lite

Resource Center Glossary

The AI Glossary

Every AI term, in plain English — from backpropagation to spicy autocomplete, plus the words lite.computer uses for its own parts.

Every term below is defined once, in plain English, with no assumed knowledge. It covers general artificial intelligence vocabulary, the slang practitioners actually use, the labs and model families worth knowing, and the words lite.computer uses for its own parts.

Model families are named by family, not by version number. Versions change every few weeks; a glossary that carries them is wrong within a month and misleading within three.

Foundations

The words underneath everything else.

Algorithm

A set of rules or instructions that a computer follows to complete a task.

Artificial General Intelligence (AGI)

A hypothetical AI that matches or exceeds human ability across essentially any intellectual task, rather than being good at one. Nobody agrees on how it would be measured, which is why the term causes more argument than it settles.

Also: AGI

See also: Singularity, Narrow AI

Artificial Intelligence (AI)

The simulation of human intelligence processes by machines, especially computer systems.

Also: AI

Artificial Neural Network (ANN)

Computing systems loosely inspired by the biological neural networks of human brains.

Also: Neural network, ANN

Bias (in AI)

When an algorithm produces results that are systematically prejudiced, because of erroneous assumptions or skewed data in the learning process.

Chatbot

A software application designed to simulate human conversation, ranging from simple rule-based systems to advanced AI.

Data Mining

The process of extracting hidden patterns and knowledge from large sets of data.

Deep Learning

A subset of machine learning that uses multi-layered neural networks to model complex patterns in data.

See also: Artificial Neural Network (ANN)

Deterministic

A system that always produces the same output for the same input. Traditional software is deterministic; language models are not, which is the single biggest adjustment for people coming from ordinary software.

Also: Deterministic data

See also: Probabilistic, Wiggly

Heuristic

A problem-solving shortcut or rule of thumb that produces a solution quickly, even if it is not mathematically perfect.

Intent

The specific goal or purpose behind a user's interaction or query — booking a flight, as opposed to checking a flight's status.

Machine Learning (ML)

A subset of AI focused on building systems that learn from data and improve their performance over time without being explicitly programmed for every task.

Also: ML

Narrow AI

An AI built for one job — spotting fraud, transcribing speech, recommending films. Every system in production today is narrow AI, however general it feels to use.

Also: Weak AI

See also: Artificial General Intelligence (AGI)

Perceptron

The simplest type of artificial neuron, and the building block used to create larger neural networks.

Probabilistic

A system whose outputs carry uncertainty and are expressed as likelihoods rather than certainties. A probabilistic model answers with a confidence, not a fact.

Also: Probabilistic data

See also: Deterministic, Temperature

Rule-based bot

A bot that follows a strict, pre-programmed script or if-then logic rather than using AI to understand nuance.

Also: Decision-tree bot

See also: Symbolic AI

Semantic Web

An extension of the web that makes data machine-readable, letting computers understand the meaning of and relationships between pieces of information.

See also: Structured data, Schema.org

Singularity

A hypothetical future point where AI surpasses human intelligence, leading to rapid and unpredictable technological growth.

Symbolic AI

The older tradition of AI built on explicit rules and logical symbols rather than learned statistics. Rule-based bots and expert systems are its surviving descendants.

Also: Good Old-Fashioned AI, GOFAI

See also: Rule-based bot

Turing Test

A test of a machine's ability to exhibit intelligent behavior equivalent to, or indistinguishable from, that of a human.

Language models

How the models that write text are built.

Attention

The mechanism that lets a model weigh which parts of its input matter most for the word it is producing now. It is what the transformer paper's title refers to.

Also: Self-attention

See also: Transformer

Autoregressive

Generating output one token at a time, each one conditioned on everything produced so far. Why text appears word by word rather than all at once.

Also: Decoder-only

See also: Token

Convolutional Neural Network (CNN)

A class of deep neural network used mainly for processing and analyzing visual data such as images and video.

Also: CNN

Diffusion model

The architecture behind most AI image and video generation. It starts from noise and removes it step by step until an image emerges. Largely replaced GANs for image work.

See also: Generative Adversarial Network (GAN)

Distillation

Training a small model to imitate a larger one, so that a cheap model inherits much of an expensive one's behavior.

Also: Knowledge distillation

Embedding

A list of numbers representing the meaning of a piece of text, image or audio, positioned so that similar things sit close together. The basis of semantic search and most retrieval.

Also: Embeddings, Vector

See also: Vector database, RAG (Retrieval-Augmented Generation)

Foundation model

A large model trained once on broad data and then adapted to many downstream jobs, rather than trained from scratch for each one.

Also: Base model

See also: Fine-tuning

Frontier model

The most capable models in existence at a given moment — the handful at the leading edge, usually from a small number of well-funded labs. The membership list changes every few months.

Also: Frontier models

See also: Frontier lab

Generative Adversarial Network (GAN)

A framework where two neural networks — a generator and a discriminator — compete against each other to produce highly realistic synthetic data.

Also: GAN

Generative AI

Deep-learning models that create new content — text, images, audio, video, code — based on patterns in their training data.

Also: GenAI

GPT (Generative Pre-trained Transformer)

An AI architecture that uses a transformer design to process language and generate human-like text. Also the name OpenAI gives its model family, which is why the acronym is often used as though it meant one company's product.

Also: GPT

See also: Transformer

Inference

Running a trained model to get an answer, as opposed to training it. Almost all the money spent on a deployed AI system is spent here.

See also: Training

KV cache

The stored intermediate state that lets a model avoid re-reading the whole conversation for every new token. It is what makes long conversations affordable, and what consumes the memory they need.

See also: Context window, Context caching

Large Language Model (LLM)

An AI trained on enormous amounts of text to understand, summarize and generate human language.

Also: LLM

Latency

How long you wait for an answer. Often split into time to first token — how long before anything appears — and total time, which matter to users in different ways.

Also: Time to first token

Long Short-Term Memory (LSTM)

A recurrent network design that holds information over longer sequences. The state of the art for language before transformers arrived.

Also: LSTM

See also: Recurrent Neural Network (RNN)

Mixture of Experts (MoE)

A design that splits a model into many specialist sub-networks and activates only a few per token. It buys the capability of a very large model at the running cost of a much smaller one.

Also: MoE

Multimodal

A model that handles more than one kind of input or output — text and images, or text and audio and video — rather than text alone.

Also: Multimodal model

Parameters

The learned numbers inside a model — the weights adjusted during training. Model sizes are quoted in billions of them, though parameter count alone stopped predicting capability some time ago.

Also: Weights

See also: Open weight

Quantization

Shrinking a model by storing its numbers at lower precision. It costs a little accuracy and buys a large amount of speed and memory, which is what lets big models run on a laptop.

See also: Local AI, Small Language Model (SLM)

Reasoning model

A model trained to work through a problem before answering, and to spend more computation on harder questions. The distinction from an ordinary model is how long it thinks, not what it knows.

See also: Chain of thought, Test-time compute

Recurrent Neural Network (RNN)

An older architecture that reads sequences one step at a time, carrying state forward. Largely displaced by transformers because it cannot be parallelized.

Also: RNN

See also: Transformer, Long Short-Term Memory (LSTM)

Small Language Model (SLM)

A deliberately compact language model, small enough to run on a laptop or a phone. It trades breadth for speed, cost and privacy.

Also: SLM

See also: Local AI, Quantization

Structured output

Constraining a model to answer in a fixed shape — valid JSON matching a schema — so a program can consume the answer instead of a person reading it.

Also: JSON mode

See also: JSON, Tool call

Temperature

A setting that controls how much randomness a model uses when choosing its next token. Low temperature is repetitive and predictable; high temperature is varied and more likely to go off the rails.

See also: Probabilistic, Wiggly

Token

The unit a language model actually reads and writes. Roughly three-quarters of a word in English, so a page of prose is about 500 tokens. Pricing, limits and speed are all measured in them.

Also: Tokens

See also: Tokenization, Context window

Tokenization

Cutting text into tokens before a model sees it. It is why a model can miscount the letters in a word: it never saw the letters.

See also: Token

Top-p

A sampling setting that limits a model's word choice to the most likely options adding up to a given probability. Used alongside temperature to control how adventurous the output is.

Also: Nucleus sampling, Top-k

See also: Temperature

Transformer

The neural network architecture introduced in the 2017 paper "Attention Is All You Need". It replaced older designs such as RNNs because it processes data in parallel, making it far faster and more powerful at scale. Nearly every modern language model is a transformer.

See also: Attention, GPT (Generative Pre-trained Transformer)

Training

How a model learns, and how it goes wrong.

Backpropagation

The algorithm that works out how much each weight in a network contributed to the error, by passing the loss backwards through it. An optimizer then uses those numbers to actually adjust the weights.

Also: Backprop

See also: Loss function

DPO (Direct Preference Optimization)

A simpler alternative to RLHF that learns from preferred and rejected answer pairs directly, with no separate reward model. Now as common as RLHF in practice.

Also: DPO

See also: RLHF (Reinforcement Learning from Human Feedback), Reward model

Epoch

One complete pass through the training data.

Feature engineering

Selecting, transforming or creating the most relevant variables — features — to improve a machine learning model's performance.

Few-shot

Giving a model a handful of worked examples inside the prompt so it can infer the pattern. Often the cheapest alternative to fine-tuning.

Also: Few-shot learning

See also: One-shot, Zero-shot learning

Fine-tuning

Further training of an existing model on a narrower dataset, so it picks up a specific style, format or domain.

See also: LoRA (Low-Rank Adaptation), Transfer learning

Gradient descent

The optimization method that walks a model's weights downhill on the loss, one small step at a time.

See also: Loss function

LoRA (Low-Rank Adaptation)

A cheap way to fine-tune: instead of retraining a whole model, train a small set of extra weights alongside it. Hours on one GPU rather than a data-center run.

Also: LoRA

See also: Fine-tuning

Loss function

The measure of how wrong a model's output is. Training is the process of making this number smaller.

Also: Objective function

See also: Backpropagation

Model collapse

A proposed failure mode where training models on the output of earlier models, generation after generation, drains the diversity of the original human data. A risk to test for rather than an automatic consequence of using synthetic data.

See also: Synthetic data, Hallucination laundering

One-shot

Giving a model exactly one worked example before asking it to do the task. Also used loosely to mean getting something right on the first attempt, with no follow-up prompt — "it one-shotted the whole feature".

Also: One-shot learning

See also: Zero-shot learning, Few-shot

Overfitting

When a model learns the training data too well, including its noise and outliers, and so performs poorly on new, unseen data.

See also: Underfitting

Post-training

Everything done to a model after pre-training to make it useful and well-behaved: instruction tuning, RLHF, safety work. Increasingly where the difference between models is decided.

See also: RLHF (Reinforcement Learning from Human Feedback), Fine-tuning

Pre-training

The first and largest training stage, where a model learns language and world knowledge from a vast general corpus. Everything after it is comparatively small.

See also: Post-training, Fine-tuning

Reinforcement learning

A type of learning where an agent learns to make decisions by taking actions in an environment to maximize a reward.

Also: RL

Reinforcement learning with verifiable rewards

Training on tasks where correctness can be checked mechanically — the tests pass, the proof holds, the number is right. It sidesteps the need for a reward model, and is much of why reasoning models improved so quickly.

Also: RLVR

See also: Reasoning model, Reward model

Reward model

A model trained to score how good an answer is, standing in for human judgment during reinforcement learning.

See also: RLHF (Reinforcement Learning from Human Feedback)

RLHF (Reinforcement Learning from Human Feedback)

Training a model using human preferences between its own outputs, to make it more helpful and better behaved. The step that turned raw language models into usable assistants.

Also: RLHF

See also: RLHF'd to death, Post-training, Alignment

Scaling laws

The observed relationships between how much compute, data and parameters go into a model and how good it turns out. The basis of the bet that bigger keeps working.

Self-supervised learning

Training where the data supplies its own labels — for example, hiding the next word and asking the model to predict it. How language models are pre-trained.

See also: Pre-training

Supervised fine-tuning (SFT)

Fine-tuning on example pairs of instruction and ideal answer. Usually the first step of post-training, before any preference learning.

Also: SFT, Instruction tuning

See also: Post-training, Fine-tuning

Supervised learning

Training a model on a labeled dataset, where the correct answers are provided so the model can learn the relationship between input and output.

Synthetic data

Training data generated by a model rather than collected from the world. Now a large share of what frontier models are trained on, and a live worry about compounding errors.

See also: Model collapse, Hallucination laundering

Training

Adjusting a model's parameters by showing it data. Pre-training is one large run; a model can be fine-tuned and post-trained afterwards. All of it is separate from inference, which is what happens when you actually use the model.

See also: Inference, Pre-training

Training cutoff

The latest period a model's training data covers, usually stated only approximately. It bounds what the model knows by itself, which is why it can be confidently wrong about recent events and why search and retrieval tools exist.

Also: Knowledge cutoff

See also: RAG (Retrieval-Augmented Generation)

Transfer learning

Taking a model trained on one task and repurposing it for a second, related task — using a model trained on photos of cars to help identify trucks.

See also: Fine-tuning

Underfitting

When a model is too simple to learn the underlying patterns in the data, and so performs poorly on both training and new data.

See also: Overfitting

Unsupervised learning

Training a model on data without explicit labels, so it must find its own patterns and structures in the information.

Zero-shot learning

The ability of a model to correctly complete a task or recognize a category it was never specifically trained on.

Also: Zero-shot

See also: One-shot, Few-shot

Agents

Software that acts, not just answers.

Agentic

Describes software that plans and acts over several steps rather than answering once. An agentic workflow decides what to do next by itself; a prompt does not.

See also: AI agent

AI agent

An autonomous system that can perceive its environment, reason about how to achieve a goal, and take actions using tools or skills, without step-by-step human instruction.

Also: Agent

See also: Agentic, Skills (or tools)

API (Application Programming Interface)

The definitions and protocols that let two software components talk to each other.

Also: API

Coding agent

An agent that reads and writes a codebase, runs commands and iterates on the result — as opposed to a chat window that suggests code you paste yourself.

Also: Agentic coding tool

See also: Vibe coding, Claude Code

Computer use

An agent operating a computer the way a person does — reading the screen, moving the pointer, typing. The general-purpose fallback for software that offers no API.

See also: AI agent, Skills (or tools)

Context engineering

Deciding what a model should have in front of it at each step, and what it should not. The successor discipline to prompt engineering, and the one that decides whether a long-running agent stays coherent.

See also: Context window, Context rot, Prompt engineering

Guardrails

The checks around a model that constrain what it may say or do — filters, allow-lists, approval gates. A guardrail is enforced outside the model, unlike alignment, which is trained into it.

See also: Alignment

Harness

The structural framework or environment around a model that gives it the tools, data and testing protocols it needs to do work or be measured. The model reasons; the harness is everything that lets the reasoning touch the world.

Also: Agent harness

See also: Agentic, Scaffolding

Human in the loop

A design where a person approves, corrects or steers the system at defined points, rather than the system running unattended.

Also: HITL

Jailbreak

A prompt designed to talk a model past its own safety training.

See also: Prompt injection, Alignment

MCP (Model Context Protocol)

An open standard that lets AI models connect to external data sources and tools — local files, databases, applications — without custom code for every integration.

Also: MCP, Model Context Protocol

See also: MCP server, Skills (or tools)

MCP client and server

The two halves of the Model Context Protocol. The server exposes tools; the client, inside the AI application, connects to it and offers those tools to the model.

Also: MCP host

See also: MCP (Model Context Protocol), MCP server

MCP server

A program that exposes a set of tools over the Model Context Protocol, so any compatible AI can use them. lite.computer ships one, which is how an agent reads and drives a Showspace.

See also: MCP (Model Context Protocol)

Multi-agent

Several agents working on one problem, often with different roles or models, and often reviewing each other. More capable and considerably harder to keep honest than one agent.

Also: Agent swarm

Prompt

The instruction given to a model. In an agentic system it is rarely one line: it carries the goal, the rules, the available tools and the relevant context.

Prompt engineering

The craft of writing prompts that get reliable results. Steadily giving ground to context engineering as models get better at instructions and worse at nothing else.

See also: Context engineering

Prompt injection

An attack where instructions hidden in content the model reads — a web page, an email, a file — are followed as though the user had given them. The central unsolved security problem of agents that read the outside world.

See also: Guardrails, Jailbreak

Scaffolding

The loop, prompts and glue code wrapped around a model to make it complete a multi-step task. Often the difference between two products using the identical model.

See also: Harness

Skills (or tools)

Specific functions or capabilities given to an AI — searching the live web, running code, reading a folder, calling a particular piece of software.

Also: Tools, Tool use, Function calling

See also: MCP (Model Context Protocol), Tool call

System prompt

The standing instruction that sits above a conversation and governs how the model behaves throughout it. Where an application's rules, persona and constraints live.

Also: Standing instructions

See also: Prompt

Tool call

One request from a model to run a named function with named arguments, and the result handed back to it. The unit an agentic loop is built from.

Also: Function call

See also: Skills (or tools)

Webhook

An automated message sent from one application to another when a specific event happens — the push notification of the data world.

Context and memory

What a model knows while it is working, and what it keeps.

Chunking

Cutting documents into passages small enough to retrieve and feed to a model. Where the chunks are cut decides how good the retrieval is, which is why it gets more attention than it looks like it deserves.

See also: RAG (Retrieval-Augmented Generation)

Citation

A pointer from a generated answer back to the source it came from, so a reader can check it. Increasingly what separates a useful AI answer from an unusable one.

See also: Grounding

Context

The information and circumstances surrounding an interaction that let the system understand and answer accurately.

Context window

How much a model can hold in mind at once, measured in tokens — the prompt, the conversation so far, every document and tool result. Everything outside it does not exist to the model.

Also: Context length

See also: Token, Context rot, Context engineering

Grounding

Supplying real evidence — retrieved documents, database rows, tool output — and holding the answer to it, rather than to the model's recollection. The main defense against hallucination, though a citation on its own does not prove an answer was grounded.

See also: Hallucination, RAG (Retrieval-Augmented Generation)

Knowledge base

An organized body of documents a person or a system can consult. When an AI is the reader rather than a person, the format matters more than the design.

Also: KB

Memory (in AI)

Anything that persists between sessions — facts a system stores about you, notes it keeps, files it writes. Different from the context window, which is what it holds while working, and which the application decides whether to carry over, summarize or clear.

Also: Persistent memory

See also: Context window, Second brain

Obsidian vault

A folder of plain Markdown files that the Obsidian app reads as a linked knowledge base. Because it is only files on disk, an AI agent can read and write it directly — which is why it became a common home for a second brain.

Also: Vault

See also: Second brain, Markdown

RAG (Retrieval-Augmented Generation)

A technique that gives a model access to specific external data — a company handbook, a folder of documents — and puts it in front of the model as it answers, rather than leaving it to remember. It is only ever as good as what the retrieval step found.

Also: RAG, Retrieval

See also: Vector database, Embedding, Training cutoff

Second brain

A personal knowledge system you offload thinking into — notes, documents and links kept so that finding something later is easier than remembering it. Popularized by note-taking practice, and now the natural place for an AI to read and write.

Also: Personal knowledge management, PKM

See also: Obsidian vault, Knowledge base

Vector database

A database that stores embeddings and finds the closest ones to a query. The retrieval half of most RAG systems.

Also: Vector store

See also: Embedding, RAG (Retrieval-Augmented Generation)

Evaluation and safety

Measuring it, and the ways it fails.

Alignment

The problem of making a system actually pursue what its operators intend, including in situations nobody wrote a rule for.

See also: RLHF (Reinforcement Learning from Human Feedback), Guardrails

Benchmark

A standard test set used to compare models. Useful, and steadily less meaningful as popular benchmarks leak into training data.

See also: Evals, Benchmark contamination

Benchmark contamination

When a benchmark's questions and answers end up in a model's training data, so a high score measures memory rather than ability.

Also: Test set contamination

See also: Benchmark

Black box

A system whose internal reasoning cannot be inspected, only its inputs and outputs.

See also: Explainability

Chain of thought

The intermediate steps a model works through before answering. Some products show a summary of them; many keep the model's actual reasoning private, so what you see is not necessarily what happened.

Also: CoT, Reasoning

See also: Test-time compute

Drift

When a system's behavior changes over time without anyone changing it on purpose — a provider updates a model, the data shifts, and outputs quietly stop matching what was tested.

Also: Model drift

See also: Evals

Evals

Short for evaluations, and often used as an imperative: don't ship on vibes, what do the evals say? The tests used to measure whether a model or a system is actually doing its job — accuracy, latency, cost, regressions. A benchmark is one kind of eval, not the whole of it.

Also: Evaluations, Eval

See also: Benchmark, Vibes

Explainability

How well a human can understand why a system produced a given output. Hard for deep learning, and often a regulatory requirement.

Also: Interpretability, XAI

See also: Black box

Hallucination

When an AI generates a response that sounds confident and fluent but is factually incorrect or disconnected from reality.

Also: Confabulation

See also: Grounding, Hallucination laundering

Red teaming

Deliberately attacking your own system to find what it will do wrong, before someone else does it for you.

See also: Jailbreak, Prompt injection

Refusal

When a model declines a request. A necessary behavior that becomes a defect when it fires on harmless prompts.

Also: Over-refusal

See also: RLHF'd to death

Sycophancy

A model's tendency to agree with the user, praise their idea and fold under pushback, because agreement was rewarded during training. Dangerous precisely when you wanted a second opinion.

See also: RLHF'd to death

Test-time compute

Spending more computation at the moment of answering — thinking longer — rather than making the model bigger. The idea behind reasoning models.

Also: Inference-time compute

See also: Chain of thought

Infrastructure

Chips, machines, and where the model runs.

Closed model

A model available only through its owner's API or product. You may use it; you may not have it.

Also: Proprietary model

See also: Open weight

Cloud AI

Running models on somebody else's machines, reached over an API. Cheap to start, scales without effort, and means your input leaves your device.

Also: Hosted AI

See also: Local AI

Compute

Processing capacity, treated as a commodity: how many chips, for how long. The binding constraint on almost everything in AI.

See also: GPU, Scaling laws

Context caching

Reusing the processed form of a prompt across calls so the same long preamble is not paid for twice. A large cost saving for agents that carry the same standing instructions all day.

Also: Prompt caching

See also: Token

Data center

The building full of machines where training and most inference actually happen. Increasingly the limiting factor is power and cooling rather than chips.

Also: Datacenter

Edge AI

Running models on the device where the data is — a phone, a camera, a car — rather than sending it away to be processed.

See also: Local AI

GPU

The graphics processor that turned out to be the right shape for training and running neural networks. The unit in which AI capacity is counted and paid for.

Also: Graphics processing unit

See also: GPU rich / GPU poor, Compute

Local AI

Running a model on hardware you control. Usually chosen for privacy, offline use or cost rather than for capability. Whether your data really stays put depends on the whole setup, not just the model.

Also: On-device AI, Local model

See also: Desktop AI, Quantization, Open weight

Model routing

Sending each request to whichever model suits it — a cheap fast one for easy work, an expensive one for hard work. A large cost saving, and a source of confusing inconsistency when it is invisible.

See also: Latency

Open source

Software published under a license that lets anyone read, change and redistribute it. The term is used loosely in AI: publishing weights alone is more precisely called open weight, because studying and rebuilding a model also needs its code and information about its data.

See also: Open weight

Open weight

A model whose trained parameters are published, so anyone can download, run and fine-tune it. Not the same as open source: the weights are given, the training data and code usually are not.

Also: Open weights, Open-weight model

See also: Open source, Local AI

Rate limit

The cap a provider puts on how much you may call an API in a given period.

Serverless

Running code without managing the machine it runs on — you are billed per request rather than per server.

TPU

Google's own AI accelerator chip, and the main alternative to Nvidia's GPUs at scale.

Also: Tensor Processing Unit

See also: GPU, Nvidia

Model families

The models themselves, by family rather than by version.

Claude

Anthropic's language models, now in their fifth generation. Named by character rather than by number: Opus, Sonnet and Haiku across capability and speed, with Fable alongside them. Strongly associated with coding and long agentic work.

See also: Anthropic, Claude Code

Command

Cohere's models, built for enterprise retrieval and generation.

See also: Cohere

DeepSeek (model family)

DeepSeek's models, notable for reaching frontier-adjacent reasoning at a fraction of the expected training cost, and for being released with open weights.

See also: DeepSeek

ERNIE

Baidu's model family.

See also: Baidu

Falcon

TII's open-weight models, among the first state-backed releases to compete with the big labs.

See also: Technology Innovation Institute (TII)

FLUX

Black Forest Labs' image models, and one of the most widely used open-weight options for image generation.

See also: Black Forest Labs

Gemini

Google DeepMind's flagship models, now in their third generation. Natively multimodal, and known for very large context windows.

See also: Google DeepMind

Gemma

Google's open-weight models, built from the same research as Gemini but small enough to run locally.

See also: Google DeepMind, Local AI

GLM

Z.ai's open-weight bilingual family, widely used where Chinese and English performance both matter.

See also: Z.ai

GPT (model family)

OpenAI's flagship language models, now in their fifth generation. The name is also the architecture, which is why people say GPT when they mean any chatbot.

See also: OpenAI, GPT (Generative Pre-trained Transformer)

Granite

IBM's open-weight family, aimed at enterprise and regulated use.

See also: IBM

Grok

xAI's models, integrated with X and marketed on fewer refusals and real-time access to the platform's content.

See also: xAI

Hunyuan

Tencent's model family, spanning language, image and 3D generation.

See also: Tencent

Jamba

AI21's family, notable for combining transformer and state-space designs for long context.

See also: AI21 Labs

Kimi

Moonshot AI's models, known for very long context and strong agentic tool use.

See also: Moonshot AI

Llama

Meta's open-weight models. The release that made running a capable model on your own hardware normal, and the base most fine-tunes were built on for years.

See also: Meta AI, Open weight

Mistral (model family)

Mistral AI's models, including the Mixtral mixture-of-experts releases. Europe's main open-weight offering.

See also: Mistral AI, Mixture of Experts (MoE)

Nemotron

Nvidia's open-weight models, released partly to demonstrate what its hardware can do.

See also: Nvidia

Nova

Amazon's own model family, offered through AWS.

See also: Amazon

Phi

Microsoft's small language models, built to show how far careful data curation can take a model that fits on a laptop.

See also: Microsoft AI, Small Language Model (SLM)

Qwen

Alibaba's open-weight family, spanning many sizes and specialisms. The most widely fine-tuned base outside Llama.

See also: Alibaba Qwen, Open weight

Sora

OpenAI's video generation model.

See also: OpenAI

Stable Diffusion

The open-weight family of diffusion image models that made local image generation possible for anyone with a decent graphics card.

See also: Stability AI, Diffusion model

Veo

Google DeepMind's video generation model.

See also: Google DeepMind

Whisper

OpenAI's open-weight speech recognition family, widely used for transcription, translation and working out which language is being spoken.

See also: OpenAI

AI labs

Who builds them.

AI21 Labs

Israeli lab, maker of the Jamba models and long-standing work on language technology.

Also: AI21

See also: Jamba

Alibaba Qwen

Alibaba Cloud's model team, in China. Publishes a very broad open-weight family — dense, mixture-of-experts, vision and coding models across many sizes.

Also: Qwen team, Tongyi

See also: Qwen

Amazon

American company; a major AI infrastructure provider through AWS, an investor in Anthropic, and the maker of the Nova models.

Also: AWS

See also: Nova

Anthropic

American lab, founded 2021 by former OpenAI researchers. Makes the Claude model family and Claude Code. Known for safety research, Constitutional AI, and for publishing the Model Context Protocol as an open standard.

See also: Claude, MCP (Model Context Protocol)

Baidu

Chinese search company and AI lab, maker of the ERNIE models.

See also: ERNIE

Black Forest Labs

German lab founded by researchers from the original Stable Diffusion team. Makes the FLUX image models.

See also: FLUX

ByteDance

Chinese technology company behind TikTok, with a substantial AI research effort of its own.

Cohere

Canadian lab aimed squarely at enterprises. Known for embedding and reranking models and for retrieval-tuned generation.

See also: Embedding, RAG (Retrieval-Augmented Generation)

DeepSeek

Chinese lab known for open-weight reasoning and mixture-of-experts models, and for cost-efficiency claims that made the industry argue about how much compute a frontier model really needs.

See also: DeepSeek (model family), Open weight

ElevenLabs

Voice and audio lab, known for speech synthesis and voice cloning good enough to have changed how audio is produced.

Frontier lab

One of the small number of organizations training models at the leading edge. The label is about capability and compute, not size or age.

See also: Frontier model

Google DeepMind

Google's AI division, formed by merging DeepMind and Google Brain. Makes the Gemini model family, the open-weight Gemma models, and the Veo video models. The 2017 transformer paper came from Google researchers, in the then-separate Google Brain team.

Also: DeepMind

See also: Gemini, Transformer

Hugging Face

The main public hub for open-weight models and datasets, and the maintainer of libraries most of the ecosystem depends on. Effectively the field's shared warehouse.

See also: Open weight

IBM

American company with a long AI research history; publishes the open-weight Granite models for enterprise use.

See also: Granite

Meta AI

Meta's AI organization, in the United States, including its FAIR research arm. Makes the Llama family, and did more than anyone to establish open-weight models as a serious alternative to closed ones.

See also: Llama, Open weight

Microsoft AI

OpenAI's largest partner and infrastructure provider, and a lab in its own right — the Phi family of small models, and the Copilot products.

See also: Phi, GitHub Copilot

Midjourney

Independent American lab making image models of the same name, known for a distinctive aesthetic and for having no free tier and no API for most of its life.

MiniMax

Chinese lab known for efficient attention architectures, long context, and multimodal and video generation.

Mistral AI

French lab, founded 2023. Europe's most prominent frontier lab, known for efficient open-weight models alongside a commercial API.

See also: Mistral (model family), Open weight

Moonshot AI

Chinese lab behind the Kimi models, known for very long context and for open-weight agentic models.

See also: Kimi

Nous Research

American open-source AI collective, known for the Hermes model family and the Hermes Agent.

See also: Hermes Agent

Nvidia

American company, and not a model lab. Its GPUs run a large share of AI training and inference, and their availability has shaped the industry more than most research results. Google's TPUs and other accelerators are the alternatives.

See also: GPU, GPU rich / GPU poor

OpenAI

American lab, founded 2015. Makes the GPT model family, the ChatGPT product, the Sora video model and the Codex coding tools. The company that made this technology a consumer category.

See also: GPT (model family), ChatGPT

Perplexity

American company building an AI answer engine — search that responds with a cited answer rather than a list of links. Frequently cited as the model for what AEO has to optimize for.

See also: Answer Engine Optimization (AEO)

Stability AI

British company behind Stable Diffusion, the release that put open-weight image generation in everyone's hands.

See also: Stable Diffusion

Technology Innovation Institute (TII)

Abu Dhabi research institute, publisher of the open-weight Falcon models.

Also: TII

See also: Falcon

Tencent

Chinese technology company and AI lab, maker of the Hunyuan models.

See also: Hunyuan

xAI

American lab founded by Elon Musk in 2023. Makes the Grok model family and the GrokBot agent product, and is known for building very large training clusters very quickly.

See also: Grok, GrokBot

Z.ai

Chinese lab out of Tsinghua University, formerly Zhipu AI. Publishes the open-weight GLM family.

Also: Zhipu AI

See also: GLM

Tools

What people actually use day to day.

Antigravity

Google's agentic development platform, launched November 2025 and built as a fork of VS Code. Agents plan, execute and verify work across the editor, the terminal and a browser, and keep a knowledge base of what they learned. Runs Gemini models and also Claude and open models.

Also: Google Antigravity

See also: Coding agent, Google DeepMind

ChatGPT

OpenAI's assistant product. The application that took this technology mainstream, and still the name most people reach for.

See also: OpenAI

ChatGPT Work

OpenAI's agent for working across your apps, files and documents, producing spreadsheets, decks and documents rather than just answers.

See also: ChatGPT, OpenAI

Claude Code

Anthropic's agentic coding tool, run in a terminal, an app or an editor. It reads and edits a real codebase, runs commands, and works over many steps rather than answering one question.

See also: Anthropic, Coding agent

Claude Cowork

A tab inside the Claude desktop app where Claude works with your own files and folders, runs scheduled tasks, and hands back finished work — the same access as a coding agent without a terminal. Its Dispatch feature lets you trigger a desktop session from your phone.

Also: Cowork, Dispatch

See also: Claude Code, Anthropic, Desktop AI

Codex CLI

OpenAI's command-line coding agent.

Also: Codex

See also: OpenAI, Coding agent

Cursor

An AI-first code editor, built as a fork of VS Code. One of the first tools to make working alongside a model feel native rather than bolted on.

See also: Coding agent, Vibe coding

Desktop AI

AI that runs as an application on your computer, with access to your files and other applications, rather than living in a browser tab. The category lite.computer belongs to.

See also: Local AI, lite.computer

Gemini CLI

Google's command-line agent for its Gemini models.

See also: Gemini, Coding agent

GitHub Copilot

The autocomplete-in-your-editor product that introduced most developers to AI assistance, since grown into an agentic tool.

Also: Copilot

See also: Microsoft AI

GrokBot

xAI's persistent agent product, launched August 2026. Each agent runs on a cloud computer of its own and signs into your applications the way a person does, rather than through an API, and keeps working after you close your laptop. For a Showspace kept in a cloud folder that means work can carry on while your Mac is asleep — and, for the same reason, it cannot reach an app running only on your own machine.

Also: Grok Bot

See also: xAI, Grok, AI agent, Regeneration

Hermes Agent

Nous Research's open-source agent, released February 2026. It runs as a persistent process on your own infrastructure rather than in a session: it keeps memory across every conversation, runs scheduled tasks while you sleep, and writes its own reusable skills from experience. Works with any model provider, including local ones.

Also: Hermes

See also: AI agent, Memory (in AI), Skills (or tools)

OpenClaw

An open-source agent that runs on your own machine and is reached through the messaging apps you already use — WhatsApp, Telegram, Slack, Signal. It acts on your behalf across shell commands, browser automation, email, calendar and files. Previously called Clawdbot and Moltbot.

Also: Open Claw

See also: AI agent, Hermes Agent

Remote control (agents)

Continuing an agent session that is running on your own machine from a phone or a browser. The work, the files and the model access all stay on the machine; the phone is only the way in. Claude Code and Codex both offer it, and Claude Cowork's Dispatch is the same idea for people who never open a terminal.

See also: Claude Code, Claude Cowork, Desktop AI

Search and AEO

Being found, and being quoted correctly.

AI Overviews

Google's AI-written summaries at the top of search results. A large part of why zero-click search grew, and one of the surfaces AEO aims at.

Also: AI Mode

See also: Zero-click search, Answer Engine Optimization (AEO)

Answer engine

A search product that returns a written answer with citations instead of a page of links.

See also: Perplexity, Answer Engine Optimization (AEO)

Answer Engine Optimization (AEO)

Writing and structuring content so that AI answer engines quote it accurately and attribute it. Where SEO competes for a click, AEO competes to be the sentence the machine repeats — which means clear definitions, direct answers near the heading, and markup a machine can parse.

Also: AEO

See also: Generative Engine Optimization (GEO), Structured data, Search Engine Optimization (SEO)

Canonical URL

The address you declare as the real one for a page, so the same content at several addresses is not treated as several pages competing with each other.

See also: URL, Search Engine Optimization (SEO)

Crawler

A program that fetches pages automatically to index them. AI companies now run their own alongside the search engines',

Also: Bot, Spider

See also: llms.txt, robots.txt

Generative Engine Optimization (GEO)

A near-synonym for AEO, used for optimizing toward generative search results specifically. The two terms are still settling and are often used interchangeably.

Also: GEO

See also: Answer Engine Optimization (AEO)

JSON-LD

The format most structured data is written in — a block of JSON in the page describing what the page contains. The form search engines prefer.

See also: Structured data, Schema.org

Knowledge graph

A structured map of entities and the relationships between them. What lets a search engine know that a term, a company and a product are three different things that relate.

See also: Semantic Web, Structured data

llms.txt

A proposed convention, not a ratified standard: a Markdown file at a site's root summarizing what the site holds, for an agent that chooses to fetch it. It does not control crawling and does not replace the site's HTML.

See also: robots.txt, Crawler

robots.txt

A file at a site's root telling crawlers what they may fetch. Honored by convention, not enforced.

See also: Crawler, llms.txt

Schema.org

The shared vocabulary for structured data, maintained jointly by the major search companies. A glossary uses its DefinedTerm and DefinedTermSet types.

See also: Structured data, JSON-LD

Search Engine Optimization (SEO)

The practice of making a page more likely to be found and ranked by a search engine.

Also: SEO

See also: Answer Engine Optimization (AEO)

Structured data

Machine-readable markup added to a page so software knows what its parts mean — that this is a definition, that is a price, this is an author.

Also: Schema markup

See also: Schema.org, JSON-LD

Things you can build

The kinds of page people actually ask an AI for.

Artifact

Anything an AI produces that you keep and look at again — a page, a report, a deck, a dashboard. The word distinguishes lasting output from a chat reply that scrolls away.

See also: Self-contained page

Brand guide

The single page holding a brand's colors, type, tone and rules. Kept as a page in a Showspace, it is a thing every model you use can read before it writes, rather than a PDF nobody opens.

See also: Showspace, Style guide

Brief

A short document that gives someone what they need to act, and nothing else. A daily brief, a creative brief, a project brief.

See also: Report, Digest

Catalog

A product, property or inventory list rendered as pages, one per item. The gallery view earns its keep here, because a thumbnail of the item is the fastest way to find it.

Also: Catalogue, Inventory

See also: Record page

Changelog

A dated record of what changed, newest first.

Content calendar

A page laying out what is being published, when, and on which channel. The kind of thing that lives in a spreadsheet until someone wants to look at it on a wall.

Also: Editorial calendar

See also: Daily brief

Customer directory

Your customer list as a page each rather than rows in a spreadsheet, grouped into folders by segment and searchable by anything written on the page.

Also: Client directory

See also: Record page, Index page

Daily brief

One page rewritten each morning with what you need to know that day. The most common recurring artifact people set up, because the value is in it being there before you ask rather than in asking for it.

Also: Morning brief

See also: Brief, Living dashboard

Dashboard

A single screen showing the numbers or the state of something, laid out so you can take it in at a glance rather than read it. In lite.computer it is one HTML page that an agent rewrites as the underlying data changes.

See also: Artifact, Report

Deck

A presentation: a sequence of slides meant to be shown rather than read. Built as one HTML page, a deck advances a heading at a time in Presentation Mode without needing a slide format.

Also: Slide deck, Presentation

See also: Presentation Mode, Self-contained page

Digest

A gathering of many sources into one readable page — the week's news, a channel's messages, a folder's changes.

See also: Daily brief

Index page

The page that lists a set of record pages with a line about each and a link to it. What turns a folder of pages into something navigable rather than a pile.

Also: Directory page

See also: Record page, lite.html

Landing page

A single page built to make one thing happen — a sign-up, a purchase, a download. Judged by whether people do that thing.

See also: Website

Live sign

A page put on a display and left there — a menu board, a notice, a status wall — that updates itself because the agent behind it keeps writing the file.

Also: Digital signage

See also: Living dashboard, Presentation Mode

Living dashboard

A dashboard that is rewritten on a schedule — usually by a small script reading the numbers underneath, sometimes by an agent — so the page on your screen keeps up instead of being a snapshot of the day it was made. The agent is what builds and changes it; a script is what refreshes it.

See also: Dashboard, Living website

Long document

A report, a contract, a piece of research, laid out to actually be read on a screen — one idea per section, room to breathe — rather than a wall of text in a word processor.

One-pager

Everything about one subject on a single screen or sheet. The discipline is the constraint, not the format.

Progress report

A page saying what shipped, what is stuck, and what decisions are needed. Made worth keeping by being answerable: you leave comments on it, and next session the agent reads them and revises the same page rather than starting a new one.

Also: Status report

See also: Report, Comments (in lite.computer)

Record page

One page for one thing — a customer, a property, a product, a member. Generated in bulk from a spreadsheet or a database export, and regenerated from it when the data moves, so a folder of them becomes a directory you can browse, search and hand to someone, with no database behind it.

See also: Index page, Customer directory, CSV, Regeneration

Regeneration

Rebuilding a page from its source data rather than editing it. The source — a spreadsheet, an export, an API — stays the thing you change; the page is only the current view of it. It is what separates a living page from a document that goes stale.

Also: Rebuild

See also: Living dashboard, Record page, Scheduled job

Report

A written account of what happened over some period, usually with a conclusion. Distinct from a dashboard: a report is read once and argues something; a dashboard is glanced at repeatedly and states.

See also: Dashboard, Brief

Runbook

Step-by-step instructions for doing a specific operational job, written to be followed under pressure by someone who is not the author.

Also: Playbook

Style guide

The rules for how something should look and sound — colors, type, spacing, tone. Increasingly written for agents to read as much as for people.

See also: Brand guide

Timeline

Events laid out in time order, so the shape of a sequence is visible at once.

Website

A set of linked pages served over the internet at a domain. A folder of HTML pages becomes a website when it is hosted; until then it is a folder of HTML pages, which is what lite.computer reads.

Also: Site

See also: Host, Static site, Domain

Website concept

A proposed site staged as real, clickable pages rather than drawn as flat mockups. The reviewer moves through it the way a visitor would, which surfaces problems a picture of a page cannot.

Also: Site concept, Prototype

See also: Website, Deck

Wiki

A body of linked pages anyone can add to, organized by its links rather than by a hierarchy.

See also: Knowledge base, lite.html

Web and computing

The words underneath a page, a folder and a site.

App

A program you install and run on your own device, as opposed to a page you visit. The difference that matters here is reach: an app can see your files and your other software; a web page cannot.

Also: Application, Native app

See also: Desktop AI, Browser

Browser

The program that fetches web pages and draws them — Chrome, Safari, Firefox, Edge. Because HTML pages are what browsers read, a page written for lite.computer also opens in one, with nothing installed.

See also: HTML, Website

CLI

Command-Line Interface: driving a program by typing commands rather than clicking. Most agentic coding tools are CLIs, because a terminal is where the files, the version control and the build already are.

Also: Command line, Command-line interface

See also: Terminal, Coding agent

CSS

Cascading Style Sheets: the rules that decide how HTML looks — color, type, spacing, layout.

Also: Cascading Style Sheets

See also: HTML

CSV

Comma-separated values: the plain-text table format nearly every system will export to. Usually the bridge between a database or spreadsheet and a set of pages made from it.

Also: Comma-separated values

See also: Record page, Spreadsheet

Data URI

A way of embedding a file — usually an image — directly inside a page as text, so the page needs no separate files beside it.

Also: Base64 image

See also: Self-contained page

Database export

A snapshot of what a database holds, written out as a file you can hand to something else. The starting point for turning records into pages, and — re-exported on a schedule — the thing that keeps them current.

Also: Data export

See also: CSV, Record page

DNS

The Domain Name System: the lookup that turns a domain name into the address of the machine serving it.

Also: Domain Name System

See also: Domain

Domain

The name you rent and point at a host, so people can reach your site by a word rather than a number. lite.computer is one.

Also: Domain name

See also: DNS, URL

File path

The address of a file on a disk. What you hand an agent so it writes in the right place, and the most common thing to get wrong when pages stop appearing.

Also: Path

See also: Showspace

Git

The version control system nearly all software uses. It records every change, so any past state can be recovered and two people can work on the same files without overwriting each other.

See also: Repository

Host

A computer that serves something to others over a network. To host a site is to put its files on such a machine so anyone with the address can fetch them. Hosting is what turns a folder of pages into a website.

Also: Hosting

See also: Website, Server, Localhost

HTML

HyperText Markup Language, the format web pages are written in. Tags wrap content to say what it is — a heading, a paragraph, a link — and a browser decides how to draw it. It is a text file, readable without any special software, which is why a page written today still opens in twenty years.

Also: HyperText Markup Language

See also: Markup, Self-contained page

JavaScript

The programming language that runs inside a web page and makes it do things after it has loaded.

Also: JS

See also: HTML

JSON

A plain-text format for structured data that both people and programs can read. What most APIs speak.

See also: JSON-LD, Structured data

Localhost

Your own machine, addressed as though it were on the network — usually at an address beginning http://localhost. A page served at localhost is visible only to you, which makes it the normal way to look at a site before anyone else can.

Also: 127.0.0.1

See also: Server, Host

Markdown

A lightweight markup format designed to be readable as plain text — a hash for a heading, asterisks for emphasis. The default output of most AI assistants, and the format an Obsidian vault is written in.

See also: Markup, Obsidian vault, HTML

Markup

Text with tags around it saying what the parts are, rather than how they look. HTML and Markdown are both markup. The point is that a machine can tell a heading from a paragraph without guessing.

See also: HTML, Markdown, Structured data

Repository

A project's files together with the full history of every change to them.

Also: Repo

See also: Git

Responsive

A page that lays itself out sensibly at any width, so the same file reads well on a phone and on a monitor.

See also: HTML

Sandbox

A restricted environment where code runs without being able to reach anything outside it. Pages in lite.computer render in one, which is why a page cannot read your files or your other pages.

See also: Self-contained page

Scheduled job

A script set to run on its own at a fixed time — every hour, every morning — without anyone starting it. On a Mac the scheduler is built in. Pairing one with a generator is how a page stays current with nobody remembering to refresh it.

Also: Cron, Cron job

See also: Regeneration, Living dashboard

Server

A program, or the machine running it, that answers requests over a network. A web server hands out pages when a browser asks for them.

See also: Host, Localhost

Spreadsheet

Rows and columns in a file. Excellent for storing and calculating, poor for reading, which is why turning one into pages is worth doing at all.

See also: CSV, Record page

Static site

A site made of files that are served exactly as they are, with no application generating pages on request. Fast, cheap, hard to break, and what a folder of self-contained HTML pages becomes when hosted.

See also: Website, Host, Self-contained page

Terminal

The window you type commands into. The CLI is the interface; the terminal is where it appears.

Also: Shell, Console

See also: CLI

URL

The full address of something on the web, including which protocol to use to fetch it.

Also: Link, Web address

See also: Domain

Web app

Software delivered through a browser rather than installed. Convenient to distribute, and fenced off from the machine it runs on.

Also: SaaS

See also: App, Browser

Slang and culture

What practitioners say to each other.

Accelerationist (e/acc)

Someone who believes AI progress should be pushed as fast as possible to unlock radical abundance. The other half of the divide, and the origin of the e/acc tag.

Also: e/acc, Effective accelerationism

See also: Doomer

Agent washing

Marketing an ordinary automation or chatbot as an agent because the word sells.

See also: AI agent, Agentic

Attention Is All You Need

The 2017 Google paper that introduced the transformer. The single most cited origin point for everything since.

See also: Transformer, Attention

Context rot

The gradual loss of focus as a model's context window fills with too many turns, tool results and irrelevant documents, until it loses the plot and starts missing explicit instructions.

See also: Context window, Context engineering

Doomer

Someone whose central concern about AI is catastrophic risk — alignment failure, existential threat, runaway frontier models. One half of the field's main ideological divide.

See also: Accelerationist (e/acc), Alignment

GPU rich / GPU poor

The divide in the industry by access to compute. Thousands of H100s or clusters on demand makes you GPU rich. Queuing on rental clouds and asking for credits makes you GPU poor.

Also: GPU rich, GPU poor

See also: GPU, Compute

Hallucination laundering

When an AI invents something plausible, a person cites it in a blog post or a draft, and a later scrape pulls that error into training data — a hallucination promoted to ground truth by being repeated.

See also: Hallucination

Human slop

The retort to slop: the observation that most of the low-quality content on the internet predates AI and was written by people.

See also: Slop

Lazy model

A model that does part of the job and stops — leaving a comment where the code should be, summarizing three of ten files, or telling you to finish the rest yourself.

See also: Nerfed

Loop engineering

Designing the loop an agent runs inside — what it does between tool calls, when it checks its own work, and how it decides it is finished — rather than hand-writing each prompt. The 2026 successor to prompt engineering, and where the difference between two products on the same model now shows up.

See also: Prompt engineering, Context engineering, Harness, Agentic

Moat

Whatever stops a competitor copying you once the underlying model is available to everyone. The question every AI product has to answer.

Nerfed

The complaint that a model has quietly got worse — slower, lazier, more likely to refuse — after a provider update.

Also: Lobotomized

See also: Drift

Prompt whisperer

Half-joking name for someone unusually good at getting what they want out of a model.

See also: Prompt engineering

RLHF'd to death

The complaint that a model has been over-aligned — by reinforcement learning from human feedback or any of the methods beside it — until it turns cautious, sterile, sycophantic, or refuses harmless prompts.

See also: RLHF (Reinforcement Learning from Human Feedback), Sycophancy, Refusal

Rug pull

When a provider changes pricing, limits or model behavior under a product built on top of it, and the product breaks through no fault of its own.

Ship it

Release it now. In AI work, often the point of tension with evals, which exist to say not yet.

See also: Evals

Slop

Low-effort, mass-produced AI content. Used pejoratively for spammy synthetic articles, filler posts, soulless generative stock art, and auto-generated sites built to catch ad traffic.

Also: AI slop

See also: Human slop

Spicy autocomplete

A tongue-in-cheek, deflationary term for large language models. A reminder that beneath the apparent reasoning, the system is predicting the next token from statistical probabilities.

See also: Large Language Model (LLM), Stochastic parrot

Stochastic parrot

A model that reproduces the statistical patterns of language without understanding it. From a 2021 paper, and still the sharpest phrase the skeptical side has.

See also: Spicy autocomplete

The bitter lesson

Rich Sutton's observation that methods which simply use more computation eventually beat methods built on human insight about the problem. Quoted constantly, usually to end an argument.

See also: Scaling laws

Token burn

The rate at which an agent spends tokens, and the money that goes with it. Long autonomous runs make this a real line item.

Also: Burn rate

See also: Token

Token discipline

The correction to tokenmaxxing: spending tokens where they earn something and measuring the result rather than the spend.

See also: Tokenmaxxing, Evals

Tokenmaxxing

Treating token consumption as a measure of productivity — the more your agents burn, the harder you are assumed to be working. The suffix comes from internet slang for pushing one number to an extreme regardless of the outcome. Companies that ranked staff this way in 2026 mostly stopped: token volume measures cost and effort, not results, and it is trivially gamed.

Also: Token maxing

See also: Token burn, Token, Evals

Vibe coding

Programming by steering an AI assistant with prompts rather than writing syntax by hand. You describe the intent, accept the diff, check whether it runs, and tweak the vibes until it works.

See also: Coding agent, Vibes

Vibes

Judging a model or an output by how it feels rather than by measurement. Half the time it is the honest answer, and the other half it is what evals exist to replace.

Also: Vibe check

See also: Evals

Wiggly

Practitioner shorthand for non-deterministic output. Traditional software is rigid and predictable; language models are wiggly, or smushy, because the same input can yield subtle variations each time.

Also: Smushy

See also: Deterministic, Probabilistic, Temperature

Wrapper

A dismissal — "it's just a GPT wrapper" — for a product that is a thin interface over someone else's model. Sometimes fair, and sometimes said about the part that was actually hard.

Also: GPT wrapper

See also: Scaffolding

Yolo mode

Letting an agent run without approving each action. Fast, occasionally spectacular, and the reason approval prompts exist.

Also: Auto-approve

See also: Human in the loop

lite.computer

Our own vocabulary, defined once.

Agent Helper

The panel in the app that hands your agent what it needs: the real path of your Showspace, and a set of copyable prompts, the first of which turns your Showspace into your agent's default place to write.

See also: lite.computer, MCP server

Canvas

The display area to the right of the sidebar, where the page or the gallery renders. Settled as the house word in August 2026, replacing display area, main area and viewport.

See also: Chrome, Sidebar

Chrome

The sidebar and top bar together: everything that is not canvas. Presentation Mode and the Presenter Frame both remove it.

See also: Canvas, Presenter Frame

Comments (in lite.computer)

Notes attached to a page without changing the page, aimed either at a heading or at the whole page. They are stored in .lite/comments.json, which is why an agent can read them and act on them.

See also: The .lite folder

Daylight

The design methodology and infrastructure Veue Management Corp uses for agent-built interfaces. lite.computer is its first trial project.

See also: Living website

Engraving

The buyer's name, signed into the license itself and shown in the sidebar. It cannot be typed, which is the point: the paid copy carries your name and the free trial carries none.

See also: lite.computer

Explode

One window per open tab, laid out across the displays you have. Every tab but the active one moves into a window of its own, and the window you started from keeps that one and becomes one of them. It acts on tabs, never on the folder: a Showspace of two hundred pages is not an instruction to open two hundred windows. ⌃⌘E.

See also: Implode, Tear off, Tony Stark effect

Folder (in a Showspace)

A filter inside a Showspace. Clicking one narrows the canvas to its contents. It is never an identity, never branded and never watched separately.

See also: Showspace, Scope

Four-scenario walk

A design check used on every lite.computer decision: render the question through one file, hundreds of files, one narrow shared screen, and several full-screen windows across displays. If any of the four fails, fix it before going on.

See also: Explode

Implode

The undo for Explode. Gathers every other window's tabs back into one window, closing them behind their last tab. ⌃⌘I.

See also: Explode

lite.computer

A macOS app that turns a folder of HTML pages into a browsable library — a file list, a gallery of rendered thumbnails, and a reader — and keeps up as an AI agent writes into that folder. Written lite.computer; said Lite.

Also: Lite

See also: Showspace, Desktop AI

lite.html

A hand-kept page at the root of a Showspace that says what is worth reading and why. A map of content rather than a file listing, usually written and updated by your agent.

See also: Showspace

Living website

A site that is regenerated by agents as its underlying material changes, rather than edited by hand between redesigns. The concept behind the Daylight work that lite.computer's methodology came out of.

See also: Daylight

Page (in lite.computer)

One HTML file. What a tab holds and the reader renders.

See also: Showspace

Presentation Mode

The page full screen with no chrome at all. A click or an arrow key steps to the next heading, one window at a time. ⌘⇧Return.

Also: Presenting

See also: Presenter Frame, Chrome

Presenter Frame

The same stripped-down reading surface as Presentation Mode, but the window stays where it is rather than taking the whole display. For a shared screen on a call, or a second monitor. ⌘⇧F.

See also: Presentation Mode, Chrome

Scope

Which subset of a Showspace the canvas is showing — the whole thing, or one folder. It has exactly two states, and Show all is the way back out of the second.

See also: Folder (in a Showspace)

Self-contained page

An HTML file that carries everything it needs inside it — styles inline, images embedded, system fonts, no build step and no server. The format lite.computer is built around, because such a page still opens in ten years and on any machine.

Also: Single-file HTML

See also: Page (in lite.computer)

Showspace

One root folder plus everything the app remembers about it: its index, its brand, its open tabs, and its health. The unit you open, switch between and brand. A Showspace is an identity; a folder inside one is only a filter.

See also: Folder (in a Showspace), lite.computer

Showspace health

The app's report on what is broken in a Showspace — links pointing nowhere, images that do not resolve, how full the index is, and how much of the folder it was able to read. The same report an agent gets when it asks.

Also: Diagnostics

See also: MCP server

Spatial anchor

The principle that people navigate by remembering where they put something, not by searching again. lite.computer's folder tree does not reorganize itself, so the order you built is the order you navigate by. Machine-scale views are layers on top of that, never replacements for it.

See also: Showspace

Tear off

Dragging a single tab out of the strip so it becomes its own window where you drop it. Explode is this done automatically to every tab at once.

See also: Explode

The .lite folder

The single hidden folder the app adds to a Showspace. It holds the app's notes about that Showspace and the comments you leave on pages. Your own pages are never altered.

Also: .lite

See also: Comments (in lite.computer)

Tony Stark effect

The internal name for the multi-window Explode experience: several windows open at once, each full-screen on its own display, all watching the same Showspace and all updating live as the AI writes. A desk of monitors becomes a wall of live dashboards. Called Explode on public surfaces.

See also: Explode

Still stuck? Write to support@lite.computer. A person reads it.