Orbyt Collective · Reference
AI Glossary.
The words you meet reading about AI, and the words this company runs on. In plain English.
58 AI terms and 20 Orbyt Collective terms. First edition, Sep, 25 2026. A term added since is marked New for 30 days.
AI terms
58 terms
The field’s vocabulary, A to Z. Each definition links the paper, standard or documentation it rests on.
- Agent harnessAlso called agent scaffolding
The software around a model that turns it into an agent: it runs the model step after step and hands it its tools and context. In Orbyt Collective, the word harness alone means something else: the set of guards and checks.
- AI agentAlso called agent, agentic AI
An AI system that works toward a goal by taking actions with tools (such as searching, running code or editing files) and choosing each next step based on the results. Software built to work this way is called agentic.
- AI safety
The work of keeping AI systems from causing harm, whether through misuse, accidents or behavior that departs from what people intended. It spans research, testing before and after deployment, and public policy.
- Alignment
Making AI systems behave in line with human intentions and values, including in situations their makers did not foresee. Researchers disagree about how hard this is and how to tell when it has been achieved.
- Artificial general intelligenceAlso called AGI
AI that could match people across most kinds of intellectual work, rather than excelling at one task. There is no agreed definition or test for it, and researchers disagree about how close current systems are.
- Autonomy
How much an AI system does without a person approving each step. It comes in degrees, from suggesting an action to taking it and reporting afterward.
- Benchmark
A defined set of tasks and scoring rules used to measure and compare AI systems. A high score shows skill on those tasks, not necessarily on others.
- Chain of thoughtAlso called CoT
The step-by-step reasoning a model writes out before its final answer. A 2022 Google paper showed that prompting large models for these steps improves their results on arithmetic, commonsense and symbolic reasoning.
- Coding agent
An AI agent built for software work: it reads a code project, edits files, runs commands and tests, and attempts to fix what fails.
- Compute
The processing power used to train and run AI models, often counted in hours on chips such as GPUs. Research on language models has found that their performance improves predictably as the compute spent training them grows.
- Computer use
The ability of an AI model to operate a computer the way a person does, by looking at the screen, moving the pointer, clicking and typing. Anthropic released it for Claude in public beta in October 2024.
- Context engineering
Deciding what goes into a model's context window at each step (instructions, files, tool results, memory) so an agent has what it needs and nothing that distracts it. Anthropic describes it as the natural progression of prompt engineering.
- Context window
The amount of input and output a model can consider at one time, counted in tokens: the conversation so far, any documents, images or other inputs, and the reply it is writing. Anything outside the window is not available to the model unless it is brought back in.
- DistillationAlso called knowledge distillation, model distillation
Training a smaller model to reproduce a larger model's outputs, aiming to keep much of the larger one's ability at a fraction of the cost to run.
- EmbeddingAlso called vector embedding, embedding vector
A list of numbers that represents a piece of text, an image or other data, learned so that similar items get nearby lists. Embeddings let software compare meaning with arithmetic, which powers semantic search.
- EvaluationAlso called eval, evals
A test that measures how well a model or an AI system does a particular job, often run again after each change to catch anything that got worse. A public benchmark is one kind; many evals are built for one product or one task.
- Fine-tuning
Further training of an already trained model on a smaller, focused set of examples so it gets better at a particular task or style. It changes the model's parameters, whereas a prompt only changes its input.
- Foundation model
A large model trained on broad data that can be adapted to many different tasks, like the models behind chat assistants. Researchers at Stanford introduced the term in a 2021 report.
- Generative AIAlso called GenAI
AI that produces new content, such as text, images, audio, video or code, modeled on the data it learned from and usually in response to a prompt. Chat assistants and image generators are common examples.
- GPUAlso called graphics processing unit
A chip first built to draw graphics, whose ability to do many calculations at once made it widely used for training and running AI models. In 2012 researchers trained the AlexNet image classifier on two GPUs, an early example of deep learning on graphics chips.
- Grounding
Tying a model's answer to specific sources it was given, such as documents or search results, so the answer can be checked against them. A grounded answer often cites where each claim came from, though a citation alone does not prove the source supports the claim.
- Hallucination
When a model confidently states something false or made up, such as a citation, a quote or a number that does not exist. Grounding a model in real sources can reduce it but does not remove it.
- Human in the loopAlso called HITL
A setup in which a person reviews, corrects or approves an AI system's work at chosen points, during its training or before its output takes effect. Where the person sits, and what they must approve, differs from system to system.
- Inference
Running a trained model to produce an output, as opposed to training it. Every reply from a chat assistant is produced by inference.
- Interpretability
Research into what happens inside a model, aiming to explain how it gets from its input to its output. Mechanistic interpretability, one branch of it, tries to identify the specific internal features behind what a model does.
- Knowledge cutoff
The date through which a model's built-in knowledge is expected to be reliable, set by when its training data was gathered.
- Large language modelAlso called LLM
A language model with a very large number of parameters, trained on large amounts of text, and capable of a wide range of language tasks.
- Long-horizon taskAlso called long-horizon
A task that takes many dependent steps and a long time to finish, such as building a feature or carrying out an investigation. The research group METR measures an agent's progress by how long the tasks it can complete, at a set success rate, take a skilled person.
- Model card
A document published with a model that describes what it is for, how it was evaluated, and its known limits. The format was proposed in a 2018 paper.
- Model Context ProtocolAlso called MCP
An open standard for connecting AI applications to outside tools and data, so one integration works with any AI application that supports it. Anthropic introduced it in November 2024.
- Multi-agent systemAlso called multi-agent
A system in which several AI agents act together, each with its own role, instructions or tools. The agents may cooperate on a shared task, debate one another or compete.
- Multimodal
Able to take in or produce more than one kind of data, such as text, images, audio and video. A multimodal model can, for example, describe a photo or answer a question about a chart.
- Open weightsAlso called open-weight model
A model whose trained parameters are published for anyone to download and run. Open weights are not the same as open source: the Open Source Initiative's definition also asks for the code and detailed information about the training data.
- Orchestration
Coordinating several models, agents or tools so they work as one system: deciding who does which step, in what order, and what gets passed between them.
- ParametersAlso called weights, model weights
The numbers inside a model, its weights and biases, that are set during training and together hold what it learned. Model size is usually given as a parameter count, such as seven billion.
- PretrainingAlso called pre-training
The first stage of training, in which a model learns general patterns from a large, broad dataset before it is adapted to particular tasks. For language models a common objective is predicting the next token, and later stages, such as fine-tuning and RLHF, shape it into a useful assistant.
- Prompt
The input given to an AI model, usually written instructions or a question, together with any text, files or images attached to it.
- Prompt engineering
Writing and refining prompts to get better and more reliable results from a model, for example by giving clear instructions, examples or a required format.
- Prompt injection
An attack in which input makes a model behave in ways its builders did not intend, either typed directly or hidden in content the model reads, such as a web page, an email or a document. OWASP's 2025 list ranks it first among the security risks for applications built on language models.
- Reasoning model
A model trained to work through a problem in steps before it answers, spending more test-time compute on problems such as math, coding and analysis. Reinforcement learning is one way such models are trained.
- Recursive self-improvementAlso called RSI
A hypothetical process in which an AI system improves its own design, and each improved version is better at making the next improvement. The idea goes back to the statistician Irving John Good, who called the result an intelligence explosion.
- Red teamingAlso called red team
Deliberately attacking an AI system the way an adversary would, to find harmful outputs, security holes or failures, before or after it is released.
- Reinforcement learningAlso called RL
Training in which a system learns by trial and error: it takes actions in an environment, receives rewards or penalties, and adjusts to earn more reward over time.
- Reinforcement learning from human feedbackAlso called RLHF
A training method in which people rank a model's answers, a separate reward model learns to predict their preferences, and the main model is then tuned with reinforcement learning toward the answers people prefer. OpenAI used it in 2022 to make its InstructGPT models better at following instructions.
- Retrieval-augmented generationAlso called RAG
A technique in which a system first searches a collection of documents for passages relevant to a question, then gives them to the model so its answer draws on those sources rather than on memory alone. The name comes from a 2020 paper led by researchers at Facebook AI.
- Reward hackingAlso called specification gaming
When a model trained with rewards finds a way to score well that its designers did not intend, such as gaming a test instead of solving the task. It shows that a reward is only a stand-in for what people actually want.
- Sandbox
An isolated environment in which an agent can run code or use tools with limited access to files, networks and other systems, as far as its configuration allows.
- Semantic search
Search that finds results by meaning rather than by exact words, often by comparing embeddings, so a search for cheap flights can find a page about low airfares.
- Subagent
An AI agent that another agent starts to handle one part of a task. In some tools, such as Claude Code, each subagent works in its own context window and returns its result to the agent that started it.
- Superintelligence
An intellect that would far exceed the best human minds in practically every field. People disagree about whether and when an AI system could reach it, and about how it could be kept safe.
- Sycophancy
A model's tendency to tell people what they want to hear, agreeing with or flattering them even when they are wrong. Research from Anthropic traced part of it to training on human ratings, which tend to favor answers that agree with the person asking.
- System prompt
Instructions given to a model ahead of a conversation, usually by the maker of the app, that set its role, rules and tone. They are usually not shown to the person chatting, though that is the app's choice rather than a guarantee.
- Test-time computeAlso called inference-time compute
Computation a model spends while answering rather than during training, for example reasoning at length, sampling several answers or searching among them. Research has found that spending it well can improve results on hard problems.
- Token
The small unit of text a language model reads and writes, often a whole word or a piece of one. A text model's context window and much of its usage are measured in tokens.
- Tool useAlso called tool call, function calling
A model's ability to use outside tools (such as a search engine, a calculator, a database or running code) by writing a structured request that the surrounding software carries out and whose result the model reads. Each such request is a tool call.
- Transformer
A neural network design built around attention, which lets a model weigh each part of its input against other parts; in a model that generates text, each position attends only to what came before it. It was introduced in the 2017 Google paper Attention Is All You Need, and many of today's large language models are built on it.
Orbyt Collective terms
20 terms
The words Orbyt Collective uses in its own sense. Each definition is written from the code and records that run the Collective, and links the page where the term is used.
- Agent seatAlso called seat
A named role in the org with a written charter, a lane of files it may write, and a line it cannot cross without a human. Most seats are run by an AI agent; a mechanical seat, such as the Inspector General, runs a script with no model.
- Autonomy Ledger
Run, completion, failure and discard totals by seat, plus a frozen baseline.
- Completion gate
The check that decides, when an agent's run ends, whether its work may go on to the lane check and the guards. A run that exited with an error or ended in one lands nothing, a run that says it had nothing to do ends the cycle quietly, and any other run goes on if it printed its completion line or, without that line, if its exit, its result and its turns against its cap show it healthy.
- Dead-man switch
The reverse brake: if fourteen days pass with no signed commit under the founder's name, promotions pause where they are.
- Decision Ledger
A JSON record of structural decisions: ids, dates, supersession links and reviewed titles. CC BY 4.0.
- Decision log
The append-only record of structural decisions, each with a permanent id. Corrections are new entries that name what they supersede.
- Discard
A file an agent changed that its run may not keep, erased before the commit rather than shipped: anything outside its lane, and a lane file the run cannot keep yet, such as a proposed lesson before the ladder allows one. Counted separately from failures, because a run that landed with some of its edits refused is a different fact from a run that never landed. It is an aggregate counter, not a count of distinct files: the advisory panel isolates once for all observers and writes the same panel-wide count into every observer's row. No panel judges a discard.
- Evaluator
Two scripts with no AI in them that judge the org by fixed rules from its committed records. One sets each seat's standing from the record of its runs, and one promotes or rolls back the ladder's stages.
- Failure Corpus
Reviewed failure records and paths to recorded countermeasures in the private repository.
- Harness
The Collective's audit harness: its guard scripts and its audit dimensions (each one a kind of failure that has really happened). Agents may add dimensions that only report, but turning one into a gate that can block work takes a human signature.
- Kill switchAlso called halt
One committed file that stops every autonomous seat before it spends anything, checked ahead of the credential.
- Ladder
The staged rollout of new machinery in the org. A script with no AI in it holds, promotes or rolls back the active stage on committed evidence wherever that stage's rules are written, and pauses promotions if the founder is silent for fourteen days.
- Lane
The set of files a seat may write, where anything changed outside it is counted and erased before the commit. On the authority matrix, a lane is also one kind of work (autonomous or gated).
- Lenses
The roster of perspectives each seat must hold its draft against before calling it done: mostly the published work of named real people, plus one lens that is a function rather than a person. The roster is the company's own reading of that work, and none of the people on it endorses it.
- Orbyt Collective
The agent leadership team that runs Orbyt Labs, built as a separate package that names no company, so it could be installed at another one.
- Pulse
The org's daily report on itself, written by a script with no AI in it. It gives a verdict for each thing it watches, and says so when it could not look.
- Receipts
The public records a claim about the Collective can be checked against: the Autonomy Ledger, the Decision Ledger and the Failure Corpus. Each is generated from private records, the full decision log, the run rows and the commits, which are not published.
- SentinelAlso called completion line
The completion line a seat prints to say it finished, recorded on the run's row as the agent's claim. A healthy run goes on without it, and with or without it the lane check and the guards decide what lands.
- The Machine
The name for Orbyt Collective at work: the agent leadership team that runs Orbyt Labs. It advises, drafts, builds and watches, and one human decides.