nagy nyelvi modell (LLM)
in English: large language model
An artificial intelligence program trained on huge amounts of text that interprets your question or request and responds in human language with coherent text.
In plain terms
The large language model is the "engine" working behind chat programs like ChatGPT, Claude, or Gemini. You type a question or request, and the model writes text in response: a letter, summary, translation, explanation, or even program code.
Imagine it as someone who has read an entire library, even a significant part of the internet. In the process, they did not memorize facts but learned how words and thoughts usually follow one another. This is why they can write fluently on almost any topic and why they sometimes err confidently: they are not reading from an encyclopedia but writing a probable continuation.
It differs from previous programs in that it was not designed for a single task. The same model can write a recipe, summarize a contract, and explain a math problem because it receives the instruction in everyday language, not through menus and buttons.
The word 'large' refers to size. Inside the model, there are billions of adjustable values (parameters), and in the largest models, far more. Training them requires huge amounts of text and computing capacity.
How it works
- Training: the model receives vast amounts of text and repeatedly practices predicting the next word. After every error, its internal settings adjust slightly, improving its guesses. Further training, including based on human feedback, then tunes it to provide helpful and safer responses.
- Tokenization: when you write to it, it breaks your text into small units, so-called tokens (these are words or word fragments), and converts them into numbers.
- Finding connections: using an architecture called the transformer, it checks, for every token, which parts of the rest of the text matter to it. This is how it can decide, for example, who "he" or "she" refers to in a sentence.
- Answer generation: the answer is generated token by token, always selecting a probable continuation. There is some randomness in this, so you may receive different answers to the same question twice.
- Additions: in today's systems, the model is often given tools, such as web search or file management. If it solves a task independently in multiple steps, we speak of an AI agent.
An everyday example
You ask the chat program on your phone to draft a polite complaint letter to your service provider, and in half a minute the text is ready. Behind the scenes, a large language model wrote it: it assembled probable sentences piece by piece to fit your request.
Who is affected
A common misconception
Limits and risks
- It may confidently state falsehoods, even fabricating non-existent sources or quotes (this is called hallucination), so verify important facts.
- Its knowledge is valid only up to the training date. It knows about events since then only if a search tool is connected to it.
- It can adopt the biases and prejudices of the training texts, so its response is not necessarily neutral.
- With online services, what you type goes to the provider's server and may be stored depending on settings. Therefore, share sensitive personal or corporate data with them only after careful consideration.
Similar, but not the same
- chatbot – The chatbot is the ready-made application you converse with; the large language model is the engine powering it.
- AI-agent – An agent is a system built on a large language model that not only answers but independently executes tasks in multiple steps using tools.
- generative AI – This is the broader category that includes image, audio, and video generation systems. The large language model is the text-based branch of this, although many modern models can also interpret images and audio.
Related concepts
transformer, token, prompt, hallucination, AI-agent
A brief history
The foundation of today's large language models is the transformer architecture, introduced by Google researchers in 2017 in the paper 'Attention Is All You Need'. The first GPT models built on this appeared from 2018. The general public first encountered the technology on a mass scale in November 2022 with the launch of ChatGPT.
Where things stand
By the end of 2026, large language models are everyday tools: they are present in chat programs, search engines, office software, and phones. The focus of development has shifted toward independently working agents, reliability, and cost reduction.
In the news
- Hungarian court makes new EU referral on press publications and LLM chatbotsNov 3, 2026 · Jog
- GitHub adds prompt-injection scanning to AI agent pull requestsOct 28, 2026 · Code
- Radiologists and LLMs fail to spot AI-generated chest x-raysOct 15, 2026 · Health
- FDA grant funds an LLM "jury" to evaluate AI-generated radiology reportsSep 28, 2026 · Health
- Open source in the LLM era: maintainers face a flood of AI-generated PRs and issuesSep 21, 2026 · Code
- LLM-written PhantomRaven malware steals npm and GitHub credentialsSep 15, 2026 · Code
- Czech analyses: LLM-generated code has been equally flawed for four years, only the volume is growingSep 14, 2026 · Code
- EFF and artist groups file amicus briefs in AI copyright casesAug 31, 2026 · Jog
Also known as
LLM · LLM-ek · nagy nyelvi modell · nagy nyelvmodell · large language model · nyelvi modell · large language model · large language models
Sources
This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.