound prefill over the prompt and a memory-bound decode of one token at a time, so throughput and cost are set by the KV cache, continuous batching, attention memory layout, and speculative decoding, where a small draft model proposes tokens a larger model verifies in one pass. Two deployments of the same weights can differ by an order of magnitude in cost per token.

Models that take camera images plus a natural-language instruction and output robot actions directly, trained across large mixed datasets of demonstrations rather than one policy per task. They port the pretrain-then-adapt pattern from language into physical control. The property under test is generalisation: whether a policy handles objects, scenes, and phrasings it never saw during training.

Splitting a task across several LLM agents with separate contexts and roles, coordinated by an orchestrator that fans out subtasks and merges what comes back. It pays off when subtasks are genuinely independent and each needs its own large context. It also multiplies token spend and makes failures harder to trace, so a single well-scoped agent is often the better answer.