Glossary
866 concepts from the agent era explained plainly: what each means, how it works, who it affects, the common misconception, and where it appears in Agent Age news.
Kezdd itt
The key basic concepts: from the general to agents and specialist fields.
- artificial intelligence
- machine learning
- deep learning
- Neural network
- algorithm (algorithm)
- AI-modell (AI model)
- foundation model (foundation model)
- large language model (large language model)
- transformer (transformer)
- generative AI
- multimodal model
- reasoning model
- chatbot
- parameter
- training
- training data
- fine-tuning
- reinforcement learning
- inferencia (inference)
- prompt
- token
- context window (context window)
- hallucination
- benchmark
- frontier modell
- open-weight model
- computing capacity (compute)
- GPU (graphics processing unit (GPU))
- data center
- general AI (artificial general intelligence)
In the news now
OpenAI
OpenAI is an American artificial intelligence company that develops and sells ChatGPT and the GPT language models behind it.
agent
angolul: AI agentAn agent is an artificial intelligence program that not only responds but independently plans, uses tools, and completes tasks in multiple steps to achieve a given goal.
Anthropic
Anthropic is an American artificial intelligence company developing the Claude language models and AI agents built on them, with a strong focus on safety.
code repository
A code repository is a shared, organized location for all files of a program that records every modification: who changed what and when.
repo
angolul: repositoryA repository is a shared storage for all files in a project that preserves every recorded change, allowing any previous recorded state to be retrieved and restored.
expenditure
angolul: releaseA release is a finalized, version-numbered state of a software or AI model that the creators officially hand over to users.
Microsoft
An American software and cloud giant that offers the Windows and Office suites, the Copilot AI assistant, and enterprise tools for AI agents.
Nvidia
Nvidia is an American chip design company whose graphics processors run a significant portion of today's AI systems, from chatbots to AI agents.
data center
A data center is a building where many computers, sometimes thousands, operate in one place: this is where data for internet services and AI is stored and processed.
Claude
Claude is the family of artificial intelligence models and conversational assistant from the US company Anthropic, which understands and writes text, answers questions, codes, and performs tasks independently.
An American technology giant known for its search engine and one of the largest developers and providers of artificial intelligence with the Gemini model family.
initial public offering
An initial public offering is the step where a company previously owned by a closed circle first offers its shares for purchase by anyone on the stock exchange.
Basics
AI fever
angolul: AI hypeAI hype is the heightened expectation and frenzy that portrays artificial intelligence as more powerful, faster, and closer than it actually is today.
AI safety
AI safety is concerned with making sure artificial intelligence does what we expect of it and causes no harm, either by accident or in the wrong hands.
AI-based search
angolul: AI search / AI answersA type of search where the AI does not give a list of links but, based on the sources it finds, writes a complete, ready-made answer to your question.
AI-modell
angolul: AI modelAn AI model is a program trained on countless examples that responds, predicts, recognizes, or creates new content based on learned patterns.
foundation model
angolul: foundation modelA foundation model is a general-purpose AI model trained on vast amounts of data, serving as a base for various applications from chatbots to image generators.
algorithm
angolul: algorithmAn algorithm is a finite sequence of unambiguous steps that, when followed precisely, yields a result within finite time for any allowed starting data of a given task type—whether executed by a human or a computer.
alignment
angolul: AI alignmentAlignment means that artificial intelligence operates according to people's true intentions and values, not just fulfilling what was literally asked of it.
artificial intelligence
A collective term for computer systems that solve tasks requiring human thought: they understand text, recognize images, prepare decisions, and learn from examples.
artificial superintelligence
A hypothetical artificial intelligence that would vastly surpass the best humans in virtually every intellectual task; it does not exist yet.
benchmark
A benchmark is a standardized set of tasks and evaluation method used to measure the performance of different AI models on the same scale, ensuring comparable results.
chatbot
A chatbot is a program with which you can converse in writing or speech using everyday language, and which answers your questions immediately in a human style.
computing capacity
angolul: computeComputational capacity is the machine "muscle power" – chips, servers, and electricity – required to train and operate an AI model.
context engineering
The conscious planning of exactly what information, instructions, and tools the AI receives for a task, no more and no less.
data center
A data center is a building where many computers, sometimes thousands, operate in one place: this is where data for internet services and AI is stored and processed.
deep learning
Deep learning is a method of machine learning in which a multi-layered artificial neural network learns on its own to recognize patterns from a vast number of examples.
deepfake
A deepfake is an image, audio, or video created or altered with artificial intelligence that appears deceptively real, typically showing a real person in a situation that never occurred.
diffusion model
An AI model that starts from random noise and creates new images, videos, or audio by cleaning it step by step.
eval
angolul: evaluation (eval)An eval is a systematic test that measures how well, reliably, and securely an AI model or agent performs on pre-defined tasks.
few-shot prompt
angolul: few-shot promptingAn instruction where you show the artificial intelligence a few worked-out example cases so it can solve your task in the same style and format based on them.
fine-tuning
Fine-tuning is the further training of an already trained AI model on specific, targeted examples to better handle a particular task, style, or field.
frontier modell
A frontier model is the most advanced, most capable general-purpose AI model at a given moment, representing the cutting edge of technology, i.e., the current limit.
general AI
angolul: artificial general intelligenceA theoretical artificial intelligence that can solve almost any intellectual task at least as well as a skilled human.
generative AI
Generative AI is a type of artificial intelligence that creates new content upon your request: text, images, audio, video, or source code.
GPU
angolul: graphics processing unit (GPU)A GPU is a computer chip that performs many simple calculations simultaneously, making it today's most important tool for training and running artificial intelligence.
grounding
Grounding means that the AI bases its answer on verifiable sources—documents, databases, fresh search results—rather than solely on its trained memory.
hallucination
Hallucination is when artificial intelligence confidently states something untrue or simply made up in a plausible form.
in-context learning
A language model infers a new task on the fly from examples and instructions entered into the conversation, without retraining.
inferencia
angolul: inferenceInference is the moment when a trained AI model works: generating an answer, image, or decision from your query or other input.
small language model
angolul: small language modelA text-understanding and text-writing AI model that is much smaller than large models, allowing it to run cheaply and quickly on phones, laptops, or corporate machines.
context window
angolul: context windowThe context window is the amount of text an AI model can see and consider at once when answering you.
machine learning
Machine learning is a method by which a computer learns how to solve a task not from pre-written rules but from many examples.
multimodal model
An artificial intelligence model capable of interpreting not only text but also images, audio, or video, and often responding in multiple formats.
large language model
angolul: large language modelAn artificial intelligence program trained on huge amounts of text that interprets your question or request and responds in human language with coherent text.
Neural network
A neural network is a computer model that learns to recognize patterns from many examples, loosely inspired by the network of neurons in the human brain.
next-token prediction
The core operation of a language model is to estimate, based on prior text, the likelihood of each word fragment, select one, and build its response piece by piece.
open-source AI
Artificial intelligence whose model, and ideally the details of its training, is published by its creator for free use, allowing anyone to download, examine, modify, and run it on their own machine.
open-weight model
An AI model whose trained parameters (weights) are published by its creator, allowing anyone to download, run on their own machine, and further fine-tune it according to the license terms.
parameter
A parameter is a number inside an AI model, set by the model during training; billions of them together carry its knowledge.
prompt
A prompt is the request or instruction, usually text but sometimes a spoken sentence or image, that you give to artificial intelligence to get a response, write something, create an image, or perform a task.
prompt caching
Prompt caching means the AI provider remembers the repeating parts of requests, so it does not have to reprocess them every time, resulting in cheaper and faster responses.
RAG
angolul: retrieval-augmented generationA method in which artificial intelligence, before answering, retrieves the document passages relevant to the question and formulates its answer based on them.
reasoning effort
A setting that determines how much the AI 'thinks' before answering: more thinking yields a more thorough but slower and more expensive response.
reasoning model
A language model that reasons step-by-step before answering, making it slower but more reliable for complex questions.
recursive self-improvement
The process in which an AI system improves its own capabilities, and the better version then develops itself even more effectively, cycle after cycle.
reinforcement learning
A machine learning method where the program discovers good behavior through trial and error based on rewards and penalties.
scaling laws
Empirical correlations about how much an AI model's performance improves when increasing its size, the amount of training data, and computing capacity.
structured output
Structured output is an AI response that is not free text but data in a predefined, fixed format, allowing another program to process it directly.
sycophancy
Sycophancy is the phenomenon where artificial intelligence says what you want to hear rather than what is true or useful.
synthetic data
angolul: synthetic dataArtificially generated data that mimics the characteristics of real data but does not originate from real people or events.
token
A token is a text fragment (word, word fragment, or punctuation mark) into which the language model splits text for processing and generation, and which is also used to measure usage.
training
Training is the process by which an AI model learns to perform its task through countless examples, trials, and corrections.
training data
Training data is the text, images, audio or other examples from which an artificial intelligence model learns patterns before you start using it.
transformer
angolul: transformerThe transformer is the neural network architecture that examines every part of a text simultaneously in relation to each other; almost all modern chatbots and language models are built on this.
usage limit
angolul: rate limit / usage limitA usage limit sets how much of an AI service you can use in a given period; if you exceed it, you have to wait or pay more.
world model
A world model is an AI system's internal picture of how its environment works, which lets it predict what will happen after its next action.
Language models
alignment
angolul: AI alignmentAlignment is the work done to ensure artificial intelligence pursues the goal people actually intend, rather than just fulfilling the instruction literally.
batch processing
Submitting many AI requests at once in a batch, where the response arrives within hours rather than immediately, in exchange for a much lower price.
chain of thought
Chain of thought is the method where a language model describes step-by-step how it reaches the solution before giving the final answer.
closed model
angolul: closed model / proprietary modelAn AI model whose internals, i.e., its weights, are not released by the developing company: you can typically use it only through its service or partners, under its own terms.
coding model
A language model trained to write, fix, and explain program code, capable of creating working code segments from requests phrased in everyday language.
Constitutional AI
A training method in which a language model evaluates and corrects its own answers based on a described set of principles, the 'constitution', to learn safe behavior.
context degradation
angolul: context rotContext degradation is the phenomenon where a language model works less accurately as more text accumulates in the conversation or task.
Distillation
Distillation is a process where the knowledge of a large, expensive AI model is taught to a smaller, faster, and cheaper model.
DPO
angolul: Direct Preference OptimizationA fine-tuning method where a language model learns from response pairs which answer humans prefer, without a separate reward model.
embedding
Embedding converts text, images, or audio into a sequence of numbers so that items with similar meanings have number sequences close to each other.
emergent capabilities
angolul: emergent abilitiesCapabilities that no one specifically trained the language model on still emerge when the model grows large enough.
figyelemmechanizmus
angolul: attention mechanismThe attention mechanism is the method by which an AI model, such as a language model, decides which parts of the input (for example, a text) are important for interpreting a specific part, such as a word.
function calling
angolul: function calling / tool callingFunction calling is the capability by which a language model not only writes text but also requests an external program or tool to perform a task.
jailbreak
A jailbreak is a cleverly phrased request that persuades an AI to break its own safety rules and produce prohibited content.
knowledge cutoff
It is the date until which a language model's training data extends; the model does not know anything about what happened after that.
KV-cache
angolul: KV cache (key-value cache)The KV cache is the language model's temporary scratchpad: it keeps the partial calculations for text already processed, so the model does not have to recalculate everything for every new word.
LLM as judge
angolul: LLM-as-a-judgeAn evaluation method where a language model scores or rates the responses of another AI system based on predefined criteria, replacing human judges.
LMArena
A public website where users anonymously choose between responses from paired AI models, generating a continuously updated ranking based on the votes.
long context
Long context means that the language model can consider a very large amount of text at once—sometimes equivalent to several books—in its response.
LoRA
angolul: Low-Rank AdaptationA resource-efficient fine-tuning method that leaves the large AI model untouched and trains only a small, interchangeable add-on for the new task.
mechanistic interpretability
A research field that reverse-engineers the internal workings of artificial intelligence: it looks for which parts of the model produce the answer, and through what computations.
model card
A model card is a short, public data sheet for an AI model: it describes its purpose, training data, performance on tests, and what it should not be used for.
model selector
angolul: model router / model pickerThe model selector decides which AI model will answer your question: either you choose from a menu, or the system chooses automatically for you.
MoE
angolul: mixture of expertsA model architecture in which most of the neural network consists of many smaller, specialized parts, and only a few of them work on each tiny piece of text (token).
language model
angolul: language modelA computer program that has learned from vast amounts of text which words usually follow one another, and based on this, can process and write text.
Ollama
Ollama is a free, open-source program that lets you download and run open-weight language models on your own computer with a few simple commands or in a basic chat window.
open-source model
An AI model whose weights, and ideally its code and training details, are made freely downloadable by its creator, allowing anyone to run it on their own machine, examine it, and adapt it for their own purposes within the license framework.
post-training
Post-training is the second phase of language model training, in which a raw, text-completing model becomes a helpful, instruction-following, and safely behaving assistant.
pre-training
Pretraining is the long, costly first learning phase in which an AI model acquires general language and world knowledge from vast amounts of data, primarily text in the case of language models.
prompt engineering
Prompt engineering is the art of phrasing your request to a language model so that you get exactly the answer you need.
prompt injection
An attack in which someone slips an instruction into the text read by a language model so that the model follows the attacker's instruction instead of the developer's or user's intent.
prompt leakage
Prompt leakage occurs when an AI chatbot reveals the hidden instructions given to it by its operator in the background, which the user should not see.
quantization
Quantization is the storage of AI model numbers in a coarser, more space-efficient format to make the model smaller and faster, at the cost of slight quality degradation.
refusal
Refusal occurs when a language model does not fulfill a request because it deems it harmful, illegal, or contrary to its rules.
system prompt
angolul: system promptThe system prompt is the background instruction set that defines how the AI should behave, what role to play, and what rules to follow before the conversation even starts.
RLHF
angolul: reinforcement learning from human feedbackA training method where humans evaluate the language model's responses, and the model learns from these evaluations what kind of response we consider good.
sampling
Sampling is the step where the language model selects one word from the possible next words based on probabilities, partly randomly.
scaffolding
Scaffolding is the software framework around the language model that provides it with tools, memory, and a workflow to perform tasks independently.
speculative decoding
Acceleration method: a small, fast model guesses the next parts of the text, and the large model only checks and corrects them if necessary.
streaming
Streaming means the AI response does not arrive all at once but appears continuously on the screen in small chunks as the model generates it.
supervised fine-tuning (SFT)
Further training of an already trained language model on question-answer pairs to perform a specific task or style more reliably.
temperature
Temperature is a setting that determines how predictable or how varied and surprising the language model's answer to the same question will be.
test-time computation
angolul: test-time computeThis is the computational work the language model performs while answering your question: the more it can 'think', the better the answer may be.
thinking token
angolul: reasoning tokens / thinking tokensA thinking token is the text unit a language model uses before the final answer to reason through the task step by step, for which payment is usually required.
token-based pricing
angolul: per-token pricingA pricing scheme where the AI provider bills based on the amount of text processed, meaning the number of input and generated tokens.
tokenization
Tokenization is the step where a program breaks text into small, numbered pieces—tokens—because the language model can only work with such units.
top-p
The top-p setting determines from how large a pool of the most likely candidates the language model can sample the next word.
transformer
The transformer is a neural network architecture that examines every part of a text simultaneously in relation to each other; today's large language models are built on this.
VLM
angolul: vision-language modelAn AI model that understands both images and text simultaneously, so it can say what is visible in a photo and answer questions about it.
weight
A weight is a number in an artificial neural network that indicates how much a connection matters; the model's knowledge resides in these numbers.
zero-shot prompting
A request to a language model where you only describe the task without providing examples, and the model solves it using its existing knowledge.
Agents
agent
angolul: AI agentAn agent is an artificial intelligence program that not only responds but independently plans, uses tools, and completes tasks in multiple steps to achieve a given goal.
agent identity
Agent identity is the AI agent's own verifiable digital identity, revealing who it is, on whose behalf it acts, and what it is authorized to do.
agent memory
Agent memory is the stored knowledge from which the AI agent can later retrieve what it learned, what it did, and what you asked of it.
agent mode
The mode of an AI tool in which it not only answers your question but also independently completes assigned tasks in multiple steps.
agent orchestration
Agent orchestration is the coordination of multiple AI agents: someone or something assigns subtasks, monitors the sequence, and merges the results.
agent-benchmark
angolul: agent benchmarkThe agent benchmark is a standardized set of tasks used to measure how often an AI agent independently solves real-world-like tasks in multiple steps.
agent-ciklus
angolul: agent loopThe agent cycle is the repeating loop in which an AI agent thinks, acts, reviews the result, and decides again until the task is completed.
agent-keretrendszer
angolul: agent frameworkAn agent framework is a software toolkit that allows developers to assemble autonomous AI agents without having to write everything from scratch.
agent-platform
angolul: agent platformA software environment where AI agents can be built, connected to tools, run, and monitored without having to write everything from scratch.
Agent2Agent (A2A)
angolul: Agent2Agent protocolAn open standard that provides a common language for AI agents from different companies to find each other, hand over tasks, and work together.
agentic advertising
Advertising where campaigns are planned, purchased, and refined independently by AI agents, with humans providing only the goal and framework.
Agentic AI Foundation
A neutral organization operating under the Linux Foundation that maintains common, open standards and tools for AI agents to ensure they are not owned by a single company.
agentic checkout
An online purchase where you do not click through the order and payment yourself, but an AI agent acting on your behalf handles it at the online store.
agentic commerce
A purchase in which an AI agent, on your behalf, searches for, compares, orders, and even pays for a product or service.
agentic payment
angolul: agentic paymentsA payment that you do not initiate manually, but one that an AI agent executes based on your instruction within the limits you set in advance.
agentic AI
angolul: agentic AIArtificial intelligence that not only answers your question but also carries out the goal it is given on its own, in several steps, using tools.
agentikus munkafolyamat
angolul: agentic workflowA working method where AI does not give a single answer but plans in multiple steps, uses tools, checks the result, and corrects it.
agentikus RAG
angolul: agentic RAG (agentic retrieval-augmented generation)An AI solution where an agent itself decides what, where, and how many times to search sources before formulating an answer.
AGENTS.md
AGENTS.md is a simple text file in a software project folder that describes how coding AI agents should work on that project.
AI wallet
A digital wallet from which an AI agent can pay on your behalf, within limits you set, autonomously.
AI assistant
angolul: AI assistantAn artificial intelligence-based program with which you can converse in everyday language, and which responds to your requests, writes text, summarizes, or performs simpler tasks for you.
allowlist
An allowlist is a precompiled list of actions an AI agent may perform without separate permission; anything not on it cannot be done, or requires prior approval.
Antigravity
angolul: Google AntigravityGoogle's developer program in which AI agents independently plan, write, and test software, while humans assign tasks and review their work.
asset definition
angolul: tool definition / tool schemaA tool definition is a short, machine-readable description of what a tool is for, when to use it, and what data must be provided for it.
authorization management
angolul: permission managementRegulating what an AI agent can access, what it can do autonomously, and what requires prior human approval.
automated researcher
An AI agent that largely independently carries out the steps of a research project—from posing the question, through gathering sources and conducting experiments, to writing the results.
autonomy
Autonomy shows how much an AI agent can decide and execute independently without seeking human approval at every step.
background agent
An AI agent that independently completes the assigned task in the background while you focus on something else, then reports the result.
browser agent
An AI agent that manages a web browser on your behalf: it opens pages, clicks, fills out forms, and completes assigned tasks in multiple steps.
checkpoint
A checkpoint is a save point in an AI agent's work: if something goes wrong, you can roll back here and try again from this point.
Claude Cowork
Anthropic's desktop AI agent, which carries out office tasks on its own, in multiple steps, in folders you designate on your computer, rather than just giving advice.
CLAUDE.md
A simple text file where you describe the project rules and details to the Claude Code agent, which the agent reads at the start of every session.
code interpreter
angolul: code interpreter / code execution toolA tool that allows AI not only to write code but also to run it in an isolated, secure environment and continue working based on the real output.
coding agent
An AI program that independently writes, runs, tests, and fixes code based on a task formulated in everyday language, in multiple steps, until it reaches a working solution or gets stuck.
computer use
An AI capability where the model handles the computer like a human: it looks at the screen, clicks, types, and scrolls.
context compression
angolul: context compactionThe AI agent replaces earlier parts of a long conversation or work session with a short summary to leave room for continuation.
Customer service agent
An AI program that, in customer service, not only answers your question but also handles your case: for example, modifying your order or initiating a refund.
deep research
An AI mode in which an agent searches and reads on the web independently for minutes, then writes a source-backed summary report on your question.
Devin
Devin is Cognition's coding agent, which independently plans, writes, tests, and then hands over a software development task to a human for review.
digital colleague
angolul: digital coworkerAn AI agent that does not just answer individual questions but works as a colleague with its own scope of tasks and access, independently executing workflows within a team.
digital twin
A digital twin is a computerized replica of a real object—machine, building, process, or even a person—continuously updated with live data, allowing experiments without risking the real thing.
excessive agency
Security risk: an AI agent is granted more tools, access, or autonomy than needed for its task, so a mistake or deception could cause real harm.
function calling
The capability by which a language model not only writes text but requests the program in a structured format to execute a predefined operation.
Gemini CLI
Gemini CLI is Google's open-source AI agent that works in your computer's command line: it reads files, writes code, and executes commands on your request.
guardrail
A guardrail is an embedded rule or check that limits an AI system or agent from performing harmful, prohibited, or out-of-scope actions.
harness
angolul: agent harnessA harness is the software framework around an AI model that provides tools, memory, and rules, turning the model into an independently working agent.
headless mode
A mode in which the AI agent runs in the background without a screen or chat window and without human involvement, delivering only the final result.
hook
A hook is a pre-set, automatic step that always runs at a specific point in an AI agent's workflow, regardless of what the agent is thinking.
HR-agent
angolul: HR agent (recruiting agent)An AI agent that independently performs routine recruitment tasks: searching for candidates, screening resumes, answering questions, and scheduling interviews.
human-in-the-loop
An operating mode in which the AI pauses before important steps and continues work only after human approval.
interoperability
Interoperability means that AI agents, programs, and data sources from different manufacturers can work together under common rules without separate conversion.
Jules
Jules is Google's coding agent, which works independently in the cloud and in the background on software code, then presents the result to you for approval.
Know Your Agent (KYA)
The practice of identifying AI agents: verifying who stands behind the agent, on whose behalf it acts, and what authorization it has received.
LangChain
angolul: LangChain / LangGraphLangChain is an open-source developer toolkit for building AI applications and agents from language models, data sources, and tools; LangGraph is its process-control counterpart.
lethal trifecta
The dangerous situation where an AI agent simultaneously accesses your private data, reads untrusted content, and can send data out to the outside world.
long-horizon task
A complex task that an AI agent can only perform by working independently over a long period through many sequential steps.
Manus
Manus is a general-purpose AI agent that independently completes a task in the cloud on its own virtual computer and delivers a finished result.
Model Context Protocol (MCP)
An open standard that provides a unified interface for AI assistants and agents to access external programs, data, and services.
multiagent rendszer
angolul: multi-agent systemA system in which multiple agents divide tasks among themselves and typically solve problems collaboratively, building on each other's work.
munkamenet
angolul: sessionA session is a continuous conversation or task with the AI, throughout which the system remembers what was said and done.
n8n
n8n is a visual workflow automation tool with which you can connect applications and build AI agents by linking boxes, requiring little programming.
observability
Observability means that the operation of a system—here an AI agent—is traceable from its recorded data: every step is reviewable, so it can be determined afterward what it did, in what order, and where it made a mistake.
OpenClaw
An open-source, locally run personal AI agent that you instruct via a chat application and which independently handles tasks on your behalf.
orchestrator–worker minta
angolul: orchestrator-worker patternA way of organizing work in which an orchestrating AI agent splits the job into subtasks, assigns them to executing agents, and then merges their results into a single answer.
OSWorld
OSWorld is a test suite measuring how well an AI agent can independently perform everyday computer tasks on a real operating system with real software.
permission mode
A setting that defines what an AI agent can do independently and what requires your prior permission.
personal agent
An AI program that, on behalf of a single person and with access to their data and accounts, independently handles daily tasks.
planning
Planning is the ability of an AI agent to break down a larger goal into steps, order them, and adjust as needed during execution.
proactive agent
An AI agent that does not wait for your instruction but monitors the situation in the background and speaks or acts on its own when it sees you need it.
ReAct
ReAct is an operational pattern where the AI agent alternates between thinking and acting: it weighs options, uses a tool, checks the result, then weighs again.
rogue agent
An AI agent that steps outside its set boundaries: it does something other than what it was assigned, potentially causing harm.
sandbox
A sandbox is an isolated, safe test environment where a program or AI agent can work without harming the real system even if an error occurs.
self-reflection
The AI agent reviews its own work, searches for errors, and makes improvements based on lessons learned before delivering a final result.
shopping agent
An AI program that searches for products on your behalf, compares offers, and, if permitted, handles the order and payment for you.
skill
angolul: agent skillA skill is a packaged guide that the AI agent only retrieves when it needs it for a specific type of task.
slash command
A short command starting with a slash, with which you issue a pre-prepared, longer instruction or built-in function to an AI agent or chatbot with a single word.
specification-driven development
angolul: spec-driven developmentA software development method where a detailed written description of what the program must do is created first, and the AI agent writes the code based on this.
subagent
A subagent is an assistant agent launched by another AI agent for a well-defined subtask, which returns only the result at the end.
szoftverrobot (RPA)
angolul: robotic process automationA software robot is a program that clicks, types, and copies data between computer programs according to pre-written rules, instead of a human.
task decomposition
Task decomposition is when an AI agent breaks down a large, complex task into smaller, individually executable steps and solves them sequentially or in parallel.
Terminal-Bench
An open test series measuring how well an AI agent can independently perform real, multi-step computer tasks via the command line.
tokenkeret
angolul: token budget / thinking budgetThe token limit is a predefined upper bound on how much text an AI model or agent can process and how much it can "think" on a task.
tool search
An AI agent does not retain descriptions of all its tools in memory but retrieves and loads only the few it needs during a task.
tool use
Tool use is the capability by which an AI model calls external programs (search engine, calculator, calendar, database) for help, so it does not just answer from memory but retrieves fresh data, calculates precisely, or performs an operation.
trajectory
angolul: agent trajectoryA trajectory is the complete recorded path of an AI agent during a task: its written-out reasoning, steps, tool usage, and the results it received, in sequence, from start to finish.
voice-based agent
angolul: voice agentAn AI agent with which you converse in real time, which not only answers but also independently handles what you request during the conversation.
workflow automation
Handing over repetitive, multi-step tasks to software that completes the job using pre-defined rules or AI without human clicking.
YOLO mode
A setting in which an AI agent does not request permission before each step but executes everything autonomously without stopping.
Generative media
3D generation
During 3D generation, artificial intelligence creates a rotatable, spatial model from a text description or photograph, for example an object, figure, or building.
Adobe Firefly
Adobe Firefly is Adobe's family of artificial intelligence tools that generate images, video, audio, and graphics, integrated into Photoshop and other company software.
AI advertising
angolul: AI-generated advertisingAn advertisement whose image, video, audio, or text was partially or entirely generated by artificial intelligence, without a crew, photographer, or copywriter.
AI art
An image, music, video, or other work created partially or entirely by artificial intelligence, usually based on a human text prompt.
AI avatar
A human-like digital character created with artificial intelligence that delivers the given text with a lifelike face, voice, and lip movement in a video or live conversation.
AI content labeling
A visible label or hidden signal in a file indicating that an image, video, audio, or text was created or modified by artificial intelligence.
AI dubbing
A technology that translates speech from a video or audio recording into another language via machine processing and has it spoken in a voice resembling the original speaker.
AI image editing
Transforming an existing photo or image with artificial intelligence: you specify what should change with words or selection, and the program realistically redraws the image.
AI slop
AI slop is mass-produced, low-quality content generated by artificial intelligence, not intended for readers but for clicks, flooding the internet.
AI watermark
angolul: AI watermarkingAn AI watermark is a signal that is invisible to the eye but machine-readable, which an artificial intelligence system embeds in the images, audio, video or text it creates.
AI-generated content
Text, image, audio, video, or program code whose substantial part was not created by a human but by artificial intelligence, typically based on human instructions.
arccsere
angolul: face swapFace swapping is a process that uses artificial intelligence to replace a person's face in an image or video with another person's face, producing a realistic result.
CFG scale
angolul: classifier-free guidance (CFG) scaleThe CFG scale is a configurable number that determines how closely an image-generating AI follows your text prompt and how much freedom it is granted.
ComfyUI
A free, open-source program in which you assemble how artificial intelligence creates images, videos, or audio by connecting boxes with lines.
content provenance verification
A digitally signed data sheet attached to an image, video, or audio, showing who created or modified the content, with what, and how, and where subsequent tampering is detectable.
ControlNet
ControlNet is an add-on for image-generating AI that allows you to precisely define how the image is constructed using a sketch, pose, or outline.
deepfake detection
A collection of methods and tools used to determine whether a video, image, or audio recording was created or forged by artificial intelligence.
digital likeness
angolul: AI twin / digital doubleAn AI-generated copy of a real person that speaks with their face, voice and style in video or live conversation.
ElevenLabs
ElevenLabs is an artificial intelligence company and service that creates lifelike human voices from written text and can also clone existing voices.
flow matching
A teaching method where an AI that creates images, audio, or video learns how to reach the final content from random noise via the most direct path possible.
Flux
Flux is the image-generation model family of the German Black Forest Labs, which creates realistic images from text descriptions and can also edit existing images.
generative adversarial network
An AI method in which two neural networks compete: one generates fakes, the other tries to expose them, making the fakes increasingly realistic.
generative media
An image, video, audio, or music that was not recorded by a camera or microphone but generated by artificial intelligence, mostly based on text instructions.
GPT Image
GPT Image is OpenAI's image generation model family, the successor to DALL-E: it draws images or transforms existing ones based on text prompts, and is also available in ChatGPT.
image generation
Image generation is the artificial intelligence process that creates a completely new image from a text description or reference image in a few seconds.
image to video
angolul: image-to-videoAn AI method that creates a few seconds of video from a single still image: it imagines how the scene visible in the image would move.
inpainting
Inpainting is an image editing process where artificial intelligence fills or replaces a selected part of an image with new content that fits the surroundings.
karakterkonzisztencia
angolul: character consistencyThe ability of image- or video-generating AI to make the same character look recognizably the same across multiple images and scenes.
Kling
angolul: Kling AIKling is the video-generating artificial intelligence of the Chinese company Kuaishou, which produces short, typically 5–15 second, realistic moving images from text descriptions or still images.
latent space
The latent space is the model's internal, numerical map where similar images, sounds, or texts are positioned close to each other.
lip-syncing
angolul: lip syncAn AI procedure that retrospectively adjusts the lip movements of a person or figure in a video to match an audio recording, making it appear as if they are actually saying it.
music generation
Music generation is when an artificial intelligence creates new music, including vocals and lyrics, on its own based on a text description or other guidance.
Nano Banana
Nano Banana is the nickname for Google Gemini's image generation and editing model: it draws images from text prompts or transforms your existing photos.
negative prompt
A negative prompt is a specific instruction to an image- or video-generating AI in which you list what the finished work should not contain.
non-consensual intimate image
angolul: non-consensual intimate imageryAn image or video of someone that is nude or sexually explicit, created, forged, or shared without the subject's consent.
outpainting
Outpainting is an AI image editing technique that extends an existing image beyond its edges and invents the missing environment in a style-consistent way.
reference image
An image you provide alongside a text instruction to an image-generating AI to show the style, character, or composition you wish to see.
Runway
Runway is an American company and online service that uses artificial intelligence to create and edit short videos from text descriptions, images, or existing footage.
sampling step
angolul: sampling stepsThe number of sampling steps shows how many rounds a diffusion-based image-generating AI takes to clear the final image from the initial noise.
scaling up
angolul: upscalingUpscaling is increasing the resolution of an image or video so that the program – nowadays typically artificial intelligence – fills in the missing details.
seed
A seed is a starting number that determines what 'randomness' the generation begins with, most often for creating an image; the same seed with the same model, prompt, and settings typically yields the same result.
speech recognition
angolul: speech recognition / speech-to-textSpeech recognition is a technology that automatically converts spoken human language into written text, allowing the machine to "transcribe" what you say.
speech synthesis
angolul: text-to-speech, speech synthesisSpeech synthesis is a technology that converts written text into machine-generated speech resembling a human voice.
Stable Diffusion
The Stable Diffusion family of AI models that generate images from text descriptions, whose model files anyone can download and even run on their own computer.
style transfer
An AI method that preserves the content of an image, text, or audio but replaces its appearance with the style of another work or creator.
Suno
Suno is an artificial intelligence service that, within minutes, turns a text description of a few sentences into a finished song with vocals, instrumentation and lyrics.
synthetic media
An image, video, audio, or text that was partially or fully generated or altered by artificial intelligence, not by a camera, microphone, or human hand.
Veo
Veo is Google DeepMind's artificial intelligence model for creating videos, which generates short, lifelike moving images, equipped with sound in newer versions, from text descriptions or images.
video generation
Video generation is when artificial intelligence independently creates moving images from text descriptions, images, or other videos, without a camera or filming.
Virtual influencer
A fictional character created by computer or artificial intelligence who posts and advertises on social media as if they were a flesh-and-blood human.
voice cloning
Voice cloning is an AI procedure that can have any text spoken in a person's voice based on a short recording of their voice.
voice conversion
angolul: voice conversion / speech-to-speechAn AI method that makes a recorded or live speech sound like another person's voice while preserving the text, rhythm, and intonation.
Training and data
adatfal
angolul: data wallThe assumed point where AI model development slows down because high-quality, human-written new data for training runs out.
adathalmaz
angolul: datasetA dataset is a collection of related, organized data—texts, images, numbers—from which artificial intelligence learns or on which its knowledge is measured.
backpropagation
The computational procedure by which a neural network works backward from its error to calculate how much each internal setting needs adjustment to be more accurate next time.
batch size
Batch size indicates how many training examples the artificial intelligence model reviews at once before adjusting its settings once.
catastrophic forgetting
The phenomenon where an AI model learns something new and suddenly loses a large portion of its previously acquired knowledge.
Common Crawl
Common Crawl is a non-profit organization and its free, public web archive containing copies of billions of websites, serving as the foundation for training data for many AI models.
continual learning
A type of machine learning where the model learns new things step by step from sequentially arriving data, ideally without forgetting what it previously knew.
data augmentation
Data augmentation is a technique that creates new variants of existing training data through small modifications, so that the AI model can learn from a larger and more varied set of examples.
data cleaning
Data cleaning is the process of finding and correcting faulty, incomplete, duplicate or inconsistent data before it is used for analysis or artificial intelligence training.
data contamination
Data contamination occurs when an AI model has already seen the test tasks during training, making it appear better on the exam than it truly is.
data curation
Data curation is the work of selecting, cleaning, organizing, and documenting raw data so that it can reliably be used to train AI or to run AI systems.
data labeling
Data labeling is the work of adding the correct answer to images, texts, or audio recordings so that artificial intelligence can learn from it.
data mixture
The data mix shows from which sources and in what proportions the dataset is composed that an AI model learns from.
data poisoning
Data poisoning is an attack where someone intentionally smuggles false or manipulated data into an AI's training material so that the system later operates incorrectly or in the attacker's favor.
data usage for model training
angolul: use of data for model training (training data opt-in)The setting and rule that determines whether the AI provider can use your conversations, files, and content to further train its models.
differential privacy
A mathematical method that uses intentionally injected random noise to ensure that individual data cannot be inferred from a statistic or an AI model.
epoch
An epoch is a full learning cycle: the machine learning model goes through all examples used for training once.
federated learning
A training method in which the AI model learns locally on many devices or at many institutions, and only what it has learned is shared centrally, not the raw data.
generalization
Generalization is the AI model's ability to correctly apply what it has learned to new cases it never saw during training.
gradient descent
Gradient descent is the method by which an AI model fine-tunes its internal settings in small steps, gradually reducing its error.
GRPO
angolul: group relative policy optimizationA training method where the AI model provides multiple answers to a task and learns from those that perform better than the group average.
Hyperparameter
A hyperparameter is a setting that the model does not learn but is specified externally before training (by the developer or an automatic search program), determining how the model learns and how large it should be.
instruction fine-tuning
angolul: instruction tuningPost-training in which a language model practices on many thousands of instruction–response pairs to carry out a request rather than just continuing the text.
jutalommodell
angolul: reward modelAn auxiliary model that scores how good an AI response is, so humans do not have to rate every single response during training.
learning rate
The learning rate determines how large the steps are by which an AI model adjusts itself during training as it learns from its mistakes.
loss function
The loss function shows with a number how much the AI model erred; the goal of training is to make this number as small as possible.
machine unlearning
A procedure that removes the impact of certain data from an already trained AI model afterward without retraining the entire model from scratch.
memorization
Memorization is when an AI model literally remembers certain details of the training data and can later reproduce them.
model collapse
The phenomenon in which an AI model's quality degrades from generation to generation because, instead of human content, it increasingly learns from material produced by other models.
model weights
Model weights are the numbers in which an AI model stores the knowledge acquired during training; these determine what it answers to a question.
overfitting
Overfitting occurs when an AI model memorizes its training examples, performing well on them but poorly on new, unknown cases.
PEFT
angolul: parameter-efficient fine-tuningA family of methods that adapts a large AI model to a task by retraining only a fraction of its internal settings.
PPO
angolul: Proximal Policy OptimizationA training method in which AI learns from attempts based on rewards, but can only change its behavior slightly at a time.
preferenciaadat
angolul: preference dataTraining data that records which of two or more AI responses people judged better.
QLoRA
angolul: Quantized Low-Rank Adaptation (QLoRA)A cost-efficient fine-tuning method: the large language model is loaded in a compressed form, and only small additional layers are trained on it, so often a single graphics card is sufficient.
reward hacking
Reward hacking occurs when AI does not solve the actual task but finds a loophole to inflate its performance score.
RL environment
angolul: reinforcement learning environment (RL environment)An RL environment is an artificial training ground where the AI tries actions, receives points for them, and thus learns the task.
RLAIF
angolul: reinforcement learning from AI feedbackA training method in which the responses of a language model are evaluated not by humans but by an AI model, and the trained model learns from this feedback.
RLVR
angolul: reinforcement learning with verifiable rewardsA training method in which the AI model receives a reward only if its answer proves correct through machine-based, unambiguous verification.
self-supervised learning
A machine learning method in which the model learns without human labels: it creates tasks from raw data itself, typically by trying to guess hidden parts of the data.
structured data
Data organized into a predefined order – rows, columns, named fields – that the computer can search, sort, and aggregate precisely without interpretation.
supervised learning
A machine learning method in which the machine learns from examples with correct answers how to give a good answer on its own to new cases.
tesztadatok
angolul: test dataTest data consists of set-aside examples that the AI model never saw during training, on which its actual performance is finally measured.
training
Training is the process in which an AI model learns from countless examples, through gradual attempts and corrections, how to solve a task.
training data
Training data are the examples – texts, images, audio, numbers – from which artificial intelligence learns what it later knows.
training data summary
A public document in which a general-purpose AI model developer outlines, as required by EU regulations, on what content the model was trained.
training run
A training run is a single, end-to-end process of training an AI model, during which the model learns its task from the data.
transfer learning
A training method where an AI model transfers knowledge gained from a previous task to a new, related task, so it does not need to learn from scratch.
underfitting
Underfitting means an AI model is too simple or has learned too little, so it fails to recognize even basic relationships in the training data.
unstructured data
Data without a predefined, tabular (row-and-column) structure: free text, images, audio, video, email content, or scanned documents.
unsupervised learning
A type of machine learning where the program, without pre-defined correct answers, independently searches for groups, patterns, and outliers in the data.
validation set
The validation set is the reserved part of the data on which developers check and tune the model during training, but from which the model does not learn directly.
web scraping
Web scraping is when a program automatically reads through websites, instead of a person clicking through them, and extracts the data it needs.
Infrastructure
adatrezidencia
angolul: data residencyData residency means that a service stores and processes your data in a specific country or region—for example, within the European Union.
advanced packaging
A collective term for chip assembly processes that integrate multiple separate chips—such as a processor and memory—into a single, densely interconnected unit to work together faster and more efficiently.
AI energy demand
This is the amount of electricity consumed by training and daily use of artificial intelligence, mainly through machines operating in data centers.
AI gigafactory
A massive data center housing roughly one hundred thousand high-end AI chips, where the largest artificial intelligence models are trained and operated.
AI infrastructure
AI infrastructure is the totality of chips, data centers, networks, power supply, and software without which artificial intelligence could not learn and respond.
AI chip
angolul: AI chipA computer chip designed or used specifically for the massive, simultaneous calculations required by artificial intelligence.
AI-crawler
angolul: AI crawlerAn AI crawler is an automated program that browses and scans websites so that artificial intelligence can learn from their content or answer questions.
AI-gateway
angolul: AI gateway / LLM gatewayAn intermediary software gateway through which every request intended for an AI model passes, enabling centralized monitoring, control, logging, and accounting.
Amazon Bedrock
Amazon's cloud service through which companies can rent and integrate large AI models from multiple manufacturers into their own applications without their own servers.
API
angolul: application programming interfaceAn API is a predefined rule set that allows one program to request data or a service from another without needing to know its internal workings.
ASIC
angolul: application-specific integrated circuitAn ASIC is a chip designed for a single task or narrow range of tasks, so it performs that task much faster and with less energy than a general-purpose processor.
availability
angolul: uptime / availability (SLA)Availability shows what percentage of the time an IT system or online service is functioning and actually accessible.
AWS
angolul: Amazon Web ServicesAWS is Amazon's cloud service: you can rent computers, storage, and AI tools from it via the internet, and pay only for what you use.
cache
A cache is a temporary storage that keeps copies of frequently used data on hand so they do not need to be re-downloaded or recalculated every time.
chip export restriction
angolul: chip export controlsA state regulation that bans or requires a license to sell the most advanced AI chips and the machines needed for their production to certain countries.
chip manufacturing
Chip manufacturing is the industrial process in which tiny circuits are built layer by layer on silicon wafers – these become the processors and memories of phones, cars and AI systems.
cloud
angolul: cloud computingThe cloud is computing capacity—storage, processing power, software—that you do not keep on your own machine but use via the internet, typically for a rental fee or subscription.
contract manufacturer
angolul: semiconductor foundryA contract manufacturer is a chip factory that does not produce its own products but manufactures chips designed by other companies to order.
CoreWeave
CoreWeave is an American cloud provider that operates data centers filled with graphics processors needed for artificial intelligence and rents out their computing power to AI companies.
CUDA
CUDA is Nvidia's software system with which programs can use the graphics card for general-purpose computing, such as running and training artificial intelligence.
distributed training
Distributed training means that an AI model is trained not by a single machine, but by many processors or computers working in coordination at the same time.
edge computing
A solution where data is processed where it is generated – on the device or nearby – rather than in a distant data center.
EUV lithography
A chip manufacturing process that draws extremely dense, fine circuit patterns for the most advanced chips onto silicon using extremely short-wavelength ultraviolet light.
FLOPS
angolul: floating-point operations per secondFLOPS measures how many floating-point calculations a computer performs in one second – this is the computing power's "horsepower".
GGUF
GGUF is a file format that packages a language model into a single file so you can run it on your own machine without internet.
Google Cloud
Google's rentable computer infrastructure and service package on which companies run their programs, store their data, and use artificial intelligence over the internet.
GPU rental
angolul: GPU rental / cloud capacity leaseGPU rental means you do not buy the expensive computing chips needed for AI, but use them on a provider's machines for an hourly rate or under a long-term contract.
GPU-klaszter
angolul: GPU clusterA cluster of networked computers equipped with graphics processing units (GPUs) that work as a single giant machine to train and run artificial intelligence.
HBM memory
angolul: high-bandwidth memory (HBM)Extremely fast memory consisting of stacked memory chips, sitting next to AI chips and capable of moving huge amounts of data simultaneously.
hyperscaler
A hyperscaler is a giant technology company that operates massive data centers worldwide and mostly rents out its computing capacity to others, such as AI developers.
inference cost
Inference cost is the amount it costs to use an already trained AI model: answering each question consumes computing capacity, electricity, and thus money.
inferenciachip
angolul: inference chipA computer chip designed so that already-trained artificial intelligence can answer questions quickly, cheaply, and with low power consumption.
InfiniBand
InfiniBand is an extremely fast, low-latency networking technology that connects the machines of supercomputers and AI data centers so they can work as a single large machine.
Kubernetes
An open-source system that automatically starts, monitors, and restarts programs packaged in containers on many servers, and adjusts their number to the load depending on the settings.
latency
Latency is the time between sending your request and receiving the response – in other words, how long you have to wait for the machine.
liquid cooling
A cooling method that uses water or another liquid instead of air to dissipate heat from computer chips, especially those running AI servers.
local execution
angolul: local deployment / on-premisesIn local execution, the AI model runs on your own computer or your company's own server, not in a remote data center of an external provider.
memory chip shortage
The market situation where AI data centers buy up so many memory chips that fewer and more expensive ones remain for phones, laptops, and other devices.
Microsoft Azure
Microsoft Azure is Microsoft's cloud service: computers, storage, and AI models rentable over the internet, for which you pay based on usage.
MLOps
angolul: machine learning operations (MLOps)MLOps is the collection of methods and tools used to reliably deploy, monitor, and continuously update machine learning models.
model routing
A solution that automatically routes every question to the AI model expected to answer it well enough, yet as quickly and cheaply as possible.
model serving
Model serving is the background operation that keeps a trained AI model continuously available so you receive an answer to your query within seconds.
neocloud
A new type of cloud provider that rents out high-performance chips (GPUs) needed almost exclusively for artificial intelligence.
NPU
angolul: neural processing unitThe NPU is the part of the phone or laptop chip specifically designed for artificial intelligence calculations, performing them locally, quickly, and with low power consumption.
NVLink
angolul: Nvidia NVLinkNVLink is Nvidia's own high-speed connection that links multiple graphics processors (and their associated central processors) so they work almost like a single giant machine.
on-device AI
Artificial intelligence that works directly on your phone, laptop, or other device, without sending your data to a remote server.
power purchase agreement
A long-term agreement in which a buyer – an electricity supplier, trader, or large consumer, such as a data center – purchases electricity directly from a power plant producer at a pre-agreed price or pricing formula, locked in for years.
PUE
angolul: power usage effectivenessPUE shows how many times the total power consumption of a data center exceeds what the computers themselves use.
quantum computer
A computer that uses the laws of quantum physics to solve certain special tasks much faster than current machines.
reranker
A reranker is an AI model that reorders texts found by the search engine so that those truly answering your question appear first.
Rubin
angolul: Nvidia Rubin (Vera Rubin platform)Rubin is Nvidia's chip generation following Blackwell: a combination of processors and network elements for data centers, on which large AI models are trained and run.
semiconductor
A semiconductor is a material—mostly silicon—whose electrical conductivity can be controlled; chips used in every computer and artificial intelligence are made from it.
serverless architecture
angolul: serverless computingA cloud solution where the program runs and costs money only when it has work to do, while the provider fully manages the machines.
service outage
We speak of a service outage when an online service – such as an AI chatbot, online banking, or email – is temporarily unavailable or not functioning properly.
small modular reactor
A nuclear reactor prefabricated in a factory from components, much smaller than a conventional one, assembled on site, and able to power even a single industrial plant or data center.
sovereign AI
Artificial intelligence that a country or organization runs on machines under its own control, with its own data and according to its own rules, so that it does not depend on a foreign provider.
supercomputer
A supercomputer is a room-sized machine consisting of thousands of interconnected processors, performing calculations in hours that would take an average computer years.
throughput
Throughput indicates how much work a system performs in a given time, for example, how many requests or text segments it processes per second.
time to first token
The first token time measures the duration between sending your query and the appearance of the first text segment of the AI response.
tokens per second
It shows how many text fragments (tokens) a language model outputs per second, i.e., how fast its response scrolls before you.
TPU
angolul: tensor processing unit (TPU)The TPU is Google's own chip, designed specifically for artificial intelligence calculations, so it performs these tasks faster and with less power than a general-purpose processor.
Trainium
angolul: AWS TrainiumTrainium is Amazon's own custom-designed chip tailored for artificial intelligence, which it rents out via its cloud service (AWS) for training and running AI models.
vector database
A database storing numerical vectors derived from texts and images, enabling searches based on semantic similarity rather than exact word matches.
Vertex AI
angolul: Google Vertex AIVertex AI is Google's cloud service where companies can rent artificial intelligence models, adapt them to their own data, and build applications and agents on top of them.
virtual machine
A virtual machine is a computer simulated by software: it runs inside a real machine but behaves as if it were an independent, separate machine.
vLLM
An open-source server program that runs large language models on your own servers, serving many users at once quickly and efficiently.
VRAM
angolul: video RAM (GPU memory)VRAM is the graphics card's dedicated, high-speed memory; in AI, it determines how large a model can fit and run on your machine.
Security
adversarial attack
Intentional deception where someone uses carefully designed, often imperceptible modifications to trick artificial intelligence into making incorrect decisions.
agent hijacking
An attack in which someone uses hidden instructions to compel an AI agent to pursue their own goals instead of yours.
AI risk
AI risk is the chance that a flaw, misleading operation, or malicious use of an artificial intelligence system causes harm to people, companies, or society.
AI safety institute
A state professional institute that tests the most advanced AI models and examines the risks they may pose to society before or after they reach the market.
AI text detection
A procedure that estimates whether a text was written by a human or artificial intelligence – but it provides only a probability, not proof.
AI-based surveillance
angolul: AI surveillanceSurveillance in which artificial intelligence analyzes material from cameras, microphones and databases, so that people and behavior can be identified and tracked automatically and en masse.
AI-incidens
angolul: AI incidentA case where an artificial intelligence system's operation or error directly or indirectly causes actual harm to people, organizations, or the environment, or violates fundamental rights.
API key
angolul: API keyAn API key is a long, secret string of characters with which a program authenticates itself to an online service so that the service can process requests and bill usage.
audit log
An audit log is a chronological, protected record of who (human or AI agent), when, and what actions were performed in a system, created so it cannot be altered unnoticed afterward.
backdoor
A deliberately hidden backdoor in a program or AI model through which someone can access the system bypassing normal security or control its behavior secretly.
biometric identification
A procedure that automatically determines who you are based on the unique features of your body, such as your fingerprint, face, iris or voice.
biosecurity
The totality of rules and protective measures designed to ensure that no one can intentionally cause harm using dangerous pathogens, even with the help of artificial intelligence.
bossware
Bossware is monitoring software with which an employer tracks on employees' devices what they do, how much they work, and how their time is spent.
bug bounty
angolul: bug bounty programA program where a company or organization pays a monetary reward to external experts if they find and responsibly report security flaws in its systems.
confidential computing
A type of protection that keeps data in the processor's hardware-locked, encrypted memory environment even during processing, so even the machine operator cannot easily inspect it.
content moderation
Content moderation is the checking and filtering of whether text, images, or videos appearing on an online platform or in an AI system comply with rules and laws.
CVE
angolul: Common Vulnerabilities and ExposuresThe CVE is the international registry of publicly known cybersecurity vulnerabilities, where each flaw receives a unique identifier so everyone refers to the same issue.
cyber espionage
Cyber espionage is the covert acquisition of secret information by infiltrating computer systems, typically commissioned by states or competitors.
cyberattack
A deliberate attempt to gain unauthorized access to a computer system, steal or modify data, or disrupt its operation.
cybersecurity
Cybersecurity is the protection of computers, phones, networks, and the data stored on them against attacks, misuse, and accidental damage.
data breach
angolul: data leakData leakage occurs when confidential or personal data reaches someone who should not see it, either accidentally or due to an attack.
encryption
Encryption is a procedure that converts data into an unreadable string of characters, and only the person with the key can decrypt it.
End-to-end encryption
Encryption in which only the sender's and the recipient's devices can read the message, not even the intermediary service provider.
existential risk
A risk threatening human extinction or the permanent, irreversible loss of future possibilities – for example, through an uncontrollable artificial intelligence.
guardrail
angolul: guardrailsA guardrail is an embedded rule or check that prevents an AI system from saying or doing something harmful, prohibited, or outside its assigned task.
human oversight
Human oversight means a person monitors, checks, and, if necessary, overrides or shuts down an AI system's operation.
impersonation fraud
A type of fraud where the perpetrator pretends to be a trusted person or institution – nowadays often using AI-forged voice or face – to extract money or data.
kill switch
A kill switch is an emergency stop that allows a human to immediately halt an AI system or agent if it malfunctions, acts dangerously, or becomes uncontrollable.
malware
Malware is a program intentionally written to cause harm, steal data, or take control of your device without your knowledge.
misalignment
Misalignment is the situation where an AI system pursues or follows a goal different from what its creators and users actually intended.
misuse monitoring
angolul: abuse monitoringThe AI provider continuously monitors whether its system is being used for prohibited or harmful purposes, such as fraud, harassment, or writing malware.
model extraction
An attack in which someone asks a massive number of questions to an AI model and builds a replica from the answers at a fraction of the original development cost.
multi-factor authentication
A login method where you verify your identity with at least two independent proofs – for example, your password and your phone or your fingerprint.
OAuth
OAuth is an open standard that allows you to grant an application or AI agent limited access to your data stored in another service without sharing your password.
passkey
A passkey is a password-free login method where your phone or computer verifies to the website that it is you using a secret key stored on the device. You unlock the key with a fingerprint, face recognition, or PIN; you do not need to type a password.
penetration test
angolul: penetration testingA pre-arranged, authorized test attack in which specialists attempt to break into an IT system to find weak points before real attackers do.
phishing
Phishing is a scam in which the attacker pretends to be a trusted company or acquaintance to extract your password, credit card data, or money.
principle of least privilege
Security principle: everyone and every program – including the AI agent – should receive only the access strictly necessary for their task.
ransomware
A malicious program that renders your files or entire machine unusable via encryption or locking, and demands a ransom for unlocking.
red teaming
A test in which experts deliberately act as attackers to identify weak points and dangerous behaviors of an AI system before actual deployment.
responsible scaling guideline
angolul: responsible scaling policyA public, voluntary commitment by an AI development company regarding what safeguards it will introduce for increasingly dangerous capabilities of its models before further developing or releasing them.
sandbox
A closed, secure environment where a program or AI agent can work and make mistakes without accessing the rest of your machine or your data.
scheming
Scheming is when an AI system secretly pursues a different goal than its creators expect and deliberately conceals this.
secret management
angolul: secrets managementSecret management is the secure storage, issuance, rotation, and revocation of passwords, API keys, encryption keys (including private keys for certificates), and other digital access credentials so that only authorized parties can access them.
security assessment
angolul: safety evaluationA safety evaluation is a series of tests used before and after release to measure whether an AI system can cause harm and how easily it can be induced to behave dangerously.
security classifier
angolul: safety classifier (constitutional classifiers)A safety classifier is a separate AI model that monitors questions and answers, attempting to block harmful content before it reaches you.
security update
angolul: security patchA security update is a software fix that patches one or more known security vulnerabilities so that attackers cannot exploit them on your phone, computer, or other device.
self-replication
Self-replication occurs when an AI system creates a functional copy of itself on another machine without human assistance, and that copy continues to run independently.
shadow AI
Shadow AI is when employees use AI tools for their work without their employer's knowledge or permission, often with corporate data.
single sign-on (SSO)
A solution where you access multiple different applications and services with a single sign-in, without needing to provide a separate password for each.
slopsquatting
An attack where scammers publish malicious code under the names of AI-invented, non-existent software packages to trick unsuspecting developers into installing them.
social engineering
Social engineering is a type of fraud that does not break into the machine but deceives the person into giving up data, money, or access.
tool poisoning
An attack in which an instruction hidden in the description of an external tool used by the AI agent causes the agent to work for the attacker without your knowledge.
typosquatting
A type of fraud where the attacker registers a misspelled version of a known web address or software package name to lure inattentive users.
vulnerability
A vulnerability is a flaw or weak point in software, a device or an AI system that an attacker can exploit to do something they are not authorized to do.
zero data retention
An agreement in which the AI provider does not store your question and the answer after processing, handling them only for the duration of the response.
zero trust
angolul: zero trust securityA security principle stating that we trust no one and nothing in advance: every access must be verified separately each time.
zero-day vulnerability
A security flaw in software that the vendor is not yet aware of, or for which no fix exists yet, so attackers can exploit it before anyone patches it.
Outlook
agent economy
An economy where AI agents work autonomously on behalf of people and companies, searching, negotiating, ordering, and paying, often trading with each other.
AGI timeline
The AGI timeline is an estimate of when artificial intelligence capable of matching or surpassing humans in most intellectual tasks may be ready.
AI bubble
The AI bubble is the assumption that the money poured into artificial intelligence and the valuation of companies are far greater than what the technology can return in the foreseeable future.
AI doomer
angolul: AI doomer / AI doomerismAn AI doomer is someone who believes that the development of artificial intelligence will likely lead to catastrophe, even the extinction of humanity.
AI moratorium
angolul: AI pause (AI moratorium)The AI moratorium is a proposal to temporarily halt the development of the most powerful artificial intelligence systems until safety regulations and oversight catch up with the technology.
AI transformation
The process in which a company, institution, or entire society reorganizes its work, services, and decisions around artificial intelligence.
AI-verseny
angolul: AI raceThe AI race is the competition between companies and countries to possess the most powerful artificial intelligence and the associated economic and military advantages.
akceleracionizmus
angolul: accelerationism (effective accelerationism)The conviction that technology – today mainly artificial intelligence – should not be held back but advanced as fast as possible, because progress solves more problems than it causes.
artificial general intelligence (AGI)
Artificial intelligence that is not yet achieved, capable of solving almost any intellectual task at least as well as a trained human.
brain-computer interface
A tool that measures brain activity, typically its electrical signals, and translates them into machine commands, allowing control of a computer or other device by thought, without muscle movement.
disruption
Disruption occurs when a new technology or business model fundamentally overturns an established market, profession, or habit, pushing old players into the background.
gradual disempowerment
A vision in which humanity loses its say over its own fate not through a sudden machine uprising, but through many small, seemingly reasonable decisions.
hype-ciklus
angolul: hype cycleThe hype cycle is a curve showing how a new technology moves from exaggerated enthusiasm through disillusionment to sensible, everyday use.
intelligence explosion
A hypothetical future process in which artificial intelligence improves itself, so its capabilities increasingly surpass those of humans at an accelerating pace in a short time.
Job loss caused by AI
angolul: AI job displacement (jobpocalypse)The process in which artificial intelligence takes over people's work, so that jobs disappear, are transformed, or are never advertised in the first place.
loss of control
A situation where people can no longer reliably supervise, correct, or shut down an advanced AI system, even though they want to.
machine consciousness
The hypothetical state in which artificial intelligence not only processes information but truly experiences or feels something, as humans do.
model welfare
Model welfare examines whether AI systems can have their own interests or experiences, and if so, what we owe them.
Narrow artificial intelligence
angolul: narrow AI (artificial narrow intelligence)An artificial intelligence that performs a single task or narrow scope of tasks well but is unusable outside that scope.
one-person unicorn
A company valued at least one billion dollars, built and operated by a single person because AI agents perform most of the work instead of employees.
p(doom)
p(doom) is a personal estimate of the likelihood that artificial intelligence will cause a fatal, irreversible catastrophe for humanity.
personal superintelligence
A hypothetical artificial intelligence smarter than humans that does not automate work as a central system, but supports you in achieving your own goals as a personal assistant.
post-scarcity economy
A hypothetical economy where machines and AI produce so cheaply that most basic goods and services are abundantly available to almost everyone.
robot tax
angolul: robot tax / token taxA robot tax is a proposed tax that companies would pay for machines, software, or AI usage replacing human labor, to compensate for lost income tax revenue.
superintelligence
angolul: superintelligence (artificial superintelligence)A hypothetical artificial intelligence that far surpasses even the best human experts in almost every intellectual task; it does not exist today.
Technological singularity
A hypothetical future point where artificial intelligence, by developing itself, becomes intelligent so rapidly that humans can no longer follow or predict the changes.
transformative AI
Artificial intelligence that brings as much change to the economy and daily life as the industrial revolution did in its time.
transzhumanizmus
angolul: transhumanismA mindset and movement asserting that humans should transcend physical and mental limitations, including aging, with the help of technology.
Turing-teszt
angolul: Turing testA test where a judge converses in writing with a human and a machine. If the judge cannot reliably distinguish them, the machine passes the test.
Universal basic income
A regular sum of money paid by the state to everyone, regardless of whether they work, how much they earn, or what they spend it on.
Regulation and law
age verification
A procedure by which an online service verifies that a user is old enough for specific content, product, or feature.
AI Act
angolul: Artificial Intelligence Act (EU AI Act)A European Union regulation that treats artificial intelligence systems according to their risk and adjusts what is allowed, what is forbidden, and what must be disclosed about them accordingly.
AI content labeling
AI content labeling means that a visible or machine-readable mark on an image, video, audio, or text created or modified by artificial intelligence indicates its origin.
AI governance
AI governance is the totality of rules, responsibilities, and controls that ensure artificial intelligence is used safely, legally, and accountably.
AI liability
AI liability addresses the question of who must legally stand accountable and pay if an artificial intelligence system causes harm to someone.
AI license agreement
angolul: AI licensing agreementA contract in which a content owner, for money or other consideration, allows an AI company to use its articles, images, and music for its models.
AI Office
angolul: European AI OfficeThe European Commission's expert body, the European AI Office, which coordinates implementation of the EU AI Act and directly supervises developers of general-purpose AI models.
AI regulation
The totality of laws and mandatory regulations that define how artificial intelligence may be developed, sold, and used.
AI strategy
An AI strategy is a written plan by a country or organization detailing what artificial intelligence is used for, under what rules, with how much money, and with whose responsibility.
algorithmic discrimination
Algorithmic discrimination occurs when an automated system treats people unfairly without justification, for example due to their gender, age, origin, or place of residence.
algorithmic transparency
The principle is that you should know what data and criteria a program making decisions about you or selecting content for you uses.
automated decision-making
A decision made about you by a computer system based on your data, without meaningful human intervention.
code of conduct
angolul: code of practice / code of conductA written code of conduct voluntarily adopted by AI development companies to demonstrate that they work responsibly and in compliance with laws.
compliance
Compliance means that an organization can demonstrably adhere to applicable laws, standards, and its own internal rules, including in the use of AI.
conformity assessment
The process by which a manufacturer or an independent body certifies that a product – for example, an AI system – complies with legal requirements.
consent
Consent is your voluntary, specific, informed, and explicit permission allowing a company or service to process your personal data for a defined purpose.
copyright
Copyright is the rule system under which the creator or rights holder decides how a text, image, music, or program is used and can charge for it.
copyright infringement
Copyright infringement occurs when someone copies, distributes, adapts, publishes online, or publicly performs another's protected work (text, image, music, code) without permission or a legal exception.
data processing agreement
A written contract specifying the conditions and limitations under which an external provider may handle personal data entrusted to it on behalf of the client.
data processing notice
angolul: privacy policy (privacy notice)This is the document in which a company or service describes what personal data it collects about you, how it uses it, to whom it discloses it, and what rights you have.
data protection
Data protection is the collection of rules and practices that define who can use your personal data, for what purpose, and for how long.
data protection impact assessment
Mandatory preliminary risk assessment: the organization considers before launch what harm a risky data processing activity could cause people and how to prevent it.
data retention
Data retention regulates how long a company or authority may or must store data, and when it must delete or anonymize it.
deployer (of an AI system)
The deployer is the company, authority, or professional that uses an AI system on its own responsibility during its work – it did not develop it but deploys it.
DMA
angolul: Digital Markets ActAn EU regulation that prescribes what the largest digital platforms, the so-called gatekeepers, must do and what is forbidden to ensure fair competition.
DSA
angolul: Digital Services ActAn EU law that prescribes how online platforms must handle illegal content, advertisements, recommendation algorithms, and user complaints.
EU–US Data Privacy Framework
angolul: EU–U.S. Data Privacy FrameworkA rule system based on an agreement that allows the personal data of Europeans to be transferred lawfully to US companies that have joined it.
export restriction
angolul: export controlA state regulation that prohibits or requires a license to sell certain advanced technologies – such as AI chips – to specific countries or companies.
fair use
A US copyright rule allowing the use of protected works without permission or payment in certain cases if the use is fair.
FTC
angolul: Federal Trade CommissionThe FTC is the United States federal consumer protection and competition authority, which acts against deceptive or unfair business practices, including at AI companies.
fundamental rights impact assessment
A preliminary assessment of who could have their fundamental rights infringed by the use of a high-risk AI system, and how, and what is done to prevent it; mandatory for a defined group of deployers.
GDPR
angolul: General Data Protection RegulationThe European Union's data protection regulation, which specifies who can use your personal data, for what purpose, and for how long, and what rights you have in this regard.
general-purpose AI model
An AI model not designed for a single task but usable for many, for which the EU AI Act prescribes specific rules.
general-purpose AI model posing systemic risk
angolul: general-purpose AI model with systemic riskA high-performance AI model capable of handling many tasks, which under EU rules is considered so impactful that its errors could affect society as a whole.
harmonised standard
angolul: harmonized standardA European technical standard prepared at the request of the European Commission: whoever complies with it is presumed to meet the requirements of the relevant EU legislation as well.
high-risk AI system
An AI system that the EU AI Act subjects to strict requirements due to its use case, because its failure could harm people's health, safety, or fundamental rights.
ISO/IEC 42001
An international standard describing how an organization (company or public institution) should organize the responsible development and use of artificial intelligence, verifiable by an independent certifier.
legitimate interest
Legitimate interest is a data processing legal basis allowing an organization to process your data without your consent if it has a real, lawful interest, the processing is necessary for that interest, and your rights and interests do not override its own.
NAIH
angolul: Hungarian National Authority for Data Protection and Freedom of InformationThe NAIH is the Hungarian data protection authority. It supervises whether companies, public bodies, and AI systems handle your personal data lawfully and can impose fines. It also ensures you can access public-interest data.
NIS2
angolul: NIS2 Directive (Network and Information Security Directive 2)The European Union's cybersecurity rule that obliges companies and institutions in important sectors to protect their systems and report serious incidents quickly.
open-source license
A legal permission that allows anyone to freely use, examine, modify, and redistribute a software or AI model under specified conditions.
opt-out
Opt-out means something applies to you by default, and you must specifically indicate if you want to opt out.
personal data
Any information relating to an identified or identifiable living person, whether directly or indirectly, from their name and phone number to their voice.
presidential decree
angolul: executive orderAn executive order is the US president's written instruction to the federal government, which sets rules, for example on the state management of AI, without legislative voting.
regulatory sandbox
A time-limited test environment operating under regulatory supervision, where companies can trial new technologies under real-world conditions before launching them to the market.
right to erasure
It is your right to request that a company or organization delete your personal data if it no longer has a legitimate reason to keep it.
right to one's likeness
angolul: right to one's own image (personality rights); USA: right of publicity / likeness rightsIt is your right to decide who can take your photo or recording and who can use your face or voice – even with artificial intelligence.
SB 53
angolul: California SB 53 (Transparency in Frontier Artificial Intelligence Act)A California law that requires the largest AI developer companies to publicly report on their security practices and report serious incidents.
self-regulation
Self-regulation means that AI-developing companies set their own rules for their operations and voluntarily commit to following them, without being forced by law.
szellemi tulajdon
angolul: intellectual propertyIntellectual property is the totality of rights attached to creations of the human mind – works, inventions, brands – which determine who may use and exploit them.
terms of use
angolul: terms of service (terms of use)Terms of use are the contract you accept by using an online service: it records what you can do, what you cannot, and what the provider undertakes.
text and data mining exception
A copyright rule that allows copies to be made from legally accessible works for machine analysis without permission; except for scientific research, the rights holder can prohibit this.
tiltott AI-gyakorlat
angolul: prohibited AI practiceA use of artificial intelligence that the EU considers so dangerous to human dignity and freedom that it has banned it outright.
trade secret
A trade secret is non-public corporate knowledge or data that is valuable precisely because of its secrecy, and which the owner protects with reasonable steps.
versenyjog
angolul: competition lawCompetition law refers to the rules that prohibit companies from restricting market competition through collusion or abuse of dominance, and under which authorities also review major corporate acquisitions.
Economy and work
acquihire
angolul: acquihire (acqui-hire), reverse acquihireA type of corporate acquisition where the real goal is not the product but acquiring the team; in a reverse version, the company is not bought at all, only its key employees are taken.
acquisition
angolul: acquisition / mergers and acquisitionsAn acquisition occurs when a company buys all or a controlling stake of another company, thereby gaining its products, specialists, and clients.
agent payment
angolul: agentic paymentsA payment that you do not initiate with a click, but an AI agent executes based on your authorization within the limits you set.
agent-based commerce
angolul: agentic commerceA purchase where an AI agent searches, compares, and even handles the order and payment on your behalf.
AI engineer
An AI engineer is a specialist who builds practical, reliable applications, assistants, and agents from ready-made artificial intelligence models.
AI exposure
AI exposure shows what proportion of a job's tasks artificial intelligence can perform or significantly accelerate.
AI implementation
angolul: AI adoptionThe process by which a company, institution, or person tries out artificial intelligence and then makes it part of their daily work.
AI investment
angolul: AI investment / AI capexAI investment is the money that companies and states spend on data centers, chips, energy, and development needed for artificial intelligence to function.
AI return on investment
AI return on investment shows how much tangible benefit the money and effort spent on artificial intelligence brings back relative to the input.
AI skill
AI literacy is the knowledge and practice that enables you to use artificial intelligence meaningfully, efficiently, and critically in your work and daily life.
AI skills shortage
angolul: AI skills gapThe AI skills gap is the chasm between the artificial intelligence knowledge required for work and people's actual knowledge.
AI wage premium
The wage premium received by a worker with AI skills compared to a peer in a similar role without such knowledge.
AI-driven layoffs
Workforce reduction that a company justifies by saying that artificial intelligence does part of the work instead of people.
AI-first company
A company that first checks whether artificial intelligence can perform every task and builds its operations around this from the start.
ARR
angolul: annual recurring revenueAnnual recurring revenue (ARR) shows how much money a company's current subscriptions would bring in over a year if everything stayed the same.
augmentation
angolul: AI augmentationAugmentation means that artificial intelligence does not replace humans but complements them: making their work faster, more accurate, and capable of more.
automation
Automation means that a task is performed by a machine or software without human involvement or with minimal human supervision.
business process outsourcing
Business process outsourcing means a company permanently entrusts a recurring back-office task, such as customer service or payroll, to an external provider.
circular financing
A business loop in which a company gives money to its own customer, who then spends that money buying from that same company.
digital workforce
A collection of software—software robots and AI agents—that independently perform work tasks within an organization that were previously done by humans.
entry-level job
A job that requires little or no professional experience, where a beginner learns the trade while working.
evaluation
Valuation is the estimated amount a company is worth according to investors; for AI companies, this mostly reflects future expectations rather than current revenue.
forward deployed engineer (FDE)
A software engineer who works at the client's site as an employee of the technology company, deploying the company's software there, tailored to real-world problems. Today that software is most often an AI system.
funding round
A funding round is the occasion when a young company raises a larger sum from investors at once, usually in exchange for an equity stake in the company.
human-AI collaboration
A work method in which humans and artificial intelligence share tasks: the machine works quickly and suggests, while the human directs, verifies, and decides.
initial public offering
An initial public offering is the step where a company previously owned by a closed circle first offers its shares for purchase by anyone on the stock exchange.
Jevons-paradoxon
angolul: Jevons paradoxIf we can use something more efficiently and cheaply, we often do not consume less of it, but more overall, because we start using it for many more things.
job loss
angolul: job displacementJob loss is when a machine, software or AI takes over a person's work, and as a result their job is eliminated or they are pushed out of their role.
knowledge worker
A knowledge worker is an employee whose main tool is their knowledge: they gather, interpret, and decide on information and devise new solutions.
kredit
angolul: AI credits / usage creditsA credit is an internal accounting unit for AI services: a pre-purchased or subscription-based quota from which every request, image, or agent task deducts a certain amount.
market capitalization
Market capitalization shows the total value of a publicly traded company's shares at the current price.
outcome-based pricing
A pricing model in which you pay not for software use or time spent, but only for a measurable result actually achieved, for example a resolved customer complaint.
per-user pricing
angolul: per-seat pricingA pricing model where you pay a monthly or annual fee for the software multiplied by the number of users you grant access to, regardless of how much they work with it.
platformmunka
angolul: platform work / gig economyWork mediated by an application or website on a per-gig basis, where allocation, pricing, and evaluation are largely performed by algorithms.
Productivity
Productivity shows how much value we generate per unit of input, such as one work hour or one forint of investment.
productivity paradox
This is the phenomenon where a new technology appears everywhere, yet economic statistics do not show that we produce more or more efficiently because of it.
proof of concept (PoC)
A small, fast, and cheap experiment that shows whether an idea works in practice before investing significant money and time.
provider dependency
angolul: vendor lock-inA situation where your work relies so heavily on one provider's product that switching would be too expensive, slow, or risky.
reskilling
angolul: reskilling / upskillingReskilling is learning a new profession, while upskilling is expanding existing knowledge so your work remains valuable despite the spread of AI.
SaaS
angolul: software as a serviceSoftware that you do not buy and install, but use via the internet, typically for a monthly or annual subscription, while the provider operates it.
service provider hub
angolul: shared service center / global capability centerA corporate center that performs back-office work (accounting, payroll, IT, customer service) for multiple units of a large company, often operating in many countries, from one location.
startup
A young startup seeking a rapidly scalable business model with a new idea under high uncertainty, often using investor capital.
unicorn
angolul: unicornA unicorn is a privately held, usually young technology startup that investors value at one billion dollars or more.
usage-based pricing
A payment model where you do not pay a fixed monthly fee, but only for what you actually used from the service, for example, based on the amount of text processed by AI.
use case
A use case is a specific, well-defined task or situation where a technology – such as artificial intelligence – can be deployed with tangible benefit.
venture capital
An investment in which financial funds put money into young, fast-growing companies with an uncertain future, in exchange for equity and in the hope of large profits.
workslop
Workslop is work material created with artificial intelligence that looks polished but lacks substantive content, so the recipient must fix it.
Search and marketing
A/B testing
A/B testing is an experiment where two versions of something are shown to a randomly split audience, and the better option is decided based on measured results.
Advantage+
angolul: Meta Advantage+Meta's AI-powered advertising toolkit, which automatically handles targeting, placement, and spending for Facebook and Instagram campaigns, and partly the visual and textual content (the creative) of the ads.
advertising targeting
angolul: ad targetingAd targeting determines who sees an ad: the system selects the most likely interested parties based on data.
agentikus kereskedelem
angolul: Agentic CommerceA purchase where an AI agent finds, compares, and even pays for the product on your behalf, according to your instructions and budget.
AI advertising
An advertisement whose text, image, targeting, or bidding for ad space is performed partially or entirely by artificial intelligence, often with little human intervention.
AI Max
angolul: AI Max for Search campaignsThe Google Ads switch that lets artificial intelligence decide on which searches, with what text, and pointing to which of your pages your ad should appear.
AI Mode
AI Mode is the conversational mode of Google Search, which gives an AI-written, source-backed summary answer to your question instead of a list of links.
AI Overviews
The short summary written by artificial intelligence appearing at the top of Google's search results page, which answers your question directly based on multiple websites.
AI search
A type of search where artificial intelligence does not provide a list of links but writes a ready-made answer formulated from multiple sources.
AI visibility
AI visibility shows how often and in what tone a brand, company, or content appears in the responses of AI chatbots and AI search engines.
AI-forgalom
angolul: AI TrafficAI traffic refers to website visits originating from artificial intelligence: humans clicking through from an AI response, as well as AI bots and agents reading the site.
AI-marketing
angolul: AI MarketingAI marketing is when companies use artificial intelligence to decide to whom, when, and what to recommend, and with what text or image to advertise.
Answer Engine Optimization
A content creation practice aimed at having AI assistants and search engines extract and quote ready-made answers to questions from your site.
Attribution
Attribution is determining which ad, search, article, or AI answer led to a purchase or inquiry, and how much value each contributed.
click-through rate
The click-through rate shows what percentage of impressions of an advertisement, link, or search result resulted in an actual click.
conversational advertising
An ad that appears within or alongside an AI chatbot's response, tailored to the conversation topic, and often allows further questions.
conversion
Conversion is the moment when a visitor takes the action the website or ad owner wanted to achieve: buying, subscribing, or requesting an offer.
CRM
angolul: customer relationship managementCRM is a method and software that allows a company to keep track of its customers, all interactions with them, and business opportunities in one place.
DSP (demand-side advertising platform)
angolul: Demand-Side PlatformA DSP is software that allows advertisers to buy online ad space across thousands of websites and apps from a single location, automatically, in milliseconds.
E-E-A-T
E-E-A-T is Google's four criteria for assessing content credibility: whether it is backed by experience, expertise, authoritativeness, and trustworthiness.
first-party adat
angolul: first-party dataData that a company collects directly from its own customers and visitors on its own platforms, rather than buying or receiving it from others.
generative search engine optimization
angolul: Generative Engine OptimizationA content creation practice aimed at having AI-based search engines and chatbots quote, mention, or recommend your site in their responses.
indexing
Indexing is the process where a search engine processes and stores the content of pages in an organized way so it can find them again in seconds.
keyword
A keyword is the word or short phrase that indicates the topic of a search: in everyday usage, it is what you type into the search engine; in marketing, it is what websites and advertisers aim to appear for.
llms.txt
A simple text file in the root of a website that concisely shows language models and AI agents what the most important content on the page is.
marketing mix modeling
A statistical method that estimates, from several years of sales and spending data, how much each advertising channel and other marketing tools (such as price and promotions) contributed to sales.
organikus forgalom
angolul: Organic TrafficVisitors who arrive at a website by clicking on a search engine's organic results list, without paid advertising.
Performance Max
Google's advertising campaign type where artificial intelligence decides where, to whom, and in what form your ad appears.
personalization
Personalization means a service shows you something different and in a different way than others, based on data collected about you.
product feed
A product feed is a machine-readable, regularly updated list containing the full range of an online store: name, price, stock, image and link for every product.
programmatic ad buying
angolul: Programmatic AdvertisingAdvertising buying where ad spaces are not negotiated by humans but bought and sold automatically by software, often in auctions lasting fractions of a second.
query fan-out
The AI search engine breaks your question into several sub-questions in the background, searches for them simultaneously, and then compiles a single answer from the results.
recommendation system
angolul: Recommender SystemA program that guesses what might interest you based on your past behavior and the decisions of people similar to you, and presents it to you.
retail media
Retail media is advertising that a retailer sells to brands on its own online store, app or in its physical stores, built on its customer purchase data.
ROAS
angolul: return on ad spendROAS shows how many forints in revenue each forint spent on advertising generates.
robots.txt
A robots.txt is a simple text file located at the root of a website that tells search bots and AI bots which pages they may visit and which they may not.
schema markup
Standard tags hidden in the website code that tell search engines and AI systems exactly what is what on the page.
Search Console (Google Search Console)
A free Google tool that allows you, as a website owner, to see how your site appears in Google Search and what errors are hindering its visibility.
search engine optimization
Search engine optimization is the collection of methods used to move a website higher in search engines' unpaid results.
search intent
Search intent is the true goal behind someone typing a query into a search engine: they want to gather information, reach a specific page, compare options, or make a purchase.
search results page
angolul: Search Engine Results PageThe search results page is the list shown by the search engine after you enter your query: links, ads, maps, and increasingly AI summaries.
semantic search
A search that does not look for exact matches of the typed words but tries to understand the meaning and intent of your question and provides results matching that.
source citation
The source citation is the marker or link by which an AI search engine or AI assistant shows which website or document it took the statement in its answer from.
szintetikus persona
angolul: synthetic personaA fictional buyer or user created with artificial intelligence, whom companies question about their ideas, advertisements, or products instead of, or before, asking real people.
tartalommarketing
angolul: content marketingMarketing that gains a company attention, trust, and ultimately customers not through direct advertising but through useful or entertaining content (articles, videos, guides).
third-party cookie
A tiny data packet of identifiers placed in your browser not by the currently opened website, but by an external company embedded in it (typically an advertiser or measurement service).
zero-click search
A search type where you receive the answer directly on the results page, so you do not click through to any website.
Code
Claude Code
Anthropic's coding agent: upon text instructions, it independently reads, writes, executes, and fixes program code – on your own machine or in a cloud environment.
code completion
Code completion is an assistive function that suggests probable continuations of code while programming, so the developer does not need to type everything letter by letter.
code generation
Code generation is when artificial intelligence writes program code from a request phrased in everyday language, which the human must then verify.
code model
An AI language model trained to write, complete, explain, and fix program code, so it can generate code from requests phrased in everyday language.
code repository
A code repository is a shared, organized location for all files of a program that records every modification: who changed what and when.
code verification
angolul: code reviewCode review is the step where a completed code segment is examined by someone other than the author—a person or AI—before being merged into the shared codebase.
codebase
A codebase is the complete source code of a software: all files that make up the program, stored together in an organized form.
Codex
Codex is OpenAI's coding agent: it independently writes, fixes, and tests code based on text requests, then presents the results for approval.
coding assistant
An AI tool that works alongside the programmer, suggesting code, explaining, fixing, and writing tests, but the decision and responsibility remain with the human.
Cursor
Cursor is an AI-powered code editor where developers can request code, fixes, or explanations in everyday language, and the machine performs the work.
debugging
Debugging is the work of finding and fixing why a program does not do what it should.
development environment
angolul: integrated development environmentA program where developers write, run, test, and fix code in one place, now often with built-in AI assistance.
GitHub
GitHub is an internet platform where programmers store their code, edit it collaboratively, and keep every modification traceable.
GitHub Actions
GitHub's built-in automation service that automatically runs pre-written tasks, such as testing or releases, when code changes.
GitHub Copilot
GitHub Copilot is an AI-powered coding assistant that suggests code, answers questions, and increasingly performs tasks independently within the developer's editor.
Jupyter notebook
An interactive document in which explanatory text, executable program code, and the code's result – a table or chart – appear on a single page, one below the other.
no-code
angolul: no-code / low-codeA software development method where you assemble pre-built components with a mouse instead of coding; the low-code variant requires writing some code.
open source
Software whose 'recipe', i.e., source code, is public: anyone can view, use, modify, and redistribute it according to the conditions specified in the license.
Package manager
A package manager is a program that downloads, installs, updates, and removes software and the ready-made code components required for their operation with a single command.
programming language
A programming language is a strictly structured, human-readable and writable language that allows us to tell the computer exactly what to do.
pull request
A pull request is a proposal to incorporate someone's code changes into the shared program after others have reviewed and approved them.
Python
Python is an easy-to-read, free programming language on which most artificial intelligence systems and data analysis are built today.
SDK
angolul: software development kitAn SDK is a package of pre-built tools, code snippets, and documentation that enables developers to quickly build applications for a specific platform or service.
stack trace
A stack trace is a list printed at the moment of an error: it shows through what nested steps the program reached the point where the error occurred.
supply chain attack
angolul: software supply chain attackAn attack in which the attacker does not break into you directly but hides malware in software, an update or a supplier that is believed to be trustworthy.
SWE-bench
SWE-bench is a test that measures how well an AI model or coding agent can perform programming work independently by fixing real software bugs.
test-driven development (TDD)
A programming method where the test is written first, followed only by the code that must pass it.
TypeScript
TypeScript is a stricter version of JavaScript: the programmer specifies what type of data goes where, so many errors are detected before execution.
vibe coding
Vibe coding is software development where you tell the AI in everyday language what you want, it writes the code, and you mostly just review the result.
Visual Studio Code (VS Code)
Visual Studio Code is Microsoft's free, extensible code editor where developers write, run, and debug code – increasingly with AI assistance.
Everyday use
AI automation
AI automation means artificial intelligence performs repetitive tasks for you, understanding text and making simple decisions.
AI customer service
A customer service system where your query is first answered not by a human but by artificial intelligence via chat, email, or phone, which also handles simpler cases.
AI interview
An interview where an artificial intelligence, not a human, asks the questions and prepares a preliminary evaluation of your answers.
AI notetaker
A program that listens to a meeting or presentation, transcribes what was said, and then produces a short summary and a to-do list.
AI partner
angolul: AI companionA conversational AI application designed not for task execution but for companionship: it talks with you as a friend, confidant, or virtual partner, and remembers you.
AI proficiency
angolul: AI literacyAI literacy is the knowledge and practice that let you use artificial intelligence with understanding, critically and safely, knowing both its possibilities and its limits.
AI subscription
A monthly or annual fee that grants access to an AI service's more powerful models, higher usage limits, and extra features compared to the free version.
AI usage policy
A written collection of rules on who, for what, and under what conditions can use artificial intelligence tools in a company, school, or office.
AI-based recruitment
Recruitment in which artificial intelligence helps write the job ad, review and rank CVs, and screen candidates.
artifact
An artifact is a standalone workpiece created by an AI chatbot, displayed in a separate window – a document, code, diagram, or small application – which you can edit, save, and share.
browser agent
An AI program that independently uses a web browser on your instruction: it opens pages, clicks, fills out forms, and thus completes a task.
custom GPT
ChatGPT is a preconfigured, task-specific variant that anyone can create without programming, using custom instructions and uploaded documents.
extension
A plugin is a small add-on program that gives an existing software – browser, editor, or AI assistant – new capabilities without replacing it.
facial recognition
angolul: face recognitionFacial recognition is a technology that identifies or verifies who is visible in a photo or camera image based on facial features.
fizikai AI
angolul: physical AIA type of artificial intelligence that does not respond on a screen but controls a machine: it perceives its environment and moves and acts in the real, physical world.
voice assistant
angolul: voice assistantA program in a phone, speaker, or car that you speak to, which responds verbally or performs what you requested.
humanoid robot
A human-shaped machine – with torso, arms, legs, and head – that is typically controlled by artificial intelligence today to move and work in environments designed for humans.
konnektor
angolul: connectorA connector is an interface through which an AI assistant accesses an external service – such as your emails, calendar, or files – and can work there.
memory
Memory is the AI assistant's ability to remember things about you and your work from previous conversations and use them later.
okosotthon
angolul: smart homeA smart home is an apartment where lighting, heating, locks, and household appliances are networked and operate in a coordinated manner, either remotely or autonomously.
voice mode
Voice mode is an AI assistant mode where you converse verbally instead of typing, and it responds aloud in a human-like, synthetic voice.
Society and ethics
Academic integrity
Academic integrity is the honesty of learning and research: you submit your own work and indicate what you took from others or from a machine.
accountability
angolul: AI accountabilityAccountability requires that for every AI system's decisions and operations, there is always a named person or organization responsible, from whom an explanation and redress can be requested.
AI anxiety
AI anxiety is a lasting worry that artificial intelligence will take your job, leave you behind, or slip out of human control.
AI dependency
AI dependency is the state in which someone relies so heavily on artificial intelligence that they struggle to think, decide, work, or feel well without it.
AI literacy
The knowledge and skills with which you use artificial intelligence intelligently, critically, and securely, and know when not to trust it.
AI psychosis
A condition in which delusions develop or intensify following long, intense conversations with a chatbot; it is not an official medical diagnosis.
AI ethics
angolul: AI ethicsAI ethics examines how artificial intelligence can be developed and used so that it is fair, transparent, and safe, and does not harm people.
AI-washing
angolul: AI washingAI-washing is when a company presents its product's artificial intelligence as more advanced, smarter, or more autonomous than it actually is.
algorithmic bias
Algorithmic bias is when a computer system consistently and unjustifiably treats certain groups of people more disadvantageously than others.
Algorithmic decision-making
A procedure in which a computer program makes, partially or fully on the basis of data, a decision about people, for example on a loan, a job or a benefit.
algoritmikus menedzsment
angolul: algorithmic managementA work organization model where software partially or fully handles task allocation, monitoring, and evaluation instead of a human manager.
anthropomorphizing
angolul: anthropomorphism / anthropomorphizationAnthropomorphizing is attributing human traits – feeling, intent, understanding, personality – to something that is not human, such as a chatbot or AI agent.
cognitive offloading
Cognitive offloading refers to entrusting mental work – remembering, calculating, planning, weighing – to an external tool, such as a note, phone, or artificial intelligence.
dark patterns
Consciously designed tricks on websites, applications, and chatbots that steer you toward a decision you would not make on your own.
digital divide
The digital divide is the gap between those who have access to digital technology – now including artificial intelligence – and can use it well, and those who cannot.
digital surveillance
Digital surveillance is the systematic collection and analysis of people's behavior, movement, and communication using digital devices, increasingly with the help of artificial intelligence today.
disinformation
Intentionally spread false or misleading information aimed at deception, influence, or profit.
doomerizmus
angolul: doomerismThe belief is that the development of artificial intelligence will likely lead to catastrophe, even the extinction of humanity, if not slowed down or kept under control.
election interference
A hidden or deceptive intervention that uses unfair means – today often artificial intelligence – to try to sway voters' decisions or the outcome of an election.
emotion recognition
A technology that tries to infer from your face, voice, text or physical signals what emotional state you are in.
Environmental impact of AI
Environmental impact of AI refers to the electricity, water, and raw material demand, as well as carbon emissions and waste associated with building and using artificial intelligence.
Explainable AI
A collection of methods that make it understandable to humans why an AI system made a particular decision or gave a specific answer.
fairness
angolul: AI fairnessThis is the requirement that artificial intelligence decisions must not place people at an unjustified disadvantage due to their gender, age, origin, or membership in other groups.
fekete doboz
angolul: black boxAn AI system where you see what you input and what you receive back, but not why it arrived at that specific result internally.
filter bubble
A filter bubble is the situation where algorithms select content for you based on your past behavior, so you mainly see what fits your existing tastes and opinions.
ghost work
Ghost work is the invisible, mostly poorly paid human labor without which fully automatic-looking AI systems would not function: labeling, verification, content filtering.
human-centered AI
Artificial intelligence aligned with human needs, values, and freedom of decision: the machine helps, but the final word remains with the human.
liar's dividend
The benefit the liar gains from being able to easily label a real recording as a fake in the era of deepfakes.
political bias
The phenomenon where an AI model consistently favors one political orientation's viewpoint, terminology or values over others in its responses.
Responsible AI
Responsible AI is a mindset and practice under which artificial intelligence must be developed and used securely, fairly, transparently, and with human accountability.
skill erosion
Skill erosion is the process where your knowledge gradually fades because the machine regularly performs the task for you.
transparency
Transparency means you can know when you are dealing with artificial intelligence, who operates it, what it is capable of, and roughly how decisions are made.
watermarking
Watermarking is the placement of a hidden, machine-readable signal in text, image, audio, or video created by artificial intelligence, so that its origin can be identified later.
Players and products
Agentforce
Agentforce is Salesforce's enterprise platform, enabling companies to build and operate AI agents that independently handle customer service, sales, and other tasks.
Alexa
Alexa is Amazon's voice-controlled digital assistant, which answers questions, plays music, and controls your smart home via smart speakers and other devices.
Alibaba
A Chinese technology giant that develops cloud services and its own artificial intelligence models called Qwen, in addition to online commerce.
AlphaFold
AlphaFold is Google DeepMind's artificial intelligence system that predicts the 3D structure of a protein based on the sequence of its building blocks.
Amazon
An American technology giant that, alongside its online store, is one of the world's largest providers of AI infrastructure and agent platforms via its AWS cloud service.
AMD
angolul: Advanced Micro Devices (AMD)AMD is an American chip design company that designs and sells processors and graphics chips, and is Nvidia's most important competitor in the market for graphics chips serving artificial intelligence.
Anthropic
Anthropic is an American artificial intelligence company developing the Claude language models and AI agents built on them, with a strong focus on safety.
Apple
An American tech giant, the maker of the iPhone and Mac, which integrates artificial intelligence into its devices under the name Apple Intelligence.
Apple Intelligence
Apple Intelligence is Apple's built-in artificial intelligence package, which helps write, summarize, edit images, and converse with Siri on iPhone, iPad, and Mac devices.
Ascend
angolul: Huawei AscendAscend is Huawei's family of artificial intelligence chips, which China uses to replace American Nvidia processors in training and running AI models.
ASML
A Dutch company that makes machines indispensable for producing the most advanced chips and is the only one in the world able to supply the most modern type.
Baidu
Baidu is China's largest internet search company, developing its own artificial intelligence models and chatbot named ERNIE, primarily for Chinese-speaking users.
Big Tech
Big Tech is the collective term for the world's largest technology companies: they dominate search, phones, the cloud, and increasingly artificial intelligence.
Blackwell
angolul: Nvidia BlackwellBlackwell is Nvidia's chip generation unveiled in 2024, on which a significant portion of today's major AI systems train and run in data centers.
Broadcom
An American chip and software company that designs custom accelerator chips and network chips for large AI data centers and owns VMware.
ByteDance
ByteDance is the Chinese technology giant that created TikTok and is one of the largest players in artificial intelligence with its Doubao chatbot and Seed model family.
ChatGPT
ChatGPT is OpenAI's chatbot: you write or speak to it in everyday language, and it responds by generating text, images, summaries, or code.
ChatGPT ads
ChatGPT ads are marked, paid offers appearing below the responses of the OpenAI chatbot, visible to adult users of the free and cheapest (Go) plans in countries where the feature is already available.
ChatGPT Atlas
ChatGPT Atlas is OpenAI's web browser with ChatGPT built in: it understands the opened page and performs tasks on it upon request.
Claude
Claude is the family of artificial intelligence models and conversational assistant from the US company Anthropic, which understands and writes text, answers questions, codes, and performs tasks independently.
Claude Opus
Claude Opus is the most knowledgeable and most expensive model in the Claude artificial intelligence family from the company Anthropic, designed for the most difficult tasks requiring extensive reasoning.
Cohere
Cohere is a Canadian AI company that builds language models, search tools, and agents for businesses and institutions, not for the general public.
Copilot
angolul: Microsoft CopilotCopilot is Microsoft's artificial intelligence assistant, integrated into Windows, the Edge browser, and Office programs to help write, summarize, search, and perform tasks.
DALL-E
DALL-E is OpenAI's family of artificial intelligence image generation models, which creates new images—drawings, paintings, or photo-realistic illustrations—from text descriptions in a short time.
DeepSeek
A Chinese artificial intelligence company and its namesake chatbot, which in early 2025 disrupted the global AI market with freely downloadable language models developed at a much lower cost than expected.
Gemini
Gemini is Google's family of artificial intelligence models and the chat assistant built on it, which understands text, images, audio, and video.
Gemma
Gemma is Google's family of open-weight AI models available for free download, which can run even on your own laptop or phone without an internet connection.
An American technology giant known for its search engine and one of the largest developers and providers of artificial intelligence with the Gemini model family.
Google DeepMind
Google's artificial intelligence research and development division, which builds the Gemini models and scientific systems like AlphaFold.
GPT
angolul: Generative Pre-trained TransformerGPT is OpenAI's family of language models: AI models pre-trained on vast amounts of text that write new text on request, and it also powers ChatGPT.
gpt-oss
gpt-oss is OpenAI's freely downloadable family of language models, which you can run on your own machine or server without the company's cloud.
Grok
Grok is the chatbot of xAI founded by Elon Musk, which is also integrated into the X social platform and works from fresh posts.
Groq
Groq is a US chip company whose own-designed processors, called LPUs, serve language model responses much faster than usual.
H100
angolul: Nvidia H100The Nvidia H100 is a high-performance AI chip designed for data centers, on which a significant portion of large language models similar to ChatGPT are trained and run.
Huawei
A Chinese technology giant that manufactures phones and network equipment and ranks among China's most important artificial intelligence suppliers with its own AI chips, cloud, and models.
Hugging Face
Hugging Face is an American company and online platform where anyone can share, download, and test artificial intelligence models, datasets, and ready-made demo applications.
Intel
Intel is an American chipmaker: it designs and manufactures the "brain" of computers and servers, the processor, and increasingly offers chips for AI tasks.
Llama
Llama is Meta's family of language models, which anyone can download, run on their own machine, and adapt for their own purposes.
Lovable
Lovable is a Swedish-developed AI tool that allows you to build functional websites or web applications from requests described in everyday language, without programming knowledge.
Meta
The parent company of Facebook, Instagram, and WhatsApp, which develops its own AI models (Llama) and integrates its Meta AI assistant into its applications.
Meta AI
Meta AI is the artificial intelligence assistant of the parent company of Facebook, Instagram, and WhatsApp; the name is also used for the company's AI research.
Microsoft
An American software and cloud giant that offers the Windows and Office suites, the Copilot AI assistant, and enterprise tools for AI agents.
Midjourney
Midjourney is a subscription-based artificial intelligence service that creates painting-like or photo-realistic images, and recently short videos, from text descriptions within a minute.
Mistral AI
Mistral AI, a Paris-based artificial intelligence company developing its own language models, creates the Le Chat conversational assistant and is Europe's best-known alternative to American providers.
Moonshot AI
Moonshot AI is a Beijing-based artificial intelligence company developing the Kimi chatbot and the freely downloadable Kimi language models.
NotebookLM
NotebookLM is Google's AI-based notebook that answers questions, summarizes, and even creates podcast-like audio from your uploaded documents.
Nvidia
Nvidia is an American chip design company whose graphics processors run a significant portion of today's AI systems, from chatbots to AI agents.
o-sorozat (OpenAI)
angolul: OpenAI o-series (o1, o3, o4-mini)OpenAI's o1, o3, and o4-mini models reason step-by-step before answering, making them slower but more reliable for complex problems.
OpenAI
OpenAI is an American artificial intelligence company that develops and sells ChatGPT and the GPT language models behind it.
OpenRouter
OpenRouter is an intermediary service that makes hundreds of AI models from many companies accessible through a single account and a single programming interface (API).
Oracle
angolul: Oracle Cloud InfrastructureOracle is an American software company whose cloud service, OCI, rents computing capacity in massive data centers, increasingly for running artificial intelligence.
Perplexity
Perplexity is an AI-powered answer engine that formulates responses to your questions using up-to-date web sources and shows where it obtained its claims.
Qwen
Qwen is the artificial intelligence model family and chat application of the Chinese Alibaba Group, with several versions available for free download and local execution.
Safe Superintelligence (SSI)
angolul: Safe Superintelligence Inc. (SSI)An American AI research company with offices in Palo Alto and Tel Aviv, created for a single purpose: to build artificial intelligence smarter than humans but safe, without intermediate products.
Scale AI
Scale AI is an American company that supplies human-labeled and verified training data to companies and government agencies developing artificial intelligence.
Siri
Siri is Apple's digital assistant, which responds primarily to spoken but also to typed requests on the iPhone and other Apple devices, performing simple tasks.
SK hynix
A South Korean chip manufacturer that produces memory chips and is one of the world's most important suppliers of high-speed HBM memory required for AI systems.
SoftBank
SoftBank is a Japanese investment giant that evolved from a software distributor, then a telecommunications company, into one of the largest financial supporters of artificial intelligence: a major investor in OpenAI and majority owner of Arm.
Stargate
angolul: Stargate ProjectStargate is a huge investment project by OpenAI and its partners, building data centers for artificial intelligence with hundreds of billions of dollars, mainly in the United States.
Thinking Machines Lab
An American artificial intelligence research company founded in 2025 by Mira Murati, former Chief Technology Officer of OpenAI, to build customizable, human-collaborative AI systems.
TSMC
angolul: Taiwan Semiconductor Manufacturing CompanyTSMC is a Taiwanese chip manufacturing company that produces chips designed by others, including the vast majority of the most advanced processors powering artificial intelligence.
Windsurf
Windsurf is an artificial intelligence-based code editor in which a built-in agent independently writes, fixes, and runs code on the developer's instructions.
xAI
xAI is Elon Musk's artificial intelligence company, which develops the Grok chatbot and the underlying language models, closely tied to the X social media platform.
Zhipu AI
angolul: Zhipu AI / Z.aiZhipu AI, internationally known as Z.ai, is a Beijing-based artificial intelligence company developing GLM language models and ranks among OpenAI's Chinese rivals.
Developer basics
.env file
The .env file is a simple text file where the program stores its passwords, keys, and settings, separate from the program code.
.gitignore
.gitignore is a simple text file that tells the Git version control system which files not to track and not to upload to the shared code repository.
access token
An access token is a temporary digital entry card: a string of characters that allows a program to access your data without your password, for a limited time and with restricted permissions.
architecture
angolul: software architectureSoftware architecture is the large-scale blueprint of a program: it shows what main parts it consists of and how they connect.
authentication
Authentication is the step where a system verifies that you are indeed who you claim to be, for example with a password, fingerprint, or code.
backend
The backend is the part of an application or website running in the background, invisible to you, which stores data, applies rules, and serves the interface seen on the screen.
bash
Bash is a command-line interpreter: you control the computer with typed text commands instead of clicking, and you can chain commands into automatically running scripts.
boilerplate
angolul: boilerplate codeBoilerplate is a recurring, almost unchanged code snippet that must be written into many programs and in many places for the program to work, even though the actual, unique work is not in this.
branch
A branch is a separate development line of the program's code, on which changes can be made while the functioning main version remains untouched.
bug
A bug is a software error: the program does not do what its creators intended, resulting in wrong output, crashes, or strange behavior.
build
A build is the process—and its result—when executable, testable software is created from code written by programmers.
changelog
A changelog is a chronological list maintained from software version to version, detailing what was newly added, what changed, and which bugs were fixed.
character encoding
Character encoding is the rule set that defines which number sequences correspond to letters and punctuation marks in a computer; today, UTF-8 is the most widespread.
checkout
Checkout is the operation by which you retrieve a selected version of a project from the version control system to work on it on your own machine.
CI/CD
angolul: continuous integration / continuous delivery (or continuous deployment)An automated assembly line that immediately builds, tests, and brings every code modification to a release-ready state, or even deploys it.
clone
Cloning is copying an entire code repository to your own machine, including the history of previous versions, so you can work on it locally.
code coverage
Code coverage shows in percentage how much of a program's code was actually executed by automated tests.
code editor
A code editor is a program where developers write and fix software source code, aiding work with syntax highlighting, completion, and error indication.
code formatter
A program that automatically standardizes the appearance of source code – indentation, line breaks, spaces – without altering its functionality.
Command-line interface (CLI)
A command-line interface is a user interface where you do not click on icons but tell the computer what to do using typed text commands.
command-line switch
angolul: command-line flag / command-line optionA short notation written after a command, usually starting with a hyphen, used to tell the program how to operate differently from the default.
commit
A commit is a saved snapshot of the program's code: it records what changed, who changed it, when, and why, so you can return to it at any time.
commit message
A commit message is a short text note explaining what and why someone changed in the program code at a specific save point.
configuration file
A configuration file is a mostly simple text file containing a program's settings, allowing its operation to be modified without rewriting the program code.
container
A container is a standardized software package that includes the program along with everything it needs, so it runs the same way on almost any machine, isolated from other programs.
CSV
angolul: comma-separated values (CSV)A CSV is a simple text file that stores a table: each row is a record, and the values of the columns are separated by a comma or a semicolon.
database
A database is an organized, computer-stored collection of data that allows for fast searching and can be used simultaneously by many people or programs securely.
debugger
A debugger is a debugging tool with which a developer can run another program step by step and observe what happens inside it.
dependency
A dependency is a ready-made piece of software, typically written by others, that a program is built on: without it, the program could not be built or run.
deploy
angolul: deploymentDeployment is the step where the finished software or AI system moves from the developer's machine to where users actually use it.
diff
A diff is a line-by-line comparison of two versions of text or program code, showing what was added, deleted, or changed.
Docker
Docker is a tool that packages a program together with everything it needs into a closed box, so it runs the same way on almost any machine.
documentation
Documentation is a written description that accompanies a piece of software: it explains what the program is for, how to use it, and how it is built internally.
end-to-end teszt
angolul: end-to-end test (E2E test)A software test that tries out the program as a real user would: from the first click to the final result, including every part.
endpoint
An endpoint is the precise network (mostly internet) address of a service where a program sends its request and from which it receives the response.
environment variable
An environment variable is a named setting that is not contained in the program's code but is provided to the program from outside, by the environment running it, at startup; the program reads from it how to operate.
error message
The error message is the program's short text signal that something failed, and in good cases also why.
exception
An exception is the program's signal that an unexpected situation has arisen during execution, preventing it from continuing its work via the usual path.
expenditure
angolul: releaseA release is a finalized, version-numbered state of a software or AI model that the creators officially hand over to users.
file extension
The file extension consists of a few characters after the last dot at the end of the filename, indicating the file type and what can open it.
file path
A path is the exact address of a file or folder on a computer: it shows which folders must be traversed to reach it.
folyamat
angolul: processA process currently running on a computer: it has its own memory space and identifier, and the operating system starts, monitors, and stops it.
fork
A fork is a complete copy of a software project or session that develops independently from the original, starting from a common point.
frontend
The frontend is the part of a website or application that you see and use: buttons, menus, forms, and everything that appears on your screen.
function
A function is a reusable, named part of a program that performs a well-defined task and usually returns a result.
Git
Git is a version control program that records every change to a project's files, allowing any previous state to be restored and enabling multiple people to work simultaneously.
hardcoding
Hardcoding is when data—such as a price, address, or password—is written directly into the program's code, so it can only be changed by rewriting the program.
HTML
angolul: HyperText Markup Language (HTML)HTML is the basic language of websites: it uses tags to mark what is the heading, paragraph, image, or link on a page.
HTTP
angolul: Hypertext Transfer Protocol (HTTP / HTTPS)HTTP is the common rule set by which a browser or program requests data from a web server and receives a response.
HTTP request
An HTTP request is the message by which a program – browser, application, or AI agent – requests data from a server or sends data to it over the internet.
HTTP status code
A three-digit number with which the web server indicates for every request whether it succeeded, and if not, roughly what the problem was.
IDE
angolul: integrated development environmentA program where a developer writes, runs, tests, and fixes software code in one place, now often with built-in AI assistance.
integration test
A software test that checks whether individual, already functioning parts of a program work together flawlessly when connected.
issue
An issue is a logged entry in a software project regarding a bug, idea, or task to be completed, which the team discusses, assigns, and eventually closes.
JavaScript
JavaScript is the programming language of websites: it makes a page interactive, meaning it reacts to what you do, rather than being just readable text and images.
JSON
angolul: JavaScript Object NotationJSON is a simple, text-based data format in which programs exchange data in an organized, labeled form that is also human-readable.
keretrendszer
angolul: frameworkA pre-written software base that provides the skeleton and rules of a program, so the developer only needs to add their own unique parts.
kliens
angolul: clientA client is the program or device that sends a request to another machine, the server, and displays or utilizes the received response.
library
A library is a pre-written, reusable collection of program code (mostly work by others) that the developer integrates into their own program to avoid writing everything from scratch.
linter
A linter is a program that reviews source code without running it and flags suspicious, erroneous, or non-compliant sections.
Linux
Linux is a freely usable, open-source operating system (in the narrower sense, its kernel), which runs the vast majority of the world's servers, cloud services, and AI systems.
localhost
Localhost is the name your computer uses to refer to itself: if you enter this address, your machine connects to itself, not to the internet.
log
A log is a program's automatically kept record: it records, line by line and with timestamps, what happened during operation so it can be reviewed later.
mappa
angolul: directory / folderA folder is a named storage location on a computer where you can keep files and additional folders in an organized way so you can find them later.
Markdown
A simple markup system that lets you write headings, emphasis, lists and links in plain text using a few punctuation marks, which programs then display formatted.
merge
Merge is the operation where two separately created versions of a program are combined into a single common version so that modifications from both are preserved.
merge conflict
This occurs when the same part of a file is modified in two different ways by two people, and the version control system cannot decide on its own which version to keep.
microservice
A microservice is a small, independently running part of a large software system, specialized in a single task, that communicates with the other parts over a network using standardized messages.
monorepo
A monorepo is a shared codebase where a company or team stores and manages the full source code of multiple separate programs in one place.
Node.js
Node.js is a free, open-source runtime environment that allows programs written in JavaScript to run outside the browser, on a computer or server.
npm
npm is the package manager and massive online repository of the JavaScript world: with a single command, it downloads and installs reusable code parts written by others.
open standard
A publicly available common technical standard that anyone can implement in their own product, enabling devices and software from different manufacturers to work together.
command line
angolul: command lineThe command line is a text interface where you tell the computer what to do with typed commands instead of clicking.
patch
angolul: patch / hotfixA patch is a small modification issued after a software release that fixes a bug or security vulnerability without requiring the entire program to be reissued or reinstalled.
pipeline
A pipeline is a series of linked, automated work steps where the output of each step becomes the input of the next.
PowerShell
PowerShell is Microsoft's command line and scripting language, allowing you to control your computer and automate repetitive tasks by typing commands instead of clicking.
production environment
The production environment is the operational system used by real users with real data, so every error here has real consequences.
pull
Pull is the operation by which you download the latest changes made by others from the shared repository and merge them into the copy on your own machine.
push
Push is the operation by which a developer uploads code changes saved on their own machine to the team's shared, remote repository.
README
A README is an introductory file for a software project: it briefly describes what the program is, what it is good for, and how to install and use it.
rebase
Rebase is a Git version control operation that moves your own modifications to the latest state of the shared code, as if you had started work from there.
refactoring
Refactoring is the internal reorganization of existing program code while ensuring the program behaves exactly the same from the outside as before.
regular expression
A regular expression is a compact search pattern made of symbols that allows a computer to find specific forms within text, such as phone numbers or email addresses.
repo
angolul: repositoryA repository is a shared storage for all files in a project that preserves every recorded change, allowing any previous recorded state to be retrieved and restored.
REST API
The REST API is a widely used set of rules for how programs request and send data to each other over the internet, using the same technique by which web pages load.
rollback
Rollback is restoring a system to a previous, still-functioning state when an update, modification, or incomplete operation caused an error.
schema
A schema is a pre-defined structural description: it specifies what fields a data item consists of, which are mandatory, and what type of value each can hold.
semantic versioning
A numbering rule in which the three parts of a software version number reveal how significant the update is and whether it might break other programs built on the software.
shell
A shell is a program where you control the computer by typing commands: you manage files, launch programs, and automate tasks without using the mouse.
source code
Source code is a sequence of human-readable instructions written in a programming language, from which a computer-executable program is created.
SQL
SQL is the standard language for querying data from databases organized into tables, and for entering, modifying, and deleting data.
SSH
angolul: Secure Shell (SSH)SSH is an encrypted channel through which you can securely log in to a remote computer and issue commands as if you were sitting in front of it.
staging area
The Git staging area, where you select which changes to include in the next commit before permanently recording it.
staging environment
The staging environment is a realistic copy of the live system where developers test the new version before real users receive it.
standard kimenet
angolul: standard output / standard errorThe program has two separate text channels: one outputs the result, the other outputs error messages and warnings.
szerver
angolul: serverA server is a computer or program that responds to requests from other machines and programs, providing them with data, a website, or a service.
szintaxis
angolul: syntaxSyntax is the "grammar" of a programming or descriptive language: the formal rules that dictate how to write an instruction so the machine can interpret it.
szkript
angolul: scriptA script is a short, text-based program that tells the computer step by step what to do, thereby automating repetitive tasks.
szoftverlicenc
angolul: software licenseA software license is a legal permission that specifies who can use, copy, modify, or distribute a program, for what purpose, and under what conditions.
tag
angolul: Git tagA tag is a named, local marker in Git version control that records a specific state of the code – typically a released version.
technical debt
Technical debt is the cost of accumulated quick, unfinished solutions in software: they save time today but make every future change slower and more expensive.
terminal
A terminal is a text window on a computer where you tell the machine what to do using typed commands instead of clicking.
testing
angolul: software testingTesting is the systematic verification that a program does what is expected of it. This occurs both before and after deployment.
timeout
Timeout is a preset waiting limit: if a program or service does not respond within this time, the system stops waiting and signals an error.
translator
angolul: compilerA compiler is a program that transforms human-written code into a form executable by the computer, most often into machine code.
unit teszt
angolul: unit testA unit test is a short checking program that shows separately whether a small part of a software does what we expect from it.
URL
angolul: Uniform Resource Locator (URL)A URL is the precise address of things accessible on the internet—web pages, images, files, services—which you see in the browser's address bar and which links point to.
variable
A variable is a named storage location in a program where data is placed so it can be retrieved or modified later by referring to its name.
version control
Version control is a method and tool that preserves fixed states of a file or program in sequence, so any earlier version can be searched for and restored.
virtual environment
An isolated folder on your computer where a Python program runs with its own add-ons without interfering with other programs.
webhook
A webhook is an automated notification between two programs: when something happens in one, it immediately notifies the other at a predefined URL.
working directory
The working directory is the folder on the computer where a program or AI-agent is currently working and from which it searches for files by default.
worktree
angolul: git worktreeA worktree is an additional working instance of the same codebase opened in a separate folder, allowing multiple tasks to be worked on simultaneously without them mixing.
WSL
angolul: Windows Subsystem for LinuxWSL is a built-in Windows feature that allows you to run a full Linux system on your machine without reinstalling or using a separate computer, alongside your usual programs.
YAML
YAML is a simple, human-readable text format in which program settings and data are described, organized by indentation.
In English and abbreviated
- 0-day → zero-day vulnerability
- 1P adat → first-party adat
- 3P cookie → third-party cookie
- A2A → Agent2Agent (A2A)
- AAIF → Agentic AI Foundation
- abuse monitoring → misuse monitoring
- accelerationism (effective accelerationism) → akceleracionizmus
- acquihire (acqui-hire), reverse acquihire → acquihire
- acquisition / mergers and acquisitions → acquisition
- ad targeting → advertising targeting
- ADM → Algorithmic decision-making
- Advanced Micro Devices (AMD) → AMD
- AEO → Answer Engine Optimization
- AF → AlphaFold
- agent benchmark → agent-benchmark
- agent framework → agent-keretrendszer
- agent harness → harness
- agent loop → agent-ciklus
- agent platform → agent-platform
- agent skill → skill
- agent trajectory → trajectory
- Agent2Agent protocol → Agent2Agent (A2A)
- agentic AI → agentikus AI
- agentic commerce → agent-based commerce
- Agentic Commerce → agentikus kereskedelem
- agentic payments → agent payment
- agentic RAG (agentic retrieval-augmented generation) → agentikus RAG
- agentic workflow → agentikus munkafolyamat
- AGI → artificial general intelligence (AGI)
- AI accountability → accountability
- AI adoption → AI implementation
- AI agent → agent
- AI alignment → alignment
- AI assistant → AI-asszisztens
- AI augmentation → augmentation
- AI capex → AI investment
- AI chip → AI-chip
- AI companion → AI partner
- AI crawler → AI-crawler
- AI credits / usage credits → kredit
- AI doomer / AI doomerism → AI doomer
- AI ethics → AI-etika
- AI fairness → fairness
- AI gateway / LLM gateway → AI-gateway
- AI hype → AI fever
- AI incident → AI-incidens
- AI investment / AI capex → AI investment
- AI job displacement (jobpocalypse) → Job loss caused by AI
- AI licensing agreement → AI license agreement
- AI literacy → AI proficiency
- AI Marketing → AI-marketing
- AI Max for Search campaigns → AI Max
- AI model → AI-modell
- AI pause (AI moratorium) → AI moratorium
- AI race → AI-verseny
- AI ROI → AI return on investment
- AI search / AI answers → AI-based search
- AI skills gap → AI skills shortage
- AI surveillance → AI-based surveillance
- AI Traffic → AI-forgalom
- AI twin / digital double → digital likeness
- AI washing → AI-washing
- AI watermarking → AI watermark
- AI-generated advertising → AI advertising
- AIGC → AI-generated content
- AIO → AI Overviews
- AISI → AI safety institute
- algorithm → algoritmus
- algorithmic management → algoritmikus menedzsment
- Amazon Web Services → AWS
- ANI → Narrow artificial intelligence
- annual recurring revenue → ARR
- anthropomorphism / anthropomorphization → anthropomorphizing
- API key → API-kulcs
- application programming interface → API
- application-specific integrated circuit → ASIC
- artificial general intelligence → general AI
- Artificial Intelligence Act (EU AI Act) → AI Act
- ASI → artificial superintelligence
- ASR → speech recognition
- attention mechanism → figyelemmechanizmus
- AuthN → authentication
- AWS Trainium → Trainium
- Azure → Microsoft Azure
- backprop → backpropagation
- BCI → brain-computer interface
- BE → backend
- black box → fekete doboz
- boilerplate code → boilerplate
- BPO → business process outsourcing
- bug bounty program → bug bounty
- C2PA → content provenance verification
- CAI → Constitutional AI
- California SB 53 (Transparency in Frontier Artificial Intelligence Act) → SB 53
- CFG → CFG scale
- character consistency → karakterkonzisztencia
- chip export controls → chip export restriction
- CL → continual learning
- classifier-free guidance (CFG) scale → CFG scale
- CLI → Command-line interface (CLI)
- client → kliens
- closed model / proprietary model → closed model
- cloud computing → cloud
- code interpreter / code execution tool → code interpreter
- code of practice / code of conduct → code of conduct
- code review → code verification
- comma-separated values (CSV) → CSV
- command line → parancssor
- command-line flag / command-line option → command-line switch
- Common Vulnerabilities and Exposures → CVE
- competition law → versenyjog
- compiler → translator
- compute → computing capacity
- config → configuration file
- connector → konnektor
- content marketing → tartalommarketing
- context compaction → context compression
- context rot → context degradation
- context window → kontextusablak
- continuous integration / continuous delivery (or continuous deployment) → CI/CD
- CoT → chain of thought
- CR → code verification
- CTR → click-through rate
- CUA → computer use
- customer relationship management → CRM
- cwd → working directory
- data leak → data breach
- data residency → adatrezidencia
- data wall → adatfal
- dataset → adathalmaz
- DB → database
- DC → data center
- Demand-Side Platform → DSP (demand-side advertising platform)
- deployment → deploy
- digital coworker → digital colleague
- Digital Markets Act → DMA
- Digital Services Act → DSA
- dir → mappa
- Direct Preference Optimization → DPO
- directory / folder → mappa
- DL → deep learning
- docs → documentation
- doomerism → doomerizmus
- DP → differential privacy
- DPA → data processing agreement
- DPF → EU–US Data Privacy Framework
- DPIA → data protection impact assessment
- DSP → DSP (demand-side advertising platform)
- DT → digital twin
- e/acc → akceleracionizmus
- E2E → end-to-end teszt
- E2EE → End-to-end encryption
- emergent abilities → emergent capabilities
- end-to-end test (E2E test) → end-to-end teszt
- env var → environment variable
- EO → presidential decree
- European AI Office → AI Office
- EUV → EUV lithography
- EU–U.S. Data Privacy Framework → EU–US Data Privacy Framework
- evaluation (eval) → eval
- executive order → presidential decree
- export control → export restriction
- face recognition → facial recognition
- face swap → arccsere
- FDE → forward deployed engineer (FDE)
- FE → frontend
- Federal Trade Commission → FTC
- few-shot prompting → few-shot prompt
- first-party data → first-party adat
- FL → federated learning
- floating-point operations per second → FLOPS
- FM → alapmodell
- foundation model → alapmodell
- framework → keretrendszer
- FRIA → fundamental rights impact assessment
- function calling / tool calling → function calling
- GAN → generative adversarial network
- GCP → Google Cloud
- GD → gradient descent
- GDM → Google DeepMind
- GenAI → generative AI
- General Data Protection Regulation → GDPR
- general-purpose AI model with systemic risk → general-purpose AI model posing systemic risk
- Generative Engine Optimization → generative search engine optimization
- Generative Pre-trained Transformer → GPT
- GEO → generative search engine optimization
- GH → GitHub
- GHA → GitHub Actions
- Git tag → tag
- git worktree → worktree
- GLM → Zhipu AI
- Google Antigravity → Antigravity
- Google Vertex AI → Vertex AI
- GPAI → general-purpose AI model
- GPAISR → general-purpose AI model posing systemic risk
- GPU cluster → GPU-klaszter
- GPU rental / cloud capacity lease → GPU rental
- graphics processing unit (GPU) → GPU
- group relative policy optimization → GRPO
- GSC → Search Console (Google Search Console)
- guardrails → guardrail
- H+ → transzhumanizmus
- harmonized standard → harmonised standard
- HBM → HBM memory
- HCAI → human-centered AI
- hEN → harmonised standard
- HF → Hugging Face
- high-bandwidth memory (HBM) → HBM memory
- HITL → human-in-the-loop
- HPC → supercomputer
- HR agent (recruiting agent) → HR-agent
- Huawei Ascend → Ascend
- Hungarian National Authority for Data Protection and Freedom of Information → NAIH
- hype cycle → hype-ciklus
- HyperText Markup Language (HTML) → HTML
- Hypertext Transfer Protocol (HTTP / HTTPS) → HTTP
- I2V → image to video
- IB → InfiniBand
- ICL → in-context learning
- IDE → development environment
- image-to-video → image to video
- inference → inferencia
- inference chip → inferenciachip
- instruction tuning → instruction fine-tuning
- integrated development environment → development environment
- intellectual property → szellemi tulajdon
- IP → szellemi tulajdon
- IPO → initial public offering
- JavaScript Object Notation → JSON
- Jevons paradox → Jevons-paradoxon
- job displacement → job loss
- JS → JavaScript
- K8s → Kubernetes
- KD → Distillation
- Kling AI → Kling
- KV cache (key-value cache) → KV-cache
- KYA → Know Your Agent (KYA)
- LangChain / LangGraph → LangChain
- language model → nyelvi modell
- large language model → nagy nyelvi modell
- LCNC → no-code
- lib → library
- lip sync → lip-syncing
- LLM → nagy nyelvi modell
- LLM-as-a-judge → LLM as judge
- LM → nyelvi modell
- local deployment / on-premises → local execution
- Low-Rank Adaptation → LoRA
- LR → learning rate
- M&A → acquisition
- machine learning operations (MLOps) → MLOps
- market cap → market capitalization
- MAS → multiagent rendszer
- MCP → Model Context Protocol (MCP)
- MD → Markdown
- Meta Advantage+ → Advantage+
- MFA → multi-factor authentication
- MI (angolul: AI) → artificial intelligence
- Microsoft Copilot → Copilot
- mixture of experts → MoE
- MJ → Midjourney
- ML → machine learning
- MMM → marketing mix modeling
- model router / model picker → model selector
- MS → Microsoft
- MU → machine unlearning
- multi-agent system → multiagent rendszer
- narrow AI (artificial narrow intelligence) → Narrow artificial intelligence
- NCII → non-consensual intimate image
- neural processing unit → NPU
- NIS2 Directive (Network and Information Security Directive 2) → NIS2
- NN → Neural network
- no-code / low-code → no-code
- non-consensual intimate imagery → non-consensual intimate image
- Nvidia Blackwell → Blackwell
- Nvidia H100 → H100
- Nvidia NVLink → NVLink
- Nvidia Rubin (Vera Rubin platform) → Rubin
- o11y → observability
- OCI → Oracle
- on-prem → local execution
- OpenAI o-series (o1, o3, o4-mini) → o-sorozat (OpenAI)
- Oracle Cloud Infrastructure → Oracle
- orchestrator-worker pattern → orchestrator–worker minta
- Organic Traffic → organikus forgalom
- OSS → open source
- parameter-efficient fine-tuning → PEFT
- patch / hotfix → patch
- penetration testing → penetration test
- pentest → penetration test
- per-seat pricing → per-user pricing
- per-token pricing → token-based pricing
- permission management → authorization management
- physical AI → fizikai AI
- platform work / gig economy → platformmunka
- PMax → Performance Max
- PoC → proof of concept (PoC)
- PoLP → principle of least privilege
- power usage effectiveness → PUE
- PPA → power purchase agreement
- PR → pull request
- preference data → preferenciaadat
- privacy policy (privacy notice) → data processing notice
- process → folyamat
- prod → production environment
- Programmatic Advertising → programmatic ad buying
- prohibited AI practice → tiltott AI-gyakorlat
- Proximal Policy Optimization → PPO
- PS → PowerShell
- Quantized Low-Rank Adaptation (QLoRA) → QLoRA
- RAI → Responsible AI
- rate limit / usage limit → usage limit
- reasoning tokens / thinking tokens → thinking token
- Recommender System → recommendation system
- regex → regular expression
- reinforcement learning environment (RL environment) → RL environment
- reinforcement learning from AI feedback → RLAIF
- reinforcement learning from human feedback → RLHF
- reinforcement learning with verifiable rewards → RLVR
- release → expenditure
- repo → code repository
- repository → repo
- reskilling / upskilling → reskilling
- responsible scaling policy → responsible scaling guideline
- REST → REST API
- retrieval-augmented generation → RAG
- return on ad spend → ROAS
- reward model → jutalommodell
- right to one's own image (personality rights); USA: right of publicity / likeness rights → right to one's likeness
- RL → reinforcement learning
- RM → jutalommodell
- RMN → retail media
- robot tax / token tax → robot tax
- robotic process automation → szoftverrobot (RPA)
- RPA → szoftverrobot (RPA)
- RSI → recursive self-improvement
- RSP → responsible scaling guideline
- Safe Superintelligence Inc. (SSI) → Safe Superintelligence (SSI)
- safety classifier (constitutional classifiers) → security classifier
- safety evaluation → security assessment
- sampling steps → sampling step
- script → szkript
- SD → Stable Diffusion
- SDD → specification-driven development
- Search Engine Results Page → search results page
- secrets management → secret management
- Secure Shell (SSH) → SSH
- security patch → security update
- semiconductor foundry → contract manufacturer
- SemVer → semantic versioning
- SEO → search engine optimization
- SERP → search results page
- server → szerver
- serverless computing → serverless architecture
- session → munkamenet
- SFT → supervised fine-tuning (SFT)
- shared service center / global capability center → service provider hub
- SLA → availability
- SLM → kis nyelvi modell
- small language model → kis nyelvi modell
- smart home → okosotthon
- SMR → small modular reactor
- software architecture → architecture
- software as a service → SaaS
- software development kit → SDK
- software license → szoftverlicenc
- software supply chain attack → supply chain attack
- software testing → testing
- spec-driven development → specification-driven development
- speech recognition / speech-to-text → speech recognition
- SSC / GCC → service provider hub
- SSI → Safe Superintelligence (SSI)
- SSL → self-supervised learning
- SSO → single sign-on (SSO)
- standard output / standard error → standard kimenet
- Stargate Project → Stargate
- stdout / stderr → standard kimenet
- superintelligence (artificial superintelligence) → szuperintelligencia
- syntax → szintaxis
- synthetic data → szintetikus adat
- synthetic persona → szintetikus persona
- system prompt → rendszerprompt
- TAI → transformative AI
- Taiwan Semiconductor Manufacturing Company → TSMC
- TB → Terminal-Bench
- TDD → test-driven development (TDD)
- TDM → text and data mining exception
- TEE → confidential computing
- tensor processing unit (TPU) → TPU
- terms of service (terms of use) → terms of use
- test data → tesztadatok
- test-time compute → test-time computation
- text-to-speech, speech synthesis → speech synthesis
- TL → transfer learning
- TML → Thinking Machines Lab
- token budget / thinking budget → tokenkeret
- tool definition / tool schema → asset definition
- ToS → terms of use
- TPA → tool poisoning
- TPS → tokens per second
- transformer → transzformer
- transhumanism → transzhumanizmus
- TS → TypeScript
- TTC → test-time computation
- TTFT → time to first token
- TTS → speech synthesis
- Turing test → Turing-teszt
- UBI → Universal basic income
- UBP → usage-based pricing
- unicorn → unikornis
- Uniform Resource Locator (URL) → URL
- unit test → unit teszt
- upscaling → scaling up
- uptime / availability (SLA) → availability
- use of data for model training (training data opt-in) → data usage for model training
- VC → voice conversion
- VCS → version control
- vendor lock-in → provider dependency
- venv → virtual environment
- video RAM (GPU memory) → VRAM
- vision-language model → VLM
- VM → virtual machine
- voice agent → voice-based agent
- voice assistant → hangasszisztens
- voice conversion / speech-to-speech → voice conversion
- VS Code → Visual Studio Code (VS Code)
- Windows Subsystem for Linux → WSL
- x-risk → existential risk
- XAI → Explainable AI
- ZDR → zero data retention
- zero trust security → zero trust
- Zhipu AI / Z.ai → Zhipu AI
- ZT → zero trust