AI Glossary

Autonomous software system that uses an LLM as its core reasoning engine to perceive environments, formulate multi-step plans, execute tools, and iteratively achieve goals.

AI Agents extend language models beyond one-shot question answering into closed-loop goal execution. An agent maintains state, breaks complex objectives into subtasks, queries external APIs or executes shell commands, inspects execution errors, and adjusts its strategy until the target condition is satisfied. Frameworks utilize architectures like ReAct (Reasoning + Acting), Plan-and-Solve, and multi-agent coordination.

Hypothetical AI system possessing the ability to understand, learn, and apply intellectual capabilities across any domain at or above human expert level.

Artificial General Intelligence (AGI) represents the milestone where an autonomous software system can match or surpass human cognitive versatility. Unlike narrow AI systems optimized for specific tasks (like chess or speech recognition), AGI can generalize across disparate domains—from scientific discovery and complex software engineering to creative strategy and philosophical reasoning. Modern benchmarks like Humanity's Last Exam and ARC-AGI are specifically designed to measure progress toward AGI by testing reasoning resilience on novel problems.

The maximum quantity of tokens (words, subwords, symbols) a model can ingest and consider simultaneously in a single prompt and response sequence.

The Context Window defines the short-term working memory capacity of an LLM. Early transformer models operated with 2,048 or 4,096 tokens, whereas modern frontier models support 128,000 to 2,000,000 tokens (equivalent to hundreds of thousands of words or entire codebases). Maintaining accuracy across massive context spans relies on positional encoding techniques such as RoPE (Rotary Position Embeddings), YaRN, and FlashAttention.

Deep neural network trained on vast text corpora using self-supervised learning to predict tokens and understand natural language.

Large Language Models (LLMs) are foundational neural networks predominantly built upon the Transformer architecture. Trained on trillions of tokens via next-token prediction objectives and refined through Reinforcement Learning from Human/AI Feedback (RLHF/RLAIF), LLMs generalize across translation, summarization, logical reasoning, and programming code synthesis.

Architecture dividing feed-forward layers into specialized subnetworks (experts), routing each token to a subset of experts to reduce active compute per token.

Mixture of Experts (MoE) replaces dense feedforward network (FFN) layers with multiple parallel "expert" networks. A gating/router network dynamically determines which top-k experts (e.g., 2 out of 8 or 8 out of 256) process each incoming token. This decouples total model parameter capacity (e.g., DeepSeek-V3's 671B total parameters) from inference execution cost (only ~37B active parameters per token), drastically lowering latency and training FLOPs.

Technique that reduces the numerical precision of neural network weights (e.g., from FP16 to INT8, INT4, or FP8) to drastically cut VRAM requirements and accelerate inference.

Quantization compresses neural network parameters from 16-bit or 32-bit floating point representations down to 8-bit, 4-bit, or sub-4-bit integers. Techniques like AWQ (Activation-aware Weight Quantization), GPTQ, EXL2, and GGUF (used in llama.cpp and Ollama) minimize perplexity degradation by preserving salient outlier weights while quantizing non-critical parameters. This allows running a 70B parameter model on a single consumer GPU (24GB VRAM) or Apple Silicon Mac.

Architecture that retrieves authoritative information from external vector databases or search indexes and injects it into the prompt context before model generation.

Retrieval-Augmented Generation (RAG) blends information retrieval mechanisms with generative LLMs. When a user submits a query, a retrieval subsystem searches domain-specific knowledge bases (using vector embeddings, BM25 hybrid search, or graph databases) to find the most relevant document chunks. These chunks are inserted into the system prompt, allowing the LLM to generate precise, grounded answers without model retraining.

Capability enabling an LLM to generate structured machine-readable invocations (e.g. JSON) corresponding to external functions, APIs, or database queries.

Tool Calling (or Function Calling) transforms an LLM into an orchestration layer. When provided with function schemas describing parameters and descriptions, the model decides when external tools are needed and outputs clean structured parameters (matching JSON schemas) rather than conversational prose. The calling environment executes the tool and returns the result back to the model for subsequent synthesis.