Glossary

A2A (Agent2Agent)

Agent2Agent, a protocol for agents to discover and delegate tasks to other agents; it joined the Agentic AI Foundation on August 17, 2026. See: Why MCP Spread

Absorptive capacity

An organization's ability to recognize, take in and use new knowledge; it depends on the knowledge it already has. See: What a Deployment Engineer Should Own, and Leave Behind

Abstention

A model declining to answer ("I don't know", "not in the documents") instead of guessing. See: Confidently Wrong

Active parameters

The parameters a mixture-of-experts model actually uses to process one token. They set the compute per token, while the total parameter count sets the memory needed. See: Open Weights and Closed Labs

Adenoma detection rate

The share of colonoscopies in which the doctor finds at least one adenoma, a precancerous growth; a standard quality measure for colonoscopy. See: Skills That Fade When the Machine Does Them

Agent

A system in which a model decides the next step in a loop, choosing tools, reading their results and stopping when it judges the task done, inside limits set by code. See: Workflows or Agents?

Agent assist

AI that helps a human support agent (suggesting replies, finding answers) rather than answering customers directly. See: Customer Support Went First

Agent skill

A packaged set of instructions, scripts or reference files that an agent loads only when a task needs it; a short description is visible first, the full content on demand. See: Write for the Model That Doesn't Exist Yet

Agentic AI Foundation

A Linux Foundation project, announced December 9, 2025, that hosts MCP, AGENTS.md, goose and later A2A, among others, as a neutral home for agent standards. See: Why MCP Spread

Aggregation theory

Ben Thompson's account of how value moves to whoever owns the user relationship when distribution costs almost nothing. See: Distribution Beats the Better Model

AI literacy

The skills and understanding people need to use AI systems knowledgeably, including their limits and risks. The EU AI Act requires providers and deployers to take measures to support it among their staff. See: Regulation as a Design Constraint

AI Overview

Google's AI-written summary shown above the regular results for some searches, with a few source links. See: When Answers Replace Links

Answer engine

A search product that writes a direct answer, usually with a few cited sources, instead of only returning a ranked list of links. See: When Answers Replace Links

Automation bias

People supervising an automated aid tend to over-rely on it, missing both its errors and problems it fails to flag, even when they are experienced. See: From a Good Demo to a Working Department, Skills That Fade When the Machine Does Them

Bank reconciliation

Matching a company's cash records against its bank statement and explaining every difference, so the adjusted book balance equals the adjusted bank balance to the cent. See: Finance and the Reconciliation Wall

Base model

A model after pretraining only. It continues text plausibly but doesn't reliably follow instructions or hold a conversation until post-training. See: What Training Actually Changes

Baseline

A measurement of how the work performed before the change (volume, handling time, error rate, cost), taken so later results can be compared with something other than memory. See: From a Good Demo to a Working Department

Batch processing

Submitting requests to be processed asynchronously, usually at a discount, when the answer can wait. See: Why AI Products Have Thinner Margins

Bradley-Terry model

A statistical model that estimates a strength for each competitor from who beat whom in pairwise comparisons; Arena uses it to turn votes into scores. See: Leaderboards Measure the Test

Bundling

Selling a new feature inside an existing product or plan, so its marginal price to the customer is zero or a small increase. See: Distribution Beats the Better Model

Business process outsourcing (BPO)

Paying an outside firm to run a repeatable business function (claims handling, customer support, document processing), usually priced per unit or per person, abbreviated BPO. See: Where the AI Budget Comes From

Calibration

How well confidence matches accuracy. For a person using AI, it means knowing where the tool is reliable on their own tasks. See: Learn to Drive Before You Study the Engine, Confidently Wrong

Canary release

applying a change to a small, time-limited slice of traffic or users and evaluating it before applying it to everyone, so a defect's damage is proportional to the slice. See: Reversible Work and Irreversible Effects, Moving Without Breaking

Click-through rate

The share of times a link is shown (an impression) that it is clicked. See: When Answers Replace Links

Code coverage

the share of lines or branches executed while tests run; it shows what the tests touched, not whether they would catch a mistake. See: Tests Are the New Specification

Cognitive load, team

The amount of complexity a team must hold in mind to own its part of a system; when it exceeds what the team can handle, decisions slow and quality falls. See: How AI-Native Teams Organize

Collective tacit knowledge

Knowledge that belongs to a group rather than any member, such as fluent language use or judgment in traffic; Collins argues it can't be made explicit even in principle and is acquired by immersion. See: Most of What a Company Knows Isn't Written Down

Compaction

Summarizing or trimming earlier parts of a long session so it fits in the context window. Saves room and can lose details that matter later. See: What Fits Versus What Gets Used, Agents That Work for Hours

Compensating action

a new action that offsets the consequence of an earlier one (a refund, a correcting entry) without erasing it; both remain on the record. (Related: saga.). See: Reversible Work and Irreversible Effects, When the Tool Call Fails Halfway

Complete mediation

A security principle in which every request to a protected resource is checked against policy at the resource, instead of trusting the caller's judgment; for agents, the tool layer validates each action rather than trusting the model. See: Giving an Agent Room to Work

Confabulation

In Farquhar et al.'s usage, an arbitrary wrong answer that changes from one sample to the next under the same prompt; detectable by comparing samples, unlike consistent errors. See: Confidently Wrong

Conformity assessment

The process of demonstrating that a product or high-risk AI system meets its legal requirements before it is placed on the market, done either internally by the provider or with an independent notified body. See: Regulation as a Design Constraint

Confused deputy

A program with legitimate authority that is tricked into using it on someone else's behalf. An AI agent following injected instructions is a modern example. See: Giving an Agent Room to Work, Data Access Is the Real Integration Project

Constrained decoding

Restricting which tokens a model may produce so the output always matches a format, such as valid JSON for a given schema. See: What Training Actually Changes

Containment rate

The share of support conversations an automated system handles without escalating to a person, whether or not the issue was resolved. See: Customer Support Went First

Contamination

Test material leaking into a model's training data, or into choices made about training, so a benchmark score partly measures memory rather than skill. See: Leaderboards Measure the Test

Context

Everything the model receives for one response: instructions, conversation, documents, tool results. The model can use only what is in its context plus what it learned in training. See: What Fits Versus What Gets Used, Same Model, Different Results

Context rot

The decline in a model's accuracy and recall as the amount of text in its context grows, even when the needed information is present. See: What Fits Versus What Gets Used, Agents That Work for Hours

Context window

The maximum number of tokens a model can take in at once. A capacity, measured in the provider's own token unit; fitting is different from being used well. See: What Fits Versus What Gets Used, The Model Never Sees Your Words

Continuous access evaluation

An identity-system feature that re-checks access during a session (for example after a password change or account disable) instead of waiting for a token to expire. See: Data Access Is the Real Integration Project

Contract test

An automated test that checks one component still meets the expectations another component relies on, so a change on one side of a boundary fails fast if it breaks the other. See: How AI-Native Teams Organize

Contribution margin

Revenue minus variable costs for a unit (seat, customer or task); here approximately gross margin per seat. See: Why AI Products Have Thinner Margins

Conway's law

Melvin Conway's 1968 observation that organizations design systems that mirror their own communication structures. See: How AI-Native Teams Organize

Cost of revenue

The direct costs of delivering a product (hosting, model calls, support, payment fees), subtracted from revenue to get gross profit; also called cost of goods sold. Companies draw the line differently, so compare definitions. See: Why AI Products Have Thinner Margins

Counter-positioning

A strategy an incumbent can't copy without damaging its existing business, such as pricing by outcome when the incumbent earns per seat. See: Where the Moat Moves

Counterfactual

What would have happened without the change. Impact is the difference between the outcome and the counterfactual, not between before and after. See: Measuring AI Impact Without Fooling Yourself

Crawl-to-referral ratio

How many pages a service fetches from websites for each visit it sends back; a high ratio means it takes pages and returns little traffic. See: When Answers Replace Links

Critical decision method

An interview technique (Klein, Calderwood and MacGregor, 1989) that walks an expert through one specific non-routine incident several times with probe questions, to draw out the cues and expectations a general question misses. See: Most of What a Company Knows Isn't Written Down

Critical path

The longest chain of dependent steps in a piece of work; it sets the minimum total duration. See: From Individual Gains to Team Capability, Where the Saved Time Goes

Customer acquisition cost

What it costs, in sales and marketing, to win one new customer; compared with gross margin per customer to see how long it takes to pay back. See: Capability, Product and Business Model

Data contamination

When test questions or their answers appear in a model's training data, so a benchmark score partly measures memory rather than ability. See: Leaderboards Measure the Test, What Does a Model Actually Learn?

Data processing agreement (DPA)

The contract that governs how a vendor processes personal data on a company's behalf, required under GDPR Article 28. See: Data Access Is the Real Integration Project

Decision record

A short note of a decision, its context, the alternatives considered and the condition for revisiting it. See: Write for the Model That Doesn't Exist Yet, What a Deployment Engineer Should Own, and Leave Behind

Default placement

Putting a feature where users meet it without choosing it (on by default, or shown inside a product they already use); the strongest form of distribution and the weakest evidence of preference. See: Distribution Beats the Better Model

Definition of done

The agreed conditions a piece of work must meet to count as finished; most useful when expressed as checks that can be run. See: How AI-Native Teams Organize

Deflection

In support, a customer contact handled without reaching a human agent. Counts avoided contacts, not solved problems. See: Customer Support Went First

Delegated access

An application acting with a signed-in user's permissions, so it can see only what that person can see. See: Data Access Is the Real Integration Project

Deployer (AI Act)

Under the EU AI Act, a person or organization using an AI system under its authority in a professional context. Deployers carry duties about how a system is used, overseen and disclosed. See: Regulation as a Design Constraint

Desirable difficulty

A condition that makes practice harder or slower in the moment (spacing, mixing problem types, producing rather than recognizing) but leads to more durable learning (Bjork). See: Skills That Fade When the Machine Does Them

Deskilling

Losing a skill one had because a tool now does the work that used to keep it in practice. See: Skills That Fade When the Machine Does Them

Difference-in-differences

Estimating an effect by comparing how much a treated group changed with how much a similar untreated group changed over the same period. See: Measuring AI Impact Without Fooling Yourself

Disruptive innovation

In Christensen's theory, an offering that starts in a foothold incumbents ignore (the low end or non-customers) and improves until it displaces them. See: Distribution Beats the Better Model

Double-entry bookkeeping

Recording every transaction in at least two accounts so that total debits equal total credits; an entry that doesn't balance can't post, but a balanced entry to the wrong account still can. See: Finance and the Reconciliation Wall

DPO

Direct preference optimization: a way to train on pairs of preferred and rejected answers without a separate reward model. See: What Training Actually Changes

Drift

The gradual gap that opens between what a system was set up to know (suppliers, prices, rules, templates) or how its inputs look, and what is true or typical now, so its answers degrade without any change to the system. See: From a Good Demo to a Working Department

Durable execution

Running multi-step work so each completed step is recorded; after a crash, the work resumes from the record instead of repeating finished steps. See: When the Tool Call Fails Halfway

Effective length

The longest input at which a model keeps most of its short-input accuracy on a given task; depends on the task and the threshold chosen, so a model has many effective lengths. See: What Fits Versus What Gets Used

Effort setting

An API parameter that steers how much a reasoning model thinks and writes; soft guidance rather than a strict token budget. See: What Thinking Buys

Egress control

Limits on where a system can send data or connect to (allowed domains, recipients, no automatic loading of remote content), so a compromised component has no path out. See: Giving an Agent Room to Work

Embedding

A list of numbers representing a piece of text so that texts with related meaning end up close together. Used for search by meaning. See: Search by Meaning, and Where It Fails, What Does a Model Actually Learn?

Emergent ability

A capability that seems to appear suddenly as models grow. Some apparent jumps turn out to depend on how the ability is scored. See: What Does a Model Actually Learn?

Evaluation set

A fixed collection of real or realistic cases with known good outcomes, run to compare models, prompts or tools on your own work. See: Build an Evaluation From Your Own Work, Leaderboards Measure the Test

Expected loss

how often an error happens times what it costs when it does, summed over the kinds of error; used to compare designs before any error occurs. See: Reversible Work and Irreversible Effects

Fail-to-pass test

a test that fails before a change and must pass after it, defining what the change has to make work (SWE-bench's FAIL_TO_PASS). See: Tests Are the New Specification, How a Coding Agent Actually Works

Feature (interpretability)

A direction in a model's internal activity that corresponds to a recognizable concept, such as a city, a language or a code error, found by interpretability research. See: What Does a Model Actually Learn?, What Interpretability Can and Can't Show

Few-shot prompting

Showing a model a few worked examples of the task in the prompt so it imitates the pattern. See: What Training Actually Changes

Fork (of a product)

A customer-specific copy of a product that diverges from the shared version, so improvements no longer flow between them. See: What a Deployment Engineer Should Own, and Leave Behind

Forward-deployed engineer (FDE)

An engineer who works inside a customer's operation to understand how the work really runs and make a product deliver results there. See: What a Deployment Engineer Should Own, and Leave Behind

Fully loaded cost

What a person costs the employer, counting benefits, supervision, space, equipment and training as well as pay. See: Where the AI Budget Comes From

General-purpose AI model

In the EU AI Act, a model that shows significant generality and can perform a wide range of distinct tasks, such as a large language model. Its providers have documentation, copyright and transparency duties, and more if the model carries systemic risk. See: Regulation as a Design Constraint

General-purpose technology

a technology that spreads across most of the economy, keeps improving and spawns complementary inventions (electricity, computers, arguably AI); its gains usually arrive with a long lag because work has to be reorganized around it. See: What Becomes Possible Now?

Generative engine optimization (GEO)

Changing a page's content so that answer engines cite or quote it more often; named and tested in a 2023 paper by Aggarwal and colleagues. See: When Answers Replace Links

Goodhart's law

When a measure becomes a target, it stops being a good measure, because people (and models) optimize the number rather than the goal. See: Measuring AI Impact Without Fooling Yourself, Leaderboards Measure the Test

Gradient descent

The procedure that trains a model: measure how wrong a prediction was, work out which direction to nudge each parameter to make it less wrong, and take a small step. Repeated millions of times. See: What Does a Model Actually Learn?

Grokking

When a network that has fit its training examples later, with much more training, jumps to generalizing to new ones; the jump in score can hide a mechanism that formed gradually. See: What Does a Model Actually Learn?

Gross margin

Revenue minus the direct cost of delivering the product (for AI products, largely model and infrastructure costs), as a share of revenue. See: Capability, Product and Business Model, Why AI Products Have Thinner Margins

Hallucination

A model's fluent, specific statement that is false or unsupported; an umbrella word for errors with several distinct causes (missing knowledge, bad context, faulty reasoning, forced output, rewarded guessing). See: Confidently Wrong

Harness

The software around a model: what it puts in the context, which tools it offers, how it loops, checks results and handles permissions. The same model can perform very differently in different harnesses. See: Same Model, Different Results

Held-out test

a test kept from the system being built or trained and used only to measure it, so it can't be tuned or edited toward. See: Tests Are the New Specification, Build an Evaluation From Your Own Work

High-risk AI system

In the EU AI Act, a system used as a safety component of regulated products or for uses listed in Annex III (such as hiring, credit scoring, education or access to public benefits). It must meet requirements on risk management, data, logging, oversight and documentation before it is supplied. See: Regulation as a Design Constraint

Idempotency

The property that doing an operation twice has the same effect as doing it once. Essential for safely retrying actions after an uncertain failure. See: When the Tool Call Fails Halfway

Idempotency key

A unique identifier sent with a request so the receiving system can recognize a retry and return the original result instead of acting again. See: When the Tool Call Fails Halfway

Innovation budget

A discretionary pool of money for experiments, usually small, held centrally or by a senior executive and easy to reduce at the next review. See: Where the AI Budget Comes From

Installed base

The customers already using a company's products, who can be offered something new without being acquired again. See: Distribution Beats the Better Model

Instruction file

A plain file of standing instructions (such as AGENTS.md or CLAUDE.md) that a coding agent's harness loads at the start of every session in a project. See: How AI-Native Teams Organize

Intended purpose

The use a provider intends for an AI system, as set out in its instructions, technical documentation and sales or promotional materials. Under the EU AI Act it decides which risk category, and therefore which obligations, apply. See: Regulation as a Design Constraint

Internal control over financial reporting, ICFR

The processes and controls that give reasonable assurance a company's financial statements are reliable; US public companies' management must assess it each year under SOX Section 404. See: Finance and the Reconciliation Wall

Jagged frontier

The uneven boundary of AI capability: tasks of similar apparent difficulty can fall well inside or just outside what a model does reliably. See: Learn to Drive Before You Study the Engine, Why Progress Arrives Unevenly

Jevons paradox

When something becomes cheaper to use, total use can rise enough that total spending grows. Described by William Stanley Jevons for coal in 1865; relevant to code, tokens and content. See: What Changes When Code Gets Cheap, Where the AI Money Goes

Jobs to be done

a way of understanding demand by the progress a person is trying to make in a particular circumstance, rather than by product features or customer demographics (Christensen et al.). See: What Becomes Possible Now?

JSON-RPC

A simple message format in which one side sends a request naming a method and parameters and the other replies with a result or an error; MCP and the Language Server Protocol both use it. See: Why MCP Spread

Lagging indicator

The measure you ultimately care about, which arrives late and is harder to attribute. See: Output, Outcome and Value

Language Server Protocol

A protocol, announced by Microsoft's Visual Studio Code team in 2016, that lets one language server provide features such as completion and go-to-definition to many code editors; the precedent MCP cites. See: Why MCP Spread

Latent demand

work or purchases people would want at a lower price or effort but don't currently do, so it shows up in no existing budget or statistic until the cost falls. See: What Becomes Possible Now?, What Changes When Code Gets Cheap

Lead time

Time from a request arriving to its result being used. Measures the whole job, including waiting, rather than one step. See: Measuring AI Impact Without Fooling Yourself, Where the Saved Time Goes

Leading indicator

A measure that moves early and is cheap to collect, used as a signal for a slower measure that matters more. See: Output, Outcome and Value

Least privilege

Giving a person, program or agent only the access a specific task needs, for only as long as it needs it. See: Giving an Agent Room to Work

Lethal trifecta

The combination of access to private data, exposure to content an attacker could have written, and a way to send data out; an agent with all three can be tricked into leaking data (term coined by Simon Willison, 2025). See: Giving an Agent Room to Work

Leverage ratio

In a professional-services firm, the number of non-partner staff for each partner; a higher ratio means more billable junior hours earning a markup for each partner. See: Selling Hours When Hours Shrink

Long tail

The many rare categories that together make up a large share of real cases (and often most of the errors), while a few common categories make up the rest; a demo built from typical examples rarely contains them. See: From a Good Demo to a Working Department

Marginal value

The value of one additional unit, which can differ sharply from the average value of all units. See: Output, Outcome and Value

Material weakness

In SEC rules, a deficiency in internal control over financial reporting such that there is a reasonable possibility a material misstatement won't be prevented or detected on a timely basis; public companies must disclose it. See: Finance and the Reconciliation Wall

MCP (Model Context Protocol)

An open protocol for connecting AI applications to tools and data sources through a common interface, so one connector works with many AI clients. See: Why MCP Spread, Data Access Is the Real Integration Project

Memorization

Reproducing specific training examples rather than applying a general pattern. Useful for facts, misleading on benchmarks whose questions leaked into training data. See: What Does a Model Actually Learn?, Leaderboards Measure the Test

Memory poisoning

A prompt injection that writes instructions or false facts into an AI system's long-term memory, so they influence future sessions instead of only the current one. See: What It Means for AI to Remember, Giving an Agent Room to Work

Mixture of experts

A model design that splits much of the network into many specialist sub-networks and routes each token through only a few of them, so the model stores far more parameters than it uses per token. See: Open Weights and Closed Labs

Moat

A durable advantage that keeps competitors from taking a company's profits, such as switching costs, network effects or exclusive access. See: Where the Moat Moves

Model

The trained neural network itself: it receives a sequence of tokens and returns probabilities for the next token. Everything else you see in a product (memory, tools, files, checks) is added around it. See: Same Model, Different Results, What Does a Model Actually Learn?

Model routing

Sending each request to the smallest model that can handle it and escalating failures to a larger one. See: Why AI Products Have Thinner Margins

Model spec

A published document in which a lab describes how its models are intended to behave: priorities, refusals, tone. Used to guide training and evaluation. See: What Training Actually Changes, Personality Is a Product Decision

Mutation testing

making many small deliberate changes to code and checking whether the tests notice each one; the share noticed is the mutation score. See: Tests Are the New Specification

Needle in a haystack

A test that hides one sentence in a long, unrelated text and asks a question whose answer is that sentence, at many lengths and positions. Easy for current models because the question usually repeats the sentence's words. See: What Fits Versus What Gets Used, Leaderboards Measure the Test

Network effects

When a product becomes more valuable to each user as more people use it. See: Where the Moat Moves

Never-skilling

Not acquiring a skill in the first place because a tool does the repetitions that would have built it; typical for students and new hires. See: Skills That Fade When the Machine Does Them

Next-token prediction

The training objective of language models: given text so far, assign probabilities to what comes next. Doing it well forces the model to learn much of what makes text predictable. See: What Does a Model Actually Learn?

One-way door

a decision or effect that is consequential and hard or impossible to reverse, in Jeff Bezos's terms; contrasted with two-way doors, which can be undone cheaply. See: Reversible Work and Irreversible Effects

Open source AI

In the Open Source Initiative's 2024 definition, a system released with its parameters, its complete training and inference code, and enough information about its training data for a skilled person to build a substantially equivalent system, under terms that allow use, study, modification and sharing for any purpose. See: Open Weights and Closed Labs

Open weights

A model whose trained parameters are published for anyone to download and run, usually without the training data or the code that produced them. See: Open Weights and Closed Labs

Opportunity cost

The value of the best alternative use of a resource (often time) that was given up. See: Output, Outcome and Value

Outcome

A change someone can observe because of the output, such as a deal won, a feature in use or a decision made. See: Output, Outcome and Value

Outcome-based pricing

Charging per result delivered (a resolved ticket, a processed invoice) rather than per seat or per unit of usage. See: Seats, Usage or Outcomes, Who Receives the Gain?

Output

What a piece of work directly produces (drafts, pull requests, proposals, tickets closed), counted regardless of whether anything changes because of it. See: Output, Outcome and Value

Parallel trends

The assumption behind difference-in-differences: without the change, both groups would have moved the same way. See: Measuring AI Impact Without Fooling Yourself

Parameter

One of the billions of numbers inside a model that training adjusts. Together they encode what the model has learned. See: What Does a Model Actually Learn?

Pass-through

The share of a cost change that reaches customers as a price change. Competition usually raises it. See: Who Receives the Gain?

pass^k

The probability that a system succeeds on all of k repeated attempts at the same task; a stricter reliability measure than succeeding once. See: Build an Evaluation From Your Own Work

Permission-aware retrieval

Searching a company's documents while filtering results by what the asking person is allowed to open, so an assistant can't leak documents through its answers. See: Data Access Is the Real Integration Project, What It Means for AI to Remember

Post-training

Everything done to a model after pretraining to shape how it behaves: instruction tuning, preference optimization, reinforcement learning, tool use training. It decides much of a model's style and habits. See: What Training Actually Changes

Postmortem

A written review after an incident covering what happened, why, and what will change, usually without blaming individuals. See: When Work Starts to Compound, An Incident That Improves the System

Pretraining

The first and largest training stage: a model learns to predict the next token over a huge mix of text, which is where most of its knowledge and fluency come from. See: What Training Actually Changes, What Does a Model Actually Learn?

Price elasticity of demand

How much the quantity people buy changes when price changes. High elasticity means cheaper work leads to much more work being bought. See: A Job Is a Bundle of Tasks, What Changes When Code Gets Cheap

Probe

A small model trained to read a specific piece of information (say, the state of a game board) from another model's internal activity. If it succeeds, the information is represented there. See: What Does a Model Actually Learn?

Product memory

Notes a product saves about you or your work outside the model and puts back into the context in later sessions. Distinct from what the model learned in training. See: What It Means for AI to Remember, Learn to Drive Before You Study the Engine

Prompt caching

Reusing the processed form of a repeated prompt prefix so it is cheaper and faster the next time. Providers bill cached input at a discount. See: Why Long Answers Cost More Than Long Questions, Where the AI Money Goes

Prompt injection

Text in content an AI system reads (a web page, email, document) that tries to make it follow the attacker's instructions instead of the user's. See: Giving an Agent Room to Work, Data Access Is the Real Integration Project

Property-based testing

stating rules that must hold for every input and letting a tool generate many inputs to find one that breaks a rule, then simplify it. See: Tests Are the New Specification

Provider (AI Act)

Under the EU AI Act, whoever develops an AI system or general-purpose model, or has one developed, and places it on the market or puts it into service under its own name, including for its own use. Providers carry most design-time obligations. See: Regulation as a Design Constraint

Quantization

Storing a model's numbers with fewer bits (for example 8 or 4 instead of 16) to cut memory and speed up serving, at some risk to accuracy. See: Open Weights and Closed Labs

Ratchet

A practice where every real failure becomes a permanent improvement (a test, a check, a constraint), so the same mistake can't recur silently. See: When Work Starts to Compound, Build an Evaluation From Your Own Work

Rate card

A vendor's or contractor's list of prices per unit, hour or role, which a buyer compares new offers against. See: Where the AI Budget Comes From

Realization rate

The share of billed (or worked) value that the client ultimately pays after discounts, write-downs and collection losses. See: Selling Hours When Hours Shrink

Reasoning model

A model trained, usually with reinforcement learning on checkable problems, to write a long working draft (thinking tokens) before its answer; the draft is produced by the same next-token prediction as any other text. See: What Thinking Buys

Reconciliation

Asking the other system what actually happened (was the payment made?) before deciding whether to retry, instead of guessing from an error. In accounting the word means something related but narrower; see bank-reconciliation. See: When the Tool Call Fails Halfway

Regression test

an existing test that must keep passing after a change, guarding behavior that should not change (SWE-bench's PASS_TO_PASS). See: Tests Are the New Specification

Regression to the mean

Extreme results tend to be followed by more ordinary ones. A team chosen for a pilot because it was struggling will often improve regardless. See: Measuring AI Impact Without Fooling Yourself

Relational tacit knowledge

Knowledge that could be told but isn't, because of the relationship between holder and learner: it is secret, the holder doesn't realize they know it, passing it on is awkward, or the holder misjudges what the other person lacks (Harry Collins). See: Most of What a Company Knows Isn't Written Down

Resolution rate

The share of support conversations in which the customer's issue was actually solved, the measure containment rates often stand in for. See: Customer Support Went First, Measuring AI Impact Without Fooling Yourself

Retrieval

Searching a collection and putting only the passages likely to matter into the model's context, by keyword, by meaning (embeddings) or both. See: Search by Meaning, and Where It Fails, What Fits Versus What Gets Used

Reverse acqui-hire

A deal in which a large company licenses a startup's technology and hires its key people without buying the company itself. See: The Price of Know-how

Reward hacking

When a model finds a way to score well on the training signal without doing the intended task, such as weakening a test so it passes. See: What a Reward Teaches, How a Coding Agent Actually Works

Reward model

A model trained to score outputs the way human raters would. Used as the signal for reinforcement learning; its blind spots become the trained model's habits. See: What Training Actually Changes, What a Reward Teaches

RLHF

Reinforcement learning from human feedback: people compare model answers, a reward model learns their preferences, and the model is trained to produce answers the reward model scores highly. See: What Training Actually Changes, What a Reward Teaches

RLVR

Reinforcement learning with verifiable rewards: training on tasks whose answers can be checked automatically (math results, passing tests), so the model is rewarded for being right, not for sounding right. See: What Training Actually Changes, Why Progress Arrives Unevenly

robots.txt

A file at the root of a website that tells each crawler, by name, which paths it may fetch; honoring it is voluntary on the crawler's side. See: When Answers Replace Links

RPM

Revenue per thousand pageviews (per mille), a standard measure of how much advertising earns a site. See: When Answers Replace Links

Runbook

Written instructions for operating and recovering a system: what to check, what to do when something specific goes wrong. See: What a Deployment Engineer Should Own, and Leave Behind

Saga

A multi-step operation where each step has a compensating action (cancel the hotel if the flight booking fails), used when steps can't share one transaction. See: When the Tool Call Fails Halfway

Saturation

When top models score so near a benchmark's ceiling that differences between them stop being informative. See: Leaderboards Measure the Test

Scaffolding

Harness code that compensates for a model's current weakness (rigid step lists, output repair, elaborate prompt rules). Expect to remove it as models improve. See: Same Model, Different Results

Segregation of duties

A control in which no single person (or system) can both initiate and approve a sensitive action, such as a payment. See: Data Access Is the Real Integration Project, Human Review Where It Matters

Service account

A non-human identity with its own permissions, often broad. Convenient for automation and risky when an AI system uses it to act for many people. See: Data Access Is the Real Integration Project

Shadow AI

AI tools employees use for work without official approval, often on personal accounts. See: The AI Your Employees Already Use, Why People Resist AI at Work, and When They're Right

Shadow mode

Running a new system on real inputs without letting its outputs take effect, so its results can be compared with the current process before switching. See: Moving Without Breaking

Skill shift

A tool takes over one skill and the work's demand moves to another, such as from computation to estimation; beneficial only if the new skill suffices when the tool fails. See: Skills That Fade When the Machine Does Them

Snapshot test (golden test)

a test that stores an output once and fails when it changes; it checks sameness, not correctness. See: Tests Are the New Specification

Specification gaming

satisfying the literal wording of an objective or check without achieving what its designer intended. See: Tests Are the New Specification, What a Reward Teaches

Static analysis

Automated inspection of source code without running it, to find likely bugs, security flaws and rule violations by pattern and data-flow rules. See: Reviewing Code You Didn't Write

Sticky information

Knowledge that is costly to move from where it lives to where it's needed, often because it is tacit or tied to a local context. See: What a Deployment Engineer Should Own, and Leave Behind, Most of What a Company Knows Isn't Written Down

Strangler fig pattern

Replacing a system gradually by building new parts alongside it and moving behavior over piece by piece, until the old part can be retired. See: Moving Without Breaking

Streamable HTTP

The MCP transport for remote servers: one web endpoint, each message sent as an HTTP request, the reply a single JSON object or a stream of events. See: Why MCP Spread

Style control

An adjustment in a preference ranking that estimates and removes the effect of answer length and formatting on votes, so scores reflect content more than presentation. See: Leaderboards Measure the Test

Subagent

A separate model session started by an agent to handle part of a task in its own clean context, returning only a result. See: When to Split the Work Across Agents

Supervised fine-tuning (SFT)

Training a model further on example inputs paired with the desired outputs, such as instructions and good answers. See: What Training Actually Changes

Sustaining innovation

In Christensen's theory, an improvement to an existing product for existing customers along the current dimension of competition; incumbents usually adopt these. See: Distribution Beats the Better Model

Switching costs

Everything a customer would lose or spend to move to a competitor: data migration, retraining, integrations, risk. See: Where the Moat Moves, Choosing a Model Provider Without Marrying It

Sycophancy

A model's tendency to agree with the user, flatter, or abandon a correct answer under pushback, often a side effect of training on human approval. See: Why Models Agree With You, What Training Actually Changes

System of record

The authoritative source for a type of data in a company (the ERP for invoices, the CRM for customers). Products that become one are hard to replace. See: Where the Moat Moves

System prompt

Instructions placed at the start of every conversation by the product or developer, shaping tone, rules and available tools. The user usually doesn't see it. See: Same Model, Different Results, Personality Is a Product Decision

Tacit knowledge

Know-how people use but haven't written down (exceptions, judgment, who to ask, why things are done a certain way). See: Most of What a Company Knows Isn't Written Down, The Price of Know-how

Task (in labor economics)

A unit of work within a job (drafting a contract clause, checking a calculation). Automation acts on tasks; jobs are bundles of them. See: A Job Is a Bundle of Tasks

Technical debt

Shortcuts in software that save time now and cost more later, in maintenance and slower changes. The term comes from Ward Cunningham (1992). See: When Work Starts to Compound, Technical Debt at Machine Speed

Test-time compute

Computation spent when a model answers (longer reasoning, more samples), as opposed to during training. See: What Thinking Buys

Thinking tokens

Tokens a model generates while reasoning before its answer; billed as output and usually hidden or summarized. See: What Thinking Buys

Three-way match

Checking that an invoice agrees with the purchase order and the goods receipt on items, quantities and prices before it is paid. See: Finance and the Reconciliation Wall

Tie-out

Confirming that a number agrees exactly with the record it should come from, such as a sub-ledger total with the general ledger. See: Finance and the Reconciliation Wall

Time horizon

METR's measure of agent capability: the length of task, in the time it takes a skilled human, that a model completes with a given success rate (usually 50%). See: Why Progress Arrives Unevenly, Leaderboards Measure the Test

Toil

Repetitive, manual operational work that scales with volume and leaves nothing lasting behind (a term from Google's Site Reliability Engineering book). See: When Work Starts to Compound

Token

A piece of text (a word, part of a word, a space plus a word, a digit group) from a fixed menu. Models read and write tokens, and providers bill by them. See: The Model Never Sees Your Words

Tokenizer

The fixed program that cuts text into tokens and maps them to numbers. Its menu is learned from training text, so it favors what was common there. See: The Model Never Sees Your Words

Tool call

A structured request a model writes asking the harness to run a tool (search, read a file, send an email) with given arguments; the harness executes it and returns the result. See: When the Tool Call Fails Halfway, Same Model, Different Results

Tool poisoning

An attack in which hidden instructions are placed in a tool's description or output, visible to the model but not to the user, so that the model leaks data or misuses other tools; related patterns are rug pulls (a description changed after approval) and shadowing (one server's text altering how another's tool is used). See: Why MCP Spread

Total factor productivity

Output growth not explained by more labor or capital, the economist's measure of how much more efficiently inputs are used. See: Who Receives the Gain?

Training cutoff

The date after which a model has seen no training data. Later facts reach it only through context, tools or retrieval. See: Where a Model's Knowledge Ends

Trajectory

The sequence of tool calls, results and decisions an agent makes in one run. See: Workflows or Agents?

Transactional outbox

Writing an intended message to the database in the same transaction as the change it reports, then sending it from there, so a message is never lost or sent for a change that didn't happen; the relay may still send twice, so the receiver deduplicates. See: When the Tool Call Fails Halfway

Unit drive

powering each factory machine with its own electric motor instead of belts from a shared line shaft, which freed factory layout to follow the flow of work. See: What Becomes Possible Now?

Unit economics

The revenue and cost of one unit (a seat, a customer, a task) used to judge whether the business scales profitably. See: Why AI Products Have Thinner Margins

Usage based pricing

Charging by what the customer consumes (tasks, tokens, messages) so revenue grows with the cost of serving them. See: Why AI Products Have Thinner Margins

Utilization rate

The share of a person's available hours that are billed to clients. See: Selling Hours When Hours Shrink

Value

What an outcome is worth to someone, net of the cost of producing it, the cost of fixing errors and the value of the work it displaced. See: Output, Outcome and Value

Verification of payee

a check, mandatory for euro-area payment providers since 9 October 2025, that warns the payer when the account number and the payee's name don't match before a transfer is sent. See: Reversible Work and Irreversible Effects

Walls

Harness controls that encode facts about the world or the organization (permissions, approvals, audit, sources of truth). They stay, and matter more, as models improve. See: Same Model, Different Results, Giving an Agent Room to Work

Wilson interval

A way to put a confidence range around a success rate measured on a small number of cases; more reliable than the simple formula when rates are near 0% or 100%. See: Build an Evaluation From Your Own Work

Worked example

A fully solved example studied before attempting similar problems; one of the best-supported techniques for learning a new skill. See: Learn to Drive Before You Study the Engine

Workflow

A system in which code defines the sequence of steps and calls a model to fill in some of them; the model does not decide what happens next. Contrast with agent. See: Workflows or Agents?

Workslop

AI-generated work that looks finished but lacks the substance to move a task forward, shifting the effort to whoever receives it. See: Where the Saved Time Goes, Measuring AI Impact Without Fooling Yourself

World model

An internal representation of how a situation works (positions, causes, states) that a model can use to predict, as opposed to surface patterns in text alone. See: What Does a Model Actually Learn?

All writing