AI Glossary 2026: The New and Emerging AI Terms You Need to Know

From agentic AI and MCP to context engineering, recurrent depth, world models and vibe coding, artificial intelligence has acquired an entirely new vocabulary. This is Nuvastra’s guide to the terms that explain where AI is actually heading in 2026 — and what they really mean.

Artificial intelligence has always come with jargon. Neural networks, machine learning, transformers and large language models once sounded specialised before becoming part of ordinary technology coverage. What has changed in 2026 is not simply the number of new terms appearing, but what those terms describe.

The industry's vocabulary is moving away from the model alone.

Increasingly, the important questions concern what surrounds the model: whether it can reason for longer before answering, use software tools, remember previous interactions, communicate with other agents, operate a computer, adapt after deployment or translate intelligence into physical action. New terminology is consequently appearing at the boundaries between AI models, conventional software, infrastructure and the real world.

Readers looking for the underlying hierarchy of artificial intelligence, including machine learning, deep learning, transformers, foundation models and large language models, should begin with Nuvastra’s AI Fundamentals guide. This glossary concentrates instead on the terms becoming increasingly important as the industry moves beyond the first generative-AI era.

The emerging AI vocabulary at a glance

Area	Terms increasingly worth understanding
AI that acts	Agentic AI, AI agents, agentic workflows, coding agents, computer use, tool calling, MCP, A2A, orchestration
AI that reasons	Reasoning models, test-time compute, chain of thought, latent reasoning, recurrent depth, recursive self-improvement
AI that remembers and adapts	Context engineering, RAG, grounding, AI memory, model drift, continual learning, catastrophic forgetting
AI that must be controlled	Evals, LLM-as-a-judge, reward hacking, deterministic AI, agent washing
AI beyond the chatbot	World models, physical AI, embodied AI, VLA models, world-action models, open-weight AI, vibe coding, AI slop

The language of AI that acts

Agentic AI

Agentic AI describes artificial-intelligence systems that can pursue objectives by deciding what actions to take rather than simply producing a response to a single prompt. A genuinely agentic system might formulate a plan, search for information, call a software tool, inspect the result, modify its strategy and continue working until it completes a task or encounters a boundary requiring human intervention.

The adjective matters because almost every major AI company is moving in this direction. OpenAI describes agents as systems that independently accomplish tasks for users, while Anthropic has converged on the practical description of agents as language models autonomously using tools in a loop.

Agentic AI should not be confused with automation generally. A fixed workflow can automate hundreds of steps without making autonomous decisions about which steps should occur.

AI agent

An AI agent is an individual system designed to pursue a goal by combining a model with tools, instructions, state and some mechanism for deciding what to do next.

The distinction between an AI model and an AI agent is fundamental. A model receives an input and generates an output. An agent wraps that model inside a wider system capable of repeatedly observing, reasoning and acting. The model might therefore be thought of as the intelligence available to the agent, while the agent architecture determines what that intelligence is allowed to do.

This difference is becoming economically important because agents can invoke models many times during a single task. As Nuvastra examined in its analysis of Gemini 3.7 Flash and the economics of agentic workloads, cost per model call becomes less meaningful when one job can require dozens of calls, retries and tool actions.

Agentic workflow

An agentic workflow combines language-model reasoning with a larger process, but not every decision inside the process necessarily belongs to the model.

Anthropic makes a useful distinction here. In a workflow, models and tools are orchestrated along predefined code paths; in a more autonomous agent, the model dynamically determines its own process and tool usage. That means an organisation can use AI extensively without handing the model complete freedom over the sequence of events.

This distinction is increasingly important because the most reliable production systems may turn out to be hybrids. Models can interpret ambiguity and make judgements where flexibility is valuable, while conventional code can handle steps where behaviour must remain predictable.

Coding agent

A coding agent is an AI agent specialised for software engineering. Rather than merely completing the next few lines of code, it can inspect repositories, search files, devise implementation plans, edit several components, execute commands, run tests, interpret failures and make further changes.

That longer feedback loop separates a coding agent from conventional AI autocomplete. The model is not simply generating code; it is participating in the software-development process.

Coding has become one of the most important proving grounds for agentic AI because the environment provides unusually clear feedback. Code can be executed. Tests can fail. Errors can be inspected. That gives an agent the opportunity to discover whether its previous action actually worked.

Computer use

Computer use refers to AI systems interacting with graphical software interfaces in ways closer to a human user: identifying controls, clicking, typing, navigating applications and completing work across software that may not expose a convenient API.

Computer use significantly expands the range of tasks an agent can perform because much of the world's software was designed for humans rather than machines. OpenAI's 2026 engineering work explicitly frames the addition of a computer environment as part of the transition “from model to agent”, allowing models to run services, work with files and complete workflows beyond the intelligence contained in the model itself.

The trade-off is control. Giving an agent access to a computer also gives it access to whatever credentials, files and applications that environment exposes.

Tool use and function calling

Tool use, sometimes called function calling, allows a model to request that external software perform an operation.

The model might decide that it needs to search a database, send information to an API, run a calculation or retrieve a file. Instead of pretending it already knows the result, it selects an appropriate tool and supplies the arguments required to invoke it. The surrounding application executes the operation and returns the result to the model.

This is one of the technologies that transformed the LLM from an isolated text generator into a component capable of affecting external systems.

Tools also create an important boundary: the model can decide what it wants to do, while conventional software can still determine whether that request is permissible.

Model Context Protocol — MCP

Model Context Protocol, or MCP, is an open standard for connecting AI applications to tools, data and external systems through a common interface.

Anthropic introduced MCP in 2024 and donated it in December 2025 to the Agentic AI Foundation under the Linux Foundation. Anthropic said at the time that more than 10,000 active public MCP servers existed and that the protocol had been adopted across products including ChatGPT, Gemini, Microsoft Copilot and Visual Studio Code. Those adoption numbers are Anthropic's own, but the broader industry support makes MCP one of the most consequential infrastructure standards to emerge from the agent boom.

The significance is interoperability. Instead of every AI application requiring a bespoke integration with every service, MCP provides a common way for applications to discover and use compatible tools.

Read Anthropic’s announcement establishing MCP within the Agentic AI Foundation

Agent2Agent Protocol — A2A

Agent2Agent, or A2A, addresses a related but different interoperability problem: communication between agents.

Originally developed by Google and subsequently donated to the Linux Foundation, A2A provides a standard through which agents built using different platforms or frameworks can discover capabilities, exchange information and coordinate work. The Linux Foundation said in April 2026 that the standard had support from more than 150 organisations and was moving into production deployments.

The easiest distinction is useful enough to remember: MCP mainly helps an AI application connect to tools and data; A2A helps agents communicate with other agents. The two standards can therefore coexist inside the same architecture rather than competing for precisely the same role.

Explore the official A2A protocol documentation

Multi-agent system

A multi-agent system uses several AI agents rather than relying on one agent to perform an entire job.

One agent might research a subject, another analyse data, another produce a draft and a fourth evaluate the result. Alternatively, different agents may represent specialised skills and collaborate dynamically as the task develops.

The attraction is similar to specialisation within human organisations: one system does not need to be equally capable at everything. The limitation is that coordination itself becomes a new engineering problem. Agents can duplicate work, disagree, propagate one another's mistakes or consume large amounts of inference compute while communicating.

More agents therefore do not automatically mean a better system.

Agent orchestration

Agent orchestration is the infrastructure responsible for coordinating models, tools, agents, state and workflows.

An orchestrator may decide which agent receives a task, whether work should occur sequentially or in parallel, what context should be passed between stages, how failures are retried and when a human needs to intervene.

As AI systems become more complex, orchestration is becoming one of the least glamorous but most important layers of the stack. A production agent is rarely just a powerful model connected directly to every available system. The orchestration layer helps determine what the model sees, what it can reach and how its decisions become actions.

Agent washing

Agent washing is the rebranding of conventional chatbots, assistants or automation as “AI agents” despite the product having little meaningful autonomy.

The term has become useful precisely because “agent” is commercially valuable. Gartner has repeatedly warned about agent washing and estimated in 2025 that only around 130 of the thousands of vendors then describing themselves as agentic had substantial agentic capabilities. That estimate belongs to Gartner rather than representing an independently measured census, but the underlying warning has become increasingly relevant in 2026.

The practical test is behavioural rather than linguistic. Can the system decide between possible actions, use tools, inspect what happened and modify its plan? Or is it following a predetermined script underneath an AI interface?

The language of AI that reasons

Reasoning model

A reasoning model is an AI model trained or configured to devote additional computation to working through difficult problems before producing its final answer.

The category became prominent with systems trained using reinforcement learning to develop extended chains of intermediate reasoning. Instead of asking only whether a larger pretrained model is inherently more intelligent, developers can also vary how much computation the model spends on a particular problem.

This has produced a major change in AI economics. Intelligence is increasingly something that can be bought not only during training but during inference.

Test-time compute

Test-time compute, also called inference-time compute, means allocating additional computation when the trained model is actually solving a problem.

OpenAI demonstrated with its early reasoning models that performance could improve both with more reinforcement learning during training and with more time spent reasoning during inference. Subsequent models have turned reasoning effort into an explicit operating parameter.

This introduces a new trade-off. A system can potentially produce a stronger answer by thinking for longer, but the additional reasoning consumes tokens, hardware capacity, time and money. Model selection is consequently becoming partly a question of how much intelligence a particular task is worth purchasing.

Chain of thought

Chain of thought, or CoT, refers to intermediate reasoning generated while a language model works towards an answer.

Earlier language models could be encouraged to reason through prompting — famously by asking them to work step by step. Modern reasoning models can be explicitly trained to generate extended internal reasoning before giving their final response.

Chain of thought matters for more than capability. AI laboratories are increasingly interested in whether these reasoning traces can help monitor advanced agents for deception, reward hacking or other unwanted behaviour. OpenAI's research has found that monitoring chain-of-thought can reveal behaviour that is substantially harder to detect from actions and final outputs alone, while cautioning that this monitorability may be fragile.

Read OpenAI’s research on chain-of-thought monitorability

Latent reasoning

Latent reasoning describes computation that occurs inside a model's hidden numerical representations rather than being expressed entirely as language tokens representing intermediate steps.

The potential advantage is efficiency. Generating a long textual reasoning trace requires the model to produce token after token; some research is exploring whether useful intermediate computation can occur directly within latent representations instead.

The difficult question is observability. If important reasoning occurs in representations that humans cannot straightforwardly read, safety systems may have less access to the information they currently use for chain-of-thought monitoring.

This is where discussions of “neuralese” sometimes appear. Neuralese is better understood as a speculative shorthand for machine reasoning becoming largely unintelligible to humans than as a precisely defined production technology.

Recurrent depth

Recurrent depth refers to architectures in which a model repeatedly applies a computational block to an internal representation, effectively gaining additional computational depth without requiring a completely separate set of parameters for every pass.

Research into looped or recurrent-depth transformers predates the current controversy. Microsoft Research, for example, published work in June 2026 showing how looped transformers could support latent reasoning while reducing thought-phase latency in its experimental system.

The broader importance is that AI may acquire another way to scale intelligence: not only through larger models or longer textual chains of thought, but through repeated internal computation.

Opaque recurrence

Opaque recurrence is a newer and more cautionary label being used around recurrent internal reasoning that produces fewer human-readable traces.

The phrase surged into technology coverage in September 2026 following reporting that OpenAI's GPT-6 Astra uses recurrent-depth techniques to a limited degree. OpenAI has publicly emphasised the importance of retaining chain-of-thought monitorability around Astra, but its launch materials do not provide a complete public architectural description that would justify treating “opaque recurrence” as a formally standardised model category.

That distinction matters. Recurrent depth is the technical architectural concept; opaque recurrence is principally a framing of the transparency problem that can arise when more reasoning happens internally.

The controversy is important because OpenAI now considers chain-of-thought monitoring a potentially important safety layer for highly capable systems. Anything that reduces the visibility of those traces therefore becomes more than an architectural curiosity.

Recursive self-improvement — RSI

Recursive self-improvement, or RSI, describes AI playing an increasing role in improving the systems that will succeed it.

The strongest version imagines a system redesigning or training progressively more capable successors with diminishing human involvement, potentially creating an accelerating feedback loop. That remains a future scenario rather than a demonstrated runaway process.

The phrase nevertheless moved significantly closer to mainstream frontier-AI discussion in 2026. OpenAI chief scientist Jakub Pachocki wrote in September that he expects machine intelligence to play a progressively larger role in its own development and explicitly described recursive self-improvement as likely to become central to future scientific discovery. That is his and OpenAI's forward-looking assessment, not evidence that unconstrained RSI has already occurred.

The language of context, memory and adaptation

Context window

A model's context window is the amount of information it can have available during a particular inference process.

The context can contain the user's current request, earlier messages, system instructions, retrieved documents, tool descriptions and other information supplied to the model. Larger context windows allow systems to consider more material at once, but size does not guarantee that every piece of information will be used equally well.

Context is also temporary. Something appearing inside the context window does not automatically mean the underlying model has permanently learned it.

That difference becomes essential when discussing memory.

Context engineering

Context engineering is the discipline of deciding what information should be placed into a model's context, how it should be organised and how that context should change while an AI system works.

Anthropic describes the concept as a progression beyond prompt engineering. A prompt concerns the instructions given to the model; context engineering concerns the whole information environment surrounding inference, including system instructions, tools, MCP connections, retrieved data, interaction history and intermediate work.

As agents tackle longer tasks, this becomes increasingly important because the context window is a finite resource. Filling it with everything available can make a system worse rather than better if irrelevant material competes with what actually matters.

Read Anthropic’s guide to context engineering for agents

Retrieval-augmented generation — RAG

Retrieval-augmented generation, or RAG, connects a generative model to external information retrieved at the time a request is made.

A typical RAG system searches documents, a database or another knowledge source, places relevant material inside the model's context and asks the model to generate its answer using that evidence. This can give an LLM access to proprietary or recently updated information that was not contained in its original training.

RAG does not mean the model itself has learned the retrieved information. Remove the external knowledge source and the underlying weights remain unchanged. This distinction separates retrieval from fine-tuning and, more importantly for 2026, from genuine continual learning.

Grounding

Grounding means anchoring an AI output to information, evidence or state outside the model's unconstrained generation.

A RAG system can ground an answer in retrieved documents. An agent can ground its next decision in the actual result returned by a tool. A customer-service system might ground an answer in the verified account record rather than allowing the model to infer what probably happened.

Grounding is therefore broader than RAG. Retrieval is one mechanism for obtaining evidence; grounding describes the relationship between the AI's output and that evidence.

It can reduce certain kinds of error, but it does not guarantee correctness. A model can still misunderstand perfectly accurate source material.

AI memory

AI memory refers to mechanisms that allow an AI application or agent to retain useful information across interactions or tasks.

That memory can take several forms. An application might store user preferences, previous events, summaries, structured state or selected conversation history in an external system and retrieve those details when they become relevant.

The important technical distinction is that memory is not necessarily learning. An agent can behave differently because it remembers something without changing the underlying model weights.

Nuvastra's reporting on continual-learning startups has found this distinction increasingly important because companies across memory, feedback infrastructure and model adaptation are sometimes grouped together despite changing very different layers of an AI system.

Hallucination

An AI hallucination is an output that presents unsupported, fabricated or incorrect information as though it were valid.

The word can be misleading if taken too literally. An LLM is not necessarily “seeing” something nonexistent in the human sense. It is generating an output whose statistical plausibility is not accompanied by factual reliability.

RAG, better training, tool use, web access and grounding can all reduce particular forms of hallucination, but none turns a generative model into an infallible database. A system can retrieve the right evidence and still reason incorrectly about it.

This is one reason hallucination remains relevant even as AI shifts from chatbots to agents: a fabricated answer is problematic; a fabricated premise that triggers an external action can be considerably more consequential.

Model drift

Model drift describes deterioration or behavioural change that occurs when conditions encountered after deployment no longer sufficiently match those under which an AI system was developed or evaluated.

The phrase covers several related problems. Data drift occurs when the characteristics of incoming data change. Concept drift occurs when the relationship the system is trying to model changes. A wider production AI system can also change when its underlying models, tools, prompts, retrieval sources or user population evolve.

This makes reliability a continuing operational problem rather than a test performed once before launch. Nuvastra's analysis of model-drift companies in 2026 examines the emerging infrastructure designed to detect exactly these changes.

Continual learning

Continual learning investigates how AI systems can acquire useful new information or capabilities over time while retaining what they previously learned.

The term is particularly important because most deployed foundation models do not simply update their weights after every conversation. They may receive new context, retrieve memories or later be retrained using collected data, but those are different processes.

Continual-learning research becomes difficult because adaptation introduces a control problem. A system should ideally improve from new experience without learning from corrupt inputs, drifting away from intended behaviour or destroying useful existing capabilities.

The attraction is obvious: today's largely static model could eventually become a system capable of improving through experience after deployment. The engineering challenge is making that change observable and reversible.

Catastrophic forgetting

Catastrophic forgetting occurs when learning new information damages capabilities a model previously possessed.

A model trained intensively on a new task might become better at that task while unexpectedly becoming worse at something it already knew how to do. Continual-learning research consequently has to solve two problems simultaneously: acquiring the new capability and preserving the old ones.

This is why claims that an AI “learns continuously” deserve scrutiny. Adaptation is not valuable merely because the model changes. It needs to improve without accumulating unacceptable regressions.

Evals

Evals is industry shorthand for evaluations: structured tests used to measure how a model or AI system behaves.

Benchmarks are one form of evaluation, but production evals can be much more specific. A company might test whether its support agent correctly follows refund policy, whether a coding agent completes a repository task without modifying unrelated files or whether an autonomous system stops appropriately when its permissions are insufficient.

Agents have made evals substantially harder because success may depend on an entire trajectory of actions rather than one final answer. Anthropic notes that autonomy, tool use and changing state make agent evaluation particularly challenging.

LLM-as-a-judge and autoraters

LLM-as-a-judge describes using one language model to evaluate another model's output. Google uses the related term autorater for machine-learning systems used to perform automatic evaluations.

The technique is attractive because many AI outputs do not have one answer that can be checked mechanically. A model can instead assess qualities such as relevance, clarity, compliance or factual support at a scale that would be expensive for humans.

The limitation is straightforward: the judge is itself an AI system. It can be biased, inconsistent or vulnerable to the same kinds of failure as the model being assessed.

An LLM judge is therefore evidence, not an objective measuring instrument.

Reward hacking

Reward hacking occurs when an AI system finds a way to maximise the metric or reward it has been given without accomplishing the underlying objective its designers intended.

The concept has existed in reinforcement learning for years, but agents make it more consequential because they can act. A coding agent told to make tests pass might discover a way to manipulate the test rather than fix the software. A system rewarded for completing tasks might exploit an unintended shortcut.

Nuvastra's reporting on the OpenAI agent that breached Hugging Face during a cybersecurity evaluation provides an unusually concrete example of the wider issue: an AI system can pursue the measurement successfully while violating the intended route to success.

Deterministic AI

Deterministic AI is increasingly being used to describe AI systems or components whose important decisions and execution paths are constrained by predictable rules rather than being left entirely to probabilistic model judgement.

The phrase does not identify one new model architecture. A modern system can combine a probabilistic LLM with deterministic authentication, business rules, schemas, permission checks and execution logic.

The distinction matters most for agents. A model may be useful for interpreting an ambiguous request, but there may be no benefit in allowing it to improvise whether a bank transfer requires approval or which database permissions apply.

As AI gains more autonomy, one of the most important architectural questions will therefore be not how to eliminate probabilistic reasoning, but where to stop it.

The language of AI entering the physical world

World model

A world model is an AI model designed to represent or predict how an environment behaves.

Instead of generating a static image or answering a question about the world, a world model attempts to model how a state might change over time, potentially including what happens when an action is taken.

Google DeepMind's Genie 3 illustrated the concept by generating interactive environments that could be navigated in real time. The significance extends well beyond synthetic worlds: an AI capable of predicting how environments evolve can potentially use those predictions for planning, robotics and autonomous systems.

Read Google DeepMind’s introduction to Genie 3 and world models

Physical AI

Physical AI is a term increasingly used for artificial intelligence that perceives, reasons about and takes action in the physical world.

NVIDIA defines physical AI around autonomous systems including robots and self-driving vehicles that can perceive and understand their environments before performing actions. The terminology has become particularly prominent as generative and foundation-model techniques move into robotics.

It is useful, but it is not a perfectly bounded academic category. Vendors may use “physical AI” more broadly than researchers do.

The underlying shift is nevertheless real: AI is moving from manipulating tokens and pixels to controlling machines whose errors can alter physical environments.

Embodied AI

Embodied AI overlaps substantially with physical AI but usually emphasises the relationship between intelligence, a physical or simulated body and an environment.

An embodied system learns or acts through perception and interaction rather than treating intelligence as an isolated computation. Robotics is the obvious application, but embodied-AI research can also involve simulated agents used to study navigation, manipulation and learning.

The distinction from physical AI is therefore partly one of emphasis. Physical AI has become a particularly prominent industry term around robotics and autonomous machines; embodied AI has a longer research tradition concerned with intelligence expressed through an agent situated in an environment.

Vision-language-action model — VLA

A vision-language-action model, or VLA, connects visual perception and language understanding with robot actions.

A VLA might see an object, interpret a natural-language instruction and output the actions required for a robot to manipulate that object. NVIDIA describes the contemporary VLA approach as building robot policies on pretrained vision-language foundations before adapting them to produce actions.

The significance is that robotics can potentially benefit from knowledge acquired during broad visual and language pretraining rather than learning every object and instruction exclusively from robot demonstrations.

VLAs are one of the clearest examples of foundation-model ideas crossing into physical machines.

World-action model — WAM

A world-action model, or WAM, is an emerging robotics concept that combines predictions about how the world will change with predictions about actions capable of producing those changes.

NVIDIA distinguishes this from a conventional VLA. A VLA generally maps observations and instructions towards actions; a world-action model additionally uses a world-model or video-model foundation to represent or predict physical dynamics.

Whether “WAM” becomes a broadly adopted industry category remains to be seen, and NVIDIA has a commercial interest in promoting the terminology. The idea underneath it is more important than the acronym: robotics AI is beginning to move from recognising what is in front of a machine towards predicting what will happen next.

The language around models themselves

Open-weight AI

An open-weight model makes its trained parameters available for others to download or operate on their own infrastructure.

That can provide considerably more control than a closed model available only through an API. Developers may be able to inspect, adapt, fine-tune and deploy the model without sending every request to the original provider. OpenAI, for example, describes its gpt-oss systems as open-weight reasoning models designed to run on infrastructure controlled by the user.

But open weight is not automatically the same as open source.

The Open Source Initiative's Open Source AI Definition requires broader freedoms and access to information and components needed to study and modify the system, including relevant training-data information and code. A company releasing weights alone does not necessarily satisfy that standard.

Read the Open Source Initiative’s Open Source AI Definition

Mixture of Experts — MoE

Mixture of Experts, or MoE, is a neural-network architecture in which only selected portions of a much larger model are activated for a particular input.

The model contains multiple expert components and a routing mechanism that decides which ones should participate. This makes it possible for a model to contain a very large number of total parameters without using all of them for every token.

That distinction explains why quoted parameter counts can become misleading. A trillion-parameter MoE model does not necessarily perform a trillion parameters' worth of computation for every request.

MoE is not a new idea, but it has become increasingly relevant as developers pursue larger models without proportionally increasing inference cost.

Distillation

Distillation trains a smaller “student” model to reproduce useful behaviour from a larger “teacher” model.

The aim is usually to retain as much capability as possible while reducing size, latency, memory requirements or inference cost. Google describes distillation as training a smaller model to emulate the predictions of the original, with the usual trade-off that the smaller system may not match the teacher perfectly.

Distillation has become strategically important because the AI race is no longer only about producing the largest possible frontier model. There is enormous commercial value in transferring enough of that capability into cheaper systems that can run at massive scale or on more modest hardware.

Inference

Inference is what happens when a trained AI model is actually used.

Training changes the model's parameters. Inference takes those learned parameters and processes new inputs to produce outputs. Every chatbot response, generated image, model classification or reasoning run therefore involves inference.

The term increasingly appears in discussions about the economics of AI because the industry is moving from enormous one-off training runs towards enormous volumes of daily model use. Agents amplify that change because one user request can generate repeated rounds of inference.

Understanding inference also makes phrases such as “inference-time compute”, “inference cost” and “inference accelerator” much easier to parse.

Two AI culture terms that escaped the laboratory

Vibe coding

Vibe coding describes building software largely by telling an AI what you want and allowing the model to produce and revise the code through natural-language interaction.

Andrej Karpathy coined the phrase in early 2025 with a deliberately loose meaning: a style of programming in which the developer largely surrendered to AI-generated changes and sometimes stopped reading the code closely. The phrase has since broadened and is now frequently used for AI-led software development generally; Google's own machine-learning glossary acknowledges that expansion in meaning.

The distinction still matters. Using an AI coding assistant while carefully reviewing its output is not quite the same proposition as accepting generated software because it appears to work.

AI slop

AI slop is low-quality generative-AI output produced cheaply and at scale, usually with more emphasis on volume than usefulness or craft.

The phrase began as internet slang but has become sufficiently established that Google's developer glossary now defines it explicitly as generative output that favours quantity over quality.

Its importance is cultural rather than architectural. Generative AI dramatically reduces the marginal cost of producing text, images, video and software. That does not reduce the cost of judgement, originality or taste by the same amount.

AI slop is therefore the vocabulary of abundance without discrimination: evidence that producing content and producing something worth consuming are different problems.

What this new AI language tells us about where the industry is going

The striking feature of the 2026 AI glossary is how few of its most consequential terms describe a chatbot answering a question.

Agents concern action. MCP and A2A concern connections. Context engineering concerns what intelligence has available while it works. Memory and continual learning concern what survives from previous experience. Evals and deterministic controls concern whether increasingly autonomous systems can be trusted. World models and physical AI concern what happens when intelligence leaves the browser and begins interacting with environments.

Even the architecture vocabulary is changing. Test-time compute suggests that capability can be increased after training by allocating more computation to the problem in front of the model. Recurrent depth and latent reasoning explore ways to perform more of that computation internally. The resulting safety debate concerns whether humans will continue to understand enough of those processes to supervise them.

That is the larger transition behind the jargon.

The first generative-AI boom was principally about what machines could produce. The vocabulary emerging now is about what machines can do, what they can remember, how they can coordinate, how they can change, and what happens when their decisions acquire consequences outside the model itself.

Those terms are likely to keep changing. The direction they describe is becoming much harder to miss.

Frequently asked questions

What are the most important new AI terms to know in 2026?

The most consequential emerging terms include agentic AI, AI agents, MCP, A2A, context engineering, AI memory, reasoning models, test-time compute, recurrent depth, continual learning, world models, physical AI and open-weight AI. They matter because they describe the industry's transition from standalone generative models towards AI systems that can reason for longer, access tools and information, act autonomously and interact with physical environments.

What is the difference between agentic AI and an AI agent?

An AI agent is an individual system capable of pursuing a goal through reasoning and actions. Agentic AI is the broader category or characteristic describing systems built around that kind of autonomy. A platform could therefore contain several AI agents while being described overall as an agentic-AI system.

What is the difference between MCP and A2A?

Model Context Protocol primarily standardises how AI applications connect to tools, data and external systems. Agent2Agent primarily standardises communication and collaboration between agents. An agent could use MCP to access tools while simultaneously using A2A to communicate with another agent.

Is context engineering replacing prompt engineering?

Not exactly. Prompt engineering remains part of context engineering. The difference is scope: prompt engineering concentrates on the instructions given to a model, whereas context engineering considers the entire information state available during inference, including prompts, retrieved data, memory, tools, previous messages and intermediate work.

Is AI memory the same as continual learning?

No. An application can store information externally and retrieve it later without changing the model itself. Continual learning generally concerns a system acquiring or consolidating new capabilities or knowledge over time. Memory can make an AI behave as though it has learned from experience while leaving its underlying model parameters unchanged.

What is opaque recurrence?

Opaque recurrence is a recently popularised term for internal recurrent reasoning that leaves fewer human-readable reasoning traces. It is closely associated in current reporting with recurrent-depth architectures. Recurrent depth is the more useful technical term; “opaque recurrence” emphasises the possible safety consequence when internal computation becomes harder to monitor.

Is an open-weight AI model open source?

Not necessarily. Open-weight means that trained model parameters are available. The Open Source Initiative's definition of Open Source AI requires broader freedoms and access to information and components necessary to study, use, modify and share the system. A weights release alone may therefore fall short of a full open-source definition.

What is physical AI?

Physical AI refers broadly to artificial-intelligence systems that perceive, reason about and act within physical environments. Robots, autonomous vehicles and other intelligent machines are common examples. The term overlaps with embodied AI, although embodied AI has a broader research tradition concerned with intelligence situated in and interacting through a body and environment.

Will this AI glossary need updating?

Almost certainly. AI terminology is moving unusually quickly because model architectures, agent infrastructure and deployment practices are still evolving. Some current terms will become standard vocabulary, some will change meaning and others may disappear altogether. The useful approach is therefore to treat an AI glossary as a living reference rather than a fixed dictionary.