Adam (Optimizer)
A widely used optimization algorithm that adapts the learning rate for each parameter, making neural-network training faster and more stable.
Your session has been flagged for unusual activity.
Please confirm you're human to continue.
Plain-English, accurate definitions of every AI term that matters in 2026 — written for people who actually use the tools, not just write about them.
A widely used optimization algorithm that adapts the learning rate for each parameter, making neural-network training faster and more stable.
An AI system that plans and carries out multi-step tasks on its own — calling tools, reading the results, and deciding what to do next — rather than replying just once.
The work of making an AI system's behaviour match human intent and values — why a model follows instructions reliably and refuses harmful requests.
A programmatic interface that lets software talk to a model directly, bypassing the chat UI — returning structured data and charging per token.
The technique that lets a model weigh how much each part of the input matters when producing each part of the output — the core operation inside every Transformer.
The algorithm that trains neural networks by propagating prediction errors backward to adjust each weight — the engine behind nearly all deep learning.
A training technique that normalizes the inputs to each layer, allowing deeper networks to train faster and more stably.
A standardized test used to measure and compare models on tasks like reasoning, knowledge or coding — useful for ranking, but never a full substitute for trying a model on your own work.
An encoder-only Transformer pre-trained to read text in both directions at once, widely used for understanding tasks like classification, search and extraction.
A long-standing automatic metric for machine translation that scores output by how much its word sequences overlap with human reference translations.
The tendency of a neural network to lose previously learned abilities when trained on new data — a key challenge for continually updating models.
Prompting a model to reason through intermediate steps before giving a final answer, which improves accuracy on multi-step problems.
OpenAI's conversational assistant, launched in November 2022, which brought instruction-tuned large language models to a mass audience.
Anthropic's family of large language models, known for long context, strong reasoning, and a safety approach called Constitutional AI.
A neural architecture that uses learned filters to detect spatial patterns, foundational to modern computer vision.
The total amount of text — measured in tokens — that a model can consider in a single call, including the system prompt, conversation history, attachments and the reply being generated.
A measure of how similar two vectors are by the angle between them — the standard way to compare embeddings in search and RAG.
OpenAI's text-to-image model that generates images from natural-language descriptions, popularizing prompt-driven image creation.
A generative model that learns to create data by reversing a gradual noising process — the technology behind most modern AI image generation.
An alignment method that trains a model directly on human preference pairs with a simple classification loss, skipping the separate reward model used in RLHF.
A regularization technique that randomly disables neurons during training to prevent overfitting and improve generalization.
A list of numbers that represents the meaning of a piece of text, so that semantically similar items sit close together in vector space.
Giving a model a handful of worked examples inside the prompt so it can infer the pattern before answering.
Continuing to train a pre-trained model on additional, specialised data so it adapts to a domain, task or style without being built from scratch.
A generative architecture that trains two networks against each other — a generator that fabricates data and a discriminator that tries to spot the fakes.
Google DeepMind's family of natively multimodal models, built to reason jointly over text, images, audio and video.
A family of decoder-only Transformer models that are pre-trained to predict the next token on large text corpora, then adapted to downstream tasks.
xAI's large language model and assistant, integrated with the X platform and positioned around real-time information and a conversational style.
When a language model produces text that is fluent and confident but factually wrong or unsupported by its sources.
A model's ability to learn a task from examples or instructions placed in the prompt at inference time, without any change to its weights.
Fine-tuning a model on many tasks phrased as natural-language instructions, so it follows new instructions well even on tasks it never saw in training.
A prompt crafted to bypass a model's safety training and make it produce content it is meant to refuse.
Training a smaller "student" model to imitate a larger "teacher", transferring much of its capability into a cheaper, faster model.
A structured network of entities and the relationships between them, used to store facts in a form machines can query and reason over.
The delay between sending a prompt and receiving a response. Larger models and longer contexts increase latency — the main speed trade-off in AI products.
Meta's family of openly released large language models, influential for bringing strong, locally runnable models to researchers and developers.
A neural network trained on very large amounts of text to predict the next token, producing general-purpose language understanding and generation.
A parameter-efficient fine-tuning method that adapts a large model by training small added matrices while keeping the original weights frozen.
A recurrent neural network design with gating that can remember information over long sequences — the workhorse of sequence modelling before Transformers.
An open standard for connecting AI assistants to external tools and data sources through a common interface, so integrations are built once and reused across models.
An architecture that routes each input to a few specialised sub-networks instead of the whole model, giving large capacity at a lower cost per query.
A short standardized document that reports a model's intended use, performance across groups, and limitations — to encourage transparent, responsible release.
A model that works across more than one type of data — such as text, images, audio or video — in the same system.
A decoding method that samples the next token from the smallest set whose probabilities add up to p, avoiding both repetitive and incoherent text.
A standard intrinsic measure of a language model's quality: how surprised it is by held-out text. Lower perplexity means the model predicts real text better.
The input text you give a model to elicit a response. Phrasing, examples and structure strongly affect the quality of the output.
An attack that smuggles malicious instructions into a model's input — directly or via retrieved content — to override its system prompt or hijack its behaviour.
Storing a model's weights at lower numerical precision to cut memory use and speed up inference, usually with little loss in quality.
A technique that retrieves relevant documents at query time and feeds them to a language model, so its answer is grounded in real sources instead of memory alone.
Deliberately probing a model with adversarial inputs to find harmful, unsafe or policy-violating behaviour before deployment.
A second retrieval stage that re-scores an initial set of candidate passages with a more accurate model, pushing the most relevant ones to the top.
A training method that uses human preference ratings to teach a model which responses people actually prefer, making it more helpful and better aligned.
A family of automatic metrics for summarization that measure overlap between generated text and human reference summaries.
Empirical relationships showing how model performance improves predictably with more parameters, data and compute.
A reasoning method that samples several chains of thought for the same question and returns the answer most of them agree on.
An encoder–decoder neural architecture that maps one sequence to another — the design that made neural machine translation and, later, generative chat possible.
An inference-speedup technique where a small fast model drafts several tokens and a large model verifies them in parallel — faster output with identical results.
An open, latent-diffusion text-to-image model that made high-quality image generation runnable on consumer hardware.
Sending a model's response token by token as it is generated, instead of waiting for the whole answer, so text appears progressively.
Hidden, high-priority instructions that set a model's role, tone and rules before the conversation begins, framing every reply that follows.
A setting that controls randomness in a model's output: low temperature gives focused, repeatable answers; high temperature gives more varied, creative ones.
The unit a language model actually reads and writes — typically a sub-word fragment, not a whole word. Context limits, pricing and speed are all measured in tokens.
See: AI Token Counter.
Letting a model trigger external functions — search the web, run code, query a database — and use the returned results inside its answer.
The neural-network architecture behind virtually every modern large language model, built on a mechanism called self-attention instead of recurrence or convolution.
A reasoning framework that lets a model explore multiple branching solution paths, evaluate them, and backtrack — like deliberate search rather than a single guess.
A generative model that learns a probabilistic latent space, letting it both compress data and sample new examples from it.
A database that indexes and searches by embedding similarity instead of exact match, so it can find semantically related content in milliseconds.
A Transformer applied directly to image patches, showing that the architecture behind language models also excels at computer vision.
Embedding a hidden, statistically detectable signal in AI-generated text so it can later be identified as machine-produced.
OpenAI's speech-recognition model, trained on a very large, diverse audio set to transcribe and translate speech robustly across languages.
Asking a model to perform a task with no worked examples, relying entirely on the instruction and the model's pre-trained knowledge.
Reading about LLMs is one thing — running the same prompt across ChatGPT, Claude, Gemini and Grok is another.
Try MultipleChat FreeContinue learning
See the AI subscription that replaces separate ChatGPT, Claude, Gemini and Grok accounts.
Multi-model AIRun the same prompt through ChatGPT, Claude, Gemini and Grok before trusting one answer.
GuideChoose the best model for writing, research, coding, documents, images and business work.
Buyer intentOne app for several frontier AI models, side-by-side comparison and model collaboration.
PlatformA professional workspace for using several AI models, comparing answers and creating deliverables.
Core conceptUse several frontier models together for stronger answers and fewer blind spots.
ComparisonCompare several AI answers side by side when the output matters.
WorkflowBuild a professional AI stack for documents, research, writing, images and automation.
Free toolsStart with free tools, then upgrade when multi-model workflows save real time.