What is the difference between AI and machine learning?

Machine learning is the part of AI where software learns patterns from examples instead of following rules a person wrote. Deep learning is machine learning with large neural networks, and LLMs are deep learning models trained on text, so each one sits inside the one before it.

How many words is 1,000 tokens?

About 750 words of English, using OpenAI's rule of thumb that a token is roughly three-quarters of a word. Code, numbers and most other languages use more tokens per word, so the same token budget holds less of them.

What is the difference between generative AI and an LLM?

Generative AI is any AI that creates new content: text, images, audio or video. An LLM is the kind that produces text. An image generator such as Midjourney is generative AI that works on pictures, so it sits outside the LLM category.

Can an LLM search the internet?

A model on its own knows only what was in its training data, which stops at a cutoff date. Apps such as ChatGPT, Claude and Perplexity add a web search step that fetches pages and hands them to the model, which is how they answer about recent events and link sources.

LLM vs AI: what LLM means, in plain English

In AI, LLM means large language model, one kind of AI (it's also a law degree). LLM vs AI vs chatbot compared, and 22 AI terms in a sentence or two.

What does LLM mean, and how is it different from AI?

In AI, LLM means large language model: a program trained on vast amounts of text to predict what comes next. An LLM is one kind of AI. AI is the whole field, and LLMs are the type of AI behind ChatGPT, Claude and Gemini.

Outside tech, LLM also means Master of Laws, a postgraduate law degree. Everything on this page is about the AI meaning.

"Large" refers to size on two counts. Current LLMs have billions or trillions of parameters (the internal numbers set during training) and learn from trillions of tokens of text. They're trained first to predict the next word or word fragment, then tuned with human feedback to follow instructions and hold a conversation. Every answer you get is produced that way, one token at a time.

LLM vs AI vs chatbot, compared

Three words get used as if they mean the same thing. They describe three different layers.

	AI	LLM	Chatbot (the app)
What it is	The whole field of software that does tasks we'd say need human intelligence	One kind of AI model, trained on text to predict the next token	The product you type into, built around one or more models
Examples	Email spam filters, Netflix recommendations, face unlock, chess engines, ChatGPT	GPT-6 Sol, Gemini 3.7 Flash, DeepSeek V4 Flash, Kimi K3	ChatGPT, Claude, Gemini, Perplexity
Works with	Whatever it was built for: photos, sound, numbers, text	Text, counted in tokens (multimodal models also read images and audio)	Your messages and files, plus tools such as web search
How you reach it	Usually built into another product	Through an app, or through an API if you're a developer	Website or phone app, free or paid
Remembers you	Depends on the system	Never by itself: its parameters don't change when you chat with it	Some apps save notes about you and re-send them in later chats

Every LLM is AI, and most AI in your life is something else. The spam filter in your inbox and the feed on YouTube are machine learning systems that sort and rank.

The last row explains a lot of odd behavior. A model starts each request blank. The app makes a chat feel continuous by sending the whole conversation back to the model every time you press enter, which is why the size of the context window (below) decides how long a chat can run before it forgets.

Is ChatGPT an LLM?

ChatGPT is a chatbot app. The LLMs inside it are OpenAI's GPT models: free accounts get GPT-6 Luna and GPT-5 Thinking Mini, and paid plans add GPT-6 Sol and others (OpenAI's pricing page, October 2026). GPT stands for generative pre-trained transformer, and the transformer is the network design almost every current LLM uses, introduced by Google researchers in the 2017 paper "Attention Is All You Need."

Claude and Gemini blur the line because each name covers both the app and the model family. The app is the part you can see: the chat window, history, file uploads and web search. The model is the part that writes the words.

AI terms in plain English: the glossary

Term	What it means
AI (artificial intelligence)	The field of building software that does tasks we'd say take human intelligence, such as recognizing faces, translating or answering questions. It covers everything from spam filters to ChatGPT.
Generative AI	AI that creates new content: text, images, audio or video. LLMs are generative AI for text.
LLM (large language model)	A model trained on a huge amount of text to predict the next token, which is how it writes answers. GPT, Claude, Gemini, DeepSeek, Qwen and Llama are LLM families.
Model	The trained system itself: a very large file of numbers plus the code that runs it. "GPT-6 Sol" and "DeepSeek V4 Flash" are model names.
Chatbot, model and app	The app is the product you open, the chatbot is the conversation it offers, and the model is the engine inside. One app can run several models and swap them without changing its name.
Prompt	Everything you send the model: your question, your instructions and any files. Apps also add a hidden system prompt of their own instructions in front of yours.
Token	The unit an LLM reads and writes, usually a short word or part of a longer one. OpenAI's rule of thumb for English is that a token is about three-quarters of a word, so 100 tokens is about 75 words.
Context window	The most text a model can take in at once, counted in tokens. It has to hold the conversation so far, any attached files and the app's instructions.
Parameters	The numbers inside a model that training adjusts. "70B" means 70 billion; more usually means more capable and more expensive to run.
Mixture of experts (active parameters)	A design where only part of the model works on each token. DeepSeek V4 Flash has 284 billion parameters in total and uses about 13 billion per token (DeepSeek's model card), which makes it far cheaper to run than its size suggests.
Training	Building a model by feeding it vast amounts of text and adjusting its parameters until it predicts well. It happens before release, costs a fortune, and sets a cutoff date after which the model knows nothing new.
Inference	Running a finished model to produce an answer. The company running it pays for the computing on every message, so models that are cheap to run make cheap apps possible.
Fine-tuning	Extra training on a smaller, focused set of examples to adjust an existing model for a task, a tone or a field such as law. It changes the model's parameters.
RAG (retrieval-augmented generation)	The app first looks up relevant material (web pages, your documents, a company's help articles) and places it in the prompt so the model answers from it. Web search with cited sources is a form of RAG.
Reasoning or thinking model	A model trained to work through a problem in hidden steps before it answers. It's better at math, code and multi-step questions, and slower and costlier on easy ones.
Hallucination	A false statement delivered with full confidence, such as an invented quote, statistic or citation. More below.
Agent	An AI given a goal and tools (a browser, web search, code, your files) that takes several steps on its own, deciding each next step from the last result. ChatGPT's agent mode and Anthropic's Claude Code are examples.
Multimodal	A model that handles more than text, such as reading photos, PDFs or audio, and sometimes producing images or speech.
Open weights	The model's parameters are published, so anyone with enough hardware can download and run it under the license's terms. DeepSeek, Qwen, Kimi and Meta's Llama models are released this way.
Open source (for AI)	A stricter label. The Open Source Initiative's definition (version 1.0) also requires the complete training code and detailed information about the training data, which most open-weight releases leave out.
Closed model	The parameters stay private, so you can use the model only through its owner's app or API, at the owner's prices and data policy, and the owner can change or retire it. OpenAI's GPT models, Claude and Gemini are closed.
AGI and superintelligence	AGI (artificial general intelligence) usually means AI as capable as a person across most tasks; superintelligence means AI far beyond people. Neither has an agreed definition or test, which is the subject of the last section.

The difference between open weights and closed models is a big part of why AI prices have fallen, because any host can serve a published model and compete on price. The longer version is in open weights vs closed models, and the cost side is in why Chinese models cost less.

How much text fits in a context window?

A token is about three-quarters of an English word, so a 54,000-token window holds roughly 40,000 words. Here are ChatGPT's consumer limits, as OpenAI lists them on its pricing page (October 2026), with the word count worked out from the token figure.

ChatGPT plan and mode	Context (input max)	Roughly, in words	OpenAI's own page estimate
Free, Instant	27,000 tokens	20,000	about 12 pages
Go or Plus, Instant	54,000 tokens	40,000	about 40 pages
Pro, Instant	128,000 tokens	96,000	about 250 pages
Go or Plus, Thinking	256,000 tokens	190,000	about 320 pages
Pro, Thinking	400,000 tokens	300,000	about 680 pages

OpenAI's page counts don't scale evenly with the token numbers, so the word column is the steadier guide. The free plan's Thinking limit is listed only as "varies."

The window has to hold the whole conversation, plus anything you've attached. When a chat outgrows it, the app drops or condenses the oldest parts, and the model loses track of things you said at the start. On ChatGPT Go I've had a chat lose details I'd given it earlier in the same conversation, mid-task, on the default Instant model. Switching to a Thinking model raises the limit on Go and Plus. The full plan-by-plan picture is in is ChatGPT Go worth it.

Why do LLMs hallucinate?

An LLM writes the most likely next token, and likely-sounding text can be false. It uses the same confident tone for a wrong date as for a right one. OpenAI's September 2025 paper "Why Language Models Hallucinate" adds a second cause: training and testing reward a model for guessing over saying it doesn't know.

Web search and RAG reduce the problem by giving the model real sources to work from, and they still leave room for misreading a source. Check dates, quotations, statistics and citations before you repeat them, because those are where models slip most.

The words I'd retire: AGI and superintelligence

If I could take two words out of everyday AI talk, they'd be "AGI" and "superintelligence." Nobody agrees on what either one means, so nobody can test for them or say when one has arrived.

The definitions in use disagree with each other. OpenAI's charter defines AGI as "highly autonomous systems that outperform humans at most economically valuable work." Google DeepMind researchers proposed a ladder of AGI levels in 2023, and François Chollet, who created the ARC benchmark, measures intelligence by how efficiently a system learns skills it hasn't seen. When Microsoft and OpenAI rewrote their partnership in October 2025, they agreed that any AGI declaration by OpenAI would be checked by an independent expert panel, and parts of the deal, including the revenue share OpenAI pays Microsoft, run until that panel confirms it.

Both words also do more selling than describing. Promises to build AGI appear in fundraising pitches and job ads for products that, today, answer questions and write code. My objection is to the words. The concerns people raise under them (safety, jobs, who controls powerful systems) are real and worth discussing in specific terms, and you can't argue clearly with a word that has no agreed meaning.

A specific claim works better. "This model writes good code and gets dates wrong" can be checked. "This is a step toward AGI" can't. My runner-up for retirement is "AI-powered," which tells you nothing about what a product does.

I built Plainly, a $5-a-month chat app, so the person using it never has to learn the words on this page.