Back to Blog
May 9, 2026 • Natasha Lomas, Romain Dillet, Kyle Wiggers, Lucas Ropek

AI Terms Glossary: Your Guide to Common AI Terminology

AI Terms Glossary: Your Guide to Common AI Terminology AI Terms Glossary: Your Guide to Common AI Terminology AI Terms Glossary: Your Guide to Common AI Terminology

So You've Heard These AI Terms and Nodded Along; Let's Fix That

Artificial intelligence is changing the world, and simultaneously inventing a whole new language to describe how it's doing it. Spend five minutes reading about AI and you'll run into LLMs, RAG, RLHF, and a dozen other terms that can make even very smart people in the tech world feel insecure. This glossary is an attempt to fix that — a living document updated regularly as the field evolves.

AGI (Artificial General Intelligence)

A nebulous term referring to AI that's more capable than the average human at many, if not most, tasks. Definitions vary:

  • OpenAI CEO Sam Altman: "Equivalent of a median human that you could hire as a co-worker."
  • OpenAI's charter: "Highly autonomous systems that outperform humans at most economically valuable work."
  • Google DeepMind: "AI that's at least as capable as humans at most cognitive tasks."

AI Agent

A tool that uses AI technologies to perform a series of tasks on your behalf — beyond basic chatbot capabilities. Examples include:

  • Filing expenses
  • Booking tickets or restaurant tables
  • Writing and maintaining code

An autonomous system that may draw on multiple AI systems to carry out multistep tasks. Infrastructure is still being built out to deliver on its full capabilities.

API Endpoints

Think of these as "buttons" on the back of software that other programs can press to make it do things. Developers use these interfaces to build integrations — allowing AI agents to control third-party services directly without manual human operation.

Chain of Thought

Breaking down a problem into smaller, intermediate steps to improve the quality of AI outputs. This approach:

  • Takes longer to get an answer
  • Produces more likely correct results, especially in logic or coding contexts
  • Is optimized through reinforcement learning in reasoning models

Coding Agents

Specialized AI agents applied to software development. Rather than simply suggesting code, coding agents can:

  • Write, test, and debug code autonomously
  • Handle iterative, trial-and-error work
  • Operate across entire codebases
  • Spot bugs, run tests, and push fixes with minimal human oversight

Compute

Refers to the vital computational power that allows AI models to operate. This includes:

  • GPUs, CPUs, TPUs, and other infrastructure
  • The processing fuel that powers the AI industry
  • The hardware backbone enabling model training and deployment

Deep Learning

A subset of self-improving machine learning with multi-layered, artificial neural network (ANN) structure. Key characteristics:

  • Makes more complex correlations than simpler ML systems
  • Identifies important data characteristics automatically
  • Learns from errors through repetition and adjustment
  • Requires millions of data points and longer training times

Diffusion

The tech at the heart of many art-, music-, and text-generating AI models. The process:

  • Slowly "destroys" data structure by adding noise
  • Learns a "reverse diffusion" process to restore data from noise
  • Gains the ability to recover data and generate new content

Distillation

A technique using a 'teacher-student' model to extract knowledge from large AI models:

  • Developers send requests to a teacher model and record outputs
  • Outputs are used to train the student model
  • Creates smaller, more efficient models with minimal distillation loss
  • Likely how OpenAI developed GPT-4 Turbo

Fine-tuning

Further training of an AI model to optimize performance for a more specific task by feeding in new, specialized (task-oriented) data. Many AI startups use this approach to build commercial products from large language models.

GAN (Generative Adversarial Network)

A machine learning framework involving a pair of neural networks:

  • Generator creates outputs from training data
  • Discriminator evaluates and spots artificially generated data
  • The two models compete to optimize outputs
  • Works best for narrower applications like producing realistic photos or videos

Hallucination

The AI industry's term for AI models making stuff up — generating incorrect information. This arises from:

  • Gaps in training data
  • Knowledge limitations in the model
  • Can lead to misleading or dangerous outputs
  • Contributing to push toward specialized, domain-specific AIs

Inference

The process of running an AI model — setting it loose to make predictions or draw conclusions from previously seen data. Cannot happen without training. Different hardware performs inference at different speeds, from smartphone processors to high-end AI chips.

Large Language Model (LLM)

AI models used by popular AI assistants like ChatGPT, Claude, Google's Gemini, Meta's Llama, Microsoft Copilot, and Mistral's Le Chat. Characteristics:

  • Deep neural networks made of billions of numerical parameters
  • Learn relationships between words and phrases
  • Create multidimensional maps of language
  • Generate most likely patterns that fit prompts

Memory Cache

An optimization technique that boosts inference efficiency by:

  • Saving particular calculations for future queries
  • Reducing the number of calculations a model must run
  • KV (key value) caching works in transformer-based models
  • Drives faster results by reducing algorithmic labor

Neural Network

The multi-layered algorithmic structure underpinning deep learning and generative AI. Though the concept dates to the 1940s, GPUs from the video game industry unlocked its power, enabling:

  • Better performance across many domains
  • Voice recognition, autonomous navigation, drug discovery
  • Training algorithms with many more layers than previously possible

Open Source

Software or AI models where underlying code is publicly available for anyone to use, inspect, or modify. Examples:

  • Meta's Llama family of models
  • Linux (in operating systems)
  • Enables independent safety audits
  • Contrasts with closed source (like OpenAI's GPT models)

Parallelization

Doing many things simultaneously instead of sequentially. In AI:

  • Fundamental to both training and inference
  • Modern GPUs perform thousands of calculations in parallel
  • Ability to parallelize across many chips/machines is crucial
  • Determines how quickly and cost-effectively models can be built and deployed

RAMageddon

The ever-increasing shortage of random access memory (RAM) chips. Impacts:

  • Gaming industry (raised console prices)
  • Consumer electronics (biggest smartphone shipments dip in over a decade)
  • General enterprise computing (limited data center RAM)
  • AI industry buying most RAM for data centers
  • Prices expected to remain high until shortage ends

Reinforcement Learning

Training AI where a system learns by trying things and receiving rewards for correct answers. Key aspects:

  • Like training a pet with treats, but for neural networks
  • Model explores environment, takes actions, updates behavior based on feedback
  • Powerful for training AI to play games, control robots, sharpen reasoning
  • RLHF (Reinforcement Learning from Human Feedback) is central to fine-tuning models

Token

Basic building blocks of human-AI communication, representing discrete segments of data. Created through tokenization, which:

  • Breaks down raw text into bite-sized units
  • Similar to how compilers translate human language to binary
  • Determines cost in enterprise settings (charged per-token)
  • Roughly analogous to "words" for AI workloads

Token Throughput

A measure of how much AI work a system can handle at once. High token throughput:

  • Determines how many users a model can serve simultaneously
  • Affects how quickly each user receives a response
  • Key goal for AI infrastructure teams
  • Maximizing it has become an obsession in the field

Training

The process of data being fed into a model so it can learn from patterns and generate useful outputs. Characteristics:

  • System responds to data characteristics to adapt outputs
  • Can be expensive (requires lots of inputs)
  • Volumes required have been trending upwards
  • Hybrid approaches like fine-tuning can help manage costs

Transfer Learning

Using a previously trained AI model as the starting point for developing a new model for a different but related task. Benefits:

  • Drives efficiency savings by shortcutting model development
  • Useful when data for the new task is limited
  • Has limitations — models may require additional training for domain-specific performance

Weights

Numerical parameters that determine how much importance is given to different features in training data, thereby shaping AI model output. Process:

  • Training begins with randomly assigned weights
  • Weights adjust as model seeks to match target output
  • Reflect how much each input influences the result
  • Apply multiplication to inputs

Validation Loss

A number indicating how well an AI model is learning during training — lower is better. Used to:

  • Decide when to stop training
  • Adjust hyperparameters
  • Investigate potential problems
  • Flag overfitting (when model memorizes rather than learns)
  • Distinguish between genuine understanding and rote memorization