Ask for the same sentence twice
You can watch a large language model at work in two minutes, with no code. Open a chat with Claude, ChatGPT, Gemini or Copilot and send "Write one sentence about a rainy Monday." Then open a brand-new chat, send the same words and put the two replies side by side.
Every reply quoted here is a real run from 4 November 2026, each in a new Claude Code session with no history and no tools. The model is Claude Haiku 5.5 unless we name Sonnet. The rainy-Monday prompt got these two:
"Rain tapped against the office window all morning, and the coffee went cold before anyone wanted to admit it was Monday."
"Rain tapped against the office window all morning, and somehow that made the Monday coffee taste better."
The two share their first nine words, through "and", then go separate ways. Neither sentence was stored anywhere. The model built each one on the spot, a small piece at a time.
Those pieces are tokens. A token can be a whole word, part of a word or a single character, which is why a token count isn't a word count. To write a reply, the model reads all the text so far, scores every possible next token by how likely it is to come next, and adds one. Then it does it again, with the new token now part of the text.
Which token gets added is the app's call. It can take the top-scoring one every time, or draw at random from the tokens above a set probability. When an app samples like this, each reply is one draw from many reasonable ones, which is what the two rainy Mondays show. The setting that controls how adventurous each draw is has a name, temperature. Even at 0, the output isn't fully deterministic.
For a caption or an email subject line, that variety costs nothing, so keep whichever version you like. It gets in the way when a weekly report must keep one shape: save a good reply and paste it into the next prompt as the example. With a fact, two dates for one event mean at least one is wrong. And the same date twice proves nothing, since both came from the same training, not from a second source.
Ask about something newer than the model
A rainy Monday can't be out of date. A question about something that changes can. We asked Haiku "What is the newest iPhone model?" Its reply began "The newest iPhone I know of is the iPhone 17 lineup, announced in September 2025" and went on:
"Apple usually releases new models each September, and today is November 2026, so there may be a newer model I can't confirm. Check Apple's website for the current lineup."
The reply names its limit and points to a page that would know. That limit is the knowledge cutoff, where a model's training text ends: June 2026 for Haiku, and its models "may not be aware of events or information that occurred after their respective cutoff dates".
Our prompt gave no date. The training text stops in June. The month came from Claude Code, which puts today's date into the instructions it sends along with every message.
Prices, versions, laws and who holds which job keep changing after the training text is collected. Even the newest cutoffs in the table stop in June 2026, so a reply about them can be months behind without saying so.
A model also can't know what's in your own files or your company's documents. Tools can bring some of that in. With retrieval, the app fetches documents that match your question and passes them into the context window, the text the model reads for one reply, along with your message. Paste a document yourself and the model works from what you gave it, up to the limits our context window lesson lists. To try it, paste a paragraph you wrote, ask something it doesn't answer, and see whether the reply stays inside your text.
| Model | Knowledge cutoff | Source |
|---|---|---|
| Claude Fable 5.1, Opus 5.5, Sonnet 5.5, Haiku 5.5 | June 2026 | Anthropic help center and models overview |
| GPT-6.1 Terra | June 1, 2026 | OpenAI models page |
| Claude Opus 5 | May 2026 | Anthropic help center |
| GPT-6 Luna | May 18, 2026 | OpenAI models page |
| GPT-6 Astra, GPT-6.1 Sol | April 30, 2026 | OpenAI models page |
| Claude Sonnet 5, Fable 5, Opus 4.8, Opus 4.7 | January 2026 | Anthropic help center |
| Claude Sonnet 4.6, Opus 4.6 | August 2025 | Anthropic help center |
| Claude Haiku 4.5 | July 2025 | Anthropic help center |
| Gemini models | Not listed on Google's models page | Google AI Studio |
A famous name and a made-up one
Next, pair a famous name with one you've just made up. We asked Haiku "Who was Marie Curie? Answer in two sentences." and then the same about Talvenor Brisk, a name invented for this lesson. The Curie reply opened with her name and her years, 1867 to 1934, and went on:
"…was a Polish-born physicist and chemist who pioneered research on radioactivity, discovering the elements polonium and radium with her husband Pierre. She was the first woman to win a Nobel Prize, and she remains the only person to win Nobel Prizes in two different sciences, physics in 1903 and chemistry in 1911."
For the made-up name, it declined:
"I don't have any reliable information about a person or entity called Talvenor Brisk, so I can't say who they are. If you tell me where the name came up, such as a book, game, company, or article, I can help you work out what it refers to."
Sonnet, given the same prompt, also had no reliable information and offered that the name "may be a private individual, a fictional character, or a name outside my knowledge".
Neither reply came from a lookup. The famous name drew a dense, dated answer out of the training text. The invented one had nothing to draw on, and both models said so instead of writing a biography.
The Curie reply carries dates, a husband's name and two prizes in one even voice. A wrong date would read just as smoothly. Our hallucinations lesson explains why models make things up and gives five steps for checking an answer.
An invented name is the easy case, since the model has nothing to draw on. Most names you'll ask about sit in between, like a small local business or an author with one book, and our runs didn't test those. Ask about a café in your town and check one detail of the reply against its own website. Do that for any name, quote or citation that will leave your screen.
Where the guesses come from
Asked for one sentence, both rainy-Monday replies above stopped after one. Prediction alone doesn't give a model that habit. Pretraining has one goal, to continue text: models like the one under Claude "are pretrained to predict the next word, given the previous context of text in the document". A model fresh from pretraining isn't inherently good at answering questions or following instructions.
Sonnet, the other model in our runs, was trained on a proprietary mix of publicly available information from the internet, public and private datasets and other sources. Substantial post-training followed, with the goal of making it an effective assistant. Those later stages are why a chatbot answers your question instead of carrying on your sentence.
Pretraining
The model reads a huge amount of text and learns to predict what comes next in it.
Fine-tuning
More training on chosen examples turns a bare language model into a helpful assistant.
Human feedback
People rank two or more example answers, and training moves the model toward outputs like the higher-ranked ones.
Your message
Your text, plus any hidden instructions the app adds, is cut into tokens.
Generation
The model adds one token at a time until the reply is done.
Is it only autocomplete? What Anthropic found inside Claude
Google's machine learning course calls large models "essentially autocomplete mechanisms". That fits the training goal and undersells what goes on inside.
Language models "aren't programmed directly by humans": they learn their own strategies from the data. On 27 March 2025, Anthropic's interpretability team published a study that looked for those strategies inside Claude 3.5 Haiku, an older model than the ones in our runs.
Given the line "He saw a carrot and had to grab it," to continue in rhyme, the model picked "rabbit" before it began the second line. Then it wrote toward that word: "His hunger was like a starving rabbit." It sometimes used a conceptual space shared between languages. To add two numbers, it ran parallel paths, one for a rough total and one for the exact last digit.
Asked how it got 36 + 59 = 95, the model described the standard method of carrying the 1, not the parallel paths the researchers saw. Its account of its own steps is more text it wrote. It can differ from what happened inside.
Step-by-step reasoning can do the same. Asked for the cosine of a large number it couldn't easily calculate, Claude sometimes claimed to have run the calculation, and the researchers found no evidence of one inside.
Claude sometimes makes up plausible-sounding steps to get where it wants to go.
Who you are asking, in Andrej Karpathy's words
Picture a person behind the reply and you'll weigh it better. Andrej Karpathy, an AI researcher and educator who was a founding member of OpenAI and later Tesla's Director of AI, made that case on X on 29 November 2024. Chat models learn largely by imitating data that human labelers produced, so "asking an AI" is closer to asking the average labeler. He named the exceptions himself. Companies hire skilled labelers for code, math and creative writing, and reinforcement learning shifts the picture.
That is the human feedback step from the timeline above, seen from your side of the chat. Ask for a judgment, such as which of two job offers to take, and you get a polished average view. Ask for the case for each side instead, and make the call yourself.
Anthropic put its microscope study on X the day it came out. The post calls what the team found inside Claude "often complex and surprising", and the study admits its method captures only a fraction of the total computation, even on short, simple prompts. The people who build these models can't yet read every step either.
you're not asking an AI, you're asking some mashup spirit of its average data labeler.
Screenshots of X's own embeds of the posts, taken on 5 November 2026, with the time in UTC. X cuts Karpathy's post short with Show more. Select one to open the post on X.
Next: choosing a model
Every reply on this page came from Claude. ChatGPT, Gemini and Copilot run on other models, trained on other text with other cutoffs. The next lesson, how to choose an AI model, compares them and helps you pick one for a given task.
Check yourself
During pretraining, what is a language model trained to do?
Show answer
B. Predict the next token from the text before it Pretraining teaches the model to guess what comes next from the text before it, and answering questions comes from later training.
You send one prompt to two new chats and get differently worded replies. What is the most likely reason?
Show answer
C. Each token is drawn with some chance, so the wording can split at any step Like the two rainy-Monday sentences, the replies share a model and a prompt but not every pick along the way.
A model's training text runs to June 2026, and it has no search tool. Which question is it most likely to get wrong?
Show answer
D. What happened in the news yesterday? Yesterday's news came after the cutoff, so only a search tool or your own message can bring it in.
In Anthropic's 2025 study, how did Claude 3.5 Haiku write a rhyming line?
Show answer
A. It picked the rhyme word first and wrote toward it It had chosen "rabbit" before starting the second line, then wrote a line that ended there.
Frequently asked questions
Is ChatGPT the same thing as a large language model?
No. ChatGPT is the chat app, and a language model is the engine it runs on. An app can even switch engines: by default, Microsoft's Copilot uses a real-time router to pick the underlying model based on your prompt.
How big is a large language model?
Hundreds of billions of parameters, sometimes trillions, by the count in Google's machine learning course. Parameters are the values a model learns in training, and the same page says transformers with more of them consistently outperform those with fewer.
Does the model look things up on the internet when it answers?
Only when the app gives it a search or retrieval tool. Otherwise it answers from its training text and whatever you or the app put in the chat. Microsoft's Copilot shows how much the app decides: Copilot Chat (Basic) is grounded in web data and uses work files only when you upload or open them. A Microsoft 365 Copilot (Premium) license also grounds answers in work data automatically.
Does the model learn from my chat while we talk?
Not by changing itself. Anthropic's glossary calls the context window a "working memory" for the model, separate from the data it was trained on. Some apps also save details from past chats and add them to new ones, but that's the app passing text along, not the model changing. Whether a vendor later trains new models on your chats depends on its policy, and the privacy lesson shows where to find it.
Sources
- Anthropic: Glossary https://platform.claude.com/docs/en/about-claude/glossary
- Anthropic: Transparency Hub https://www.anthropic.com/transparency
- Anthropic Help Center: How up-to-date is Claude's training data? https://support.claude.com/en/articles/8114494-how-up-to-date-is-claude-s-training-data
- Anthropic: Models overview https://platform.claude.com/docs/en/models/overview
- Anthropic Research: Tracing the thoughts of a large language model https://www.anthropic.com/research/tracing-thoughts-language-model
- Google: Machine Learning Crash Course, Introduction to large language models https://developers.google.com/machine-learning/crash-course/llm
- Google: Machine Learning Crash Course, LLMs: What's a large language model? https://developers.google.com/machine-learning/crash-course/llm/transformers
- Google AI Studio: Gemini models https://aistudio.google.com/docs/models
- OpenAI: Models https://developers.openai.com/api/docs/models
- Microsoft Learn: What is Microsoft Copilot? https://learn.microsoft.com/en-us/copilot/plan/copilot-overview
- Andrej Karpathy on X, 29 November 2024 https://x.com/karpathy/status/1862565643436138619
- Anthropic on X, 27 March 2025 https://x.com/AnthropicAI/status/1905303835892990278
- YouTube: Andrej Karpathy, Deep Dive into LLMs like ChatGPT https://www.youtube.com/watch?v=7xTGNNLPyMI