Large Language Models and Chatbots

A large language model is a very big neural network trained on enormous amounts of text to predict what comes next, and a chatbot is simply a friendly conversation window wrapped around one.

Taking the name apart

• Large: huge in two ways: the amount of text it learned from (trillions of words) and the number of dials inside it (billions or more). Lesson 2, Section 6 covers why size mattered so much.
• Language: it works with human language, and also with computer code, math, and increasingly with images and sound.
• Model: the trained file of numbers (Lesson 2, Section 7).

The LLM vs. the chatbot

People often use these words as if they mean the same thing. They don't, quite:

The same LLM can power many different products. A customer-service bot on a bank's website, a writing assistant in your email, and a homework helper might all run on the same underlying model with different instructions.

How a chat actually works, behind the curtain

Here's something most people never learn: the model doesn't really "have a conversation." Each time you send a message, the app bundles up a script and hands the whole thing to the model:

1. Hidden instructions from the company (called a system prompt): "You are a helpful assistant. Be polite. Today's date is…"
2. The entire conversation so far, every message from you and every reply.
3. Your newest message.

The model reads this script and predicts what the "assistant" character would say next, one piece at a time. When you reply, the app sends the whole script again, now one exchange longer. It's a bit like an actor who is handed the full play every time it's their turn to speak, and improvises the next line.

What LLMs are good at

• Writing and rewriting: drafts, summaries, simpler versions, different tones.
• Explaining: breaking down hard ideas, answering follow-up questions patiently, giving analogies.
• Brainstorming: names, ideas, angles you hadn't considered.
• Language tasks: translation, grammar, reading comprehension.
• Code: writing and fixing computer programs.
• Reasoning through problems, especially newer models that "think" step by step before answering.

What they struggle with

• Precise facts they weren't sure of (see Section 4).
• Very recent events, unless they can search the web.
• Exact counting and letter-level tasks (see Section 2 for why).
• Knowing what they don't know (see Section 3).
• Your personal situation: they only know what you tell them.

"Reasoning" models

Since about 2024, many chatbots can spend extra time "thinking" before answering: writing out steps privately, checking their work, trying another approach. It's like the difference between blurting an answer and working it out on scratch paper. This makes them much better at math, logic, and planning, though it takes longer and it doesn't eliminate mistakes.

LLMs don't read letters or whole words; they read tokens, which are common chunks of text (a whole short word, part of a longer word, or a bit of punctuation), and nearly everything about a chatbot is measured in them.

What it looks like

Here's roughly how a model might split some text. (Each company's model splits a little differently, so these are illustrations.)

Rules of thumb

• In English, one token is about three-quarters of a word, or about four letters.
• 100 tokens ≈ 75 words, about one short paragraph.
• 1,000 tokens ≈ 750 words, about a page and a half of a typical book.
• Many other languages take more tokens to say the same thing, because the token "bricks" were mostly designed around English-heavy data. That can make AI slower, costlier, or a bit weaker in those languages.

Why tokens matter to you

1. Limits are counted in tokens. How long a conversation can get, how big a document you can upload, and how long an answer can be are all measured in tokens.
2. Cost is counted in tokens. Companies that sell AI to businesses usually charge per million tokens read and written. "Free" chatbots have usage limits for the same reason.
3. Some odd mistakes come from tokens. Because the model sees chunks, not letters, tasks like "How many r's are in strawberry?" or "Write a sentence with exactly 12 words" have tripped chatbots up. Newer models handle many of these better, often by working step by step, but it's a good example of how the AI "sees" text differently than you do.

Tokens beyond text

Models that handle images and sound turn those into tokens too: a picture is chopped into small patches, and audio into short slices. Everything becomes a sequence of bricks the model can predict.

A chatbot's tone comes from patterns in polished human writing and from training that rewarded helpful-sounding answers, not from checking how sure it actually is, so it can sound exactly as certain when it's guessing as when it's right.

Reason 1: it learned from confident writing

Books, encyclopedias, news articles, and textbooks are written in a clear, sure voice. People rarely publish "I think maybe this is right, but I'm not sure." So the patterns the model absorbed are mostly patterns of confident prose. Its natural "voice" is an authoritative one.

Reason 2: it was coached toward pleasing answers

In the feedback stage of training (Lesson 2, Section 5), people rated answers. Clear, direct, complete answers usually got better ratings than hedged ones or "I don't know." Without careful design, that teaches the model that sounding sure is rewarded. AI companies now work specifically to reward honest uncertainty instead, with real but incomplete success.

Reason 3: its tone isn't wired to its certainty

Inside, a model does have something like confidence: some next-word guesses are much more likely than others. But that internal certainty doesn't automatically control the words it chooses. "The answer is definitely X" and "the answer might be X" are just different word patterns, and it can pick the confident phrasing even when its underlying guess is shaky.

Reason 4: it tends to agree with you

Chatbots can be sycophantic: inclined to go along with the user. If you ask "Isn't it true that…?", it may confirm something false. If you push back on a correct answer ("Are you sure? I think it's Y"), it may cave and apologize even when it was right the first time. That can feel polite, but it makes the chatbot less trustworthy.

A hallucination is when a chatbot states something false or invented (a fake quote, a nonexistent book, a wrong date) as if it were true, because it's built to produce plausible text, and plausible and true are not the same thing.

What hallucinations look like

• Invented sources: book titles, research papers, or web links that don't exist, often with realistic-looking authors and dates.
• Fake quotes: words attributed to real people who never said them.
• Wrong details: the right person, the wrong year; the right city, the wrong population.
• Made-up people or events, especially for obscure topics.
• Misreading a document you gave it: "summarizing" a point that isn't actually in your file.
• False claims about itself: saying it did something (like checked a website) that it didn't do.

A real-world example

In a widely reported 2023 case in New York, lawyers filed a court brief that cited several earlier court decisions. The cases turned out not to exist; a chatbot had invented them, complete with realistic names and quotations. The judge sanctioned the lawyers. Since then, courts in many places have seen similar incidents, and many now require lawyers to disclose or verify AI use.

Why it happens

1. Prediction, not lookup. The model generates the most likely-sounding next words (Lesson 2, Section 3). It isn't checking a database of facts.
2. Blurry memory. Knowledge is stored as compressed patterns, not exact records (Lesson 2, Section 7). Frequently repeated facts are stored firmly; rare ones are fuzzy.
3. Pressure to answer. Training has historically rewarded giving an answer over admitting ignorance.
4. Patterns without the thing itself. The model knows what a citation looks like: author, year, title, journal. It can produce that shape perfectly, whether or not a real paper fills it.

When hallucinations are most likely

Is it getting better?

Yes, substantially, though not to zero. Newer models admit uncertainty more often, search the web and show their sources, and "think" before answering. Giving the chatbot trustworthy material to work from (a document, a web search) also helps a lot. But because generating plausible text is how these systems work at their core, most researchers expect some level of hallucination to remain. Plan on checking.

A word about the word

Some people dislike the term "hallucination" because it makes a machine error sound like a human experience, or makes it sound rare and strange when it's a normal side effect of how the system works. Other terms you may see: "confabulation" (making up plausible stories to fill gaps) or simply "fabrication." They all mean the same thing: false output presented as true.

The context window is the maximum amount of text, measured in tokens, that a chatbot can look at in one go (its hidden instructions, your whole conversation, any files, and its reply), and anything that doesn't fit simply isn't seen.

What goes on the desk

• The company's hidden instructions (system prompt)
• Any saved memories or preferences the app adds
• Every message in the current conversation, yours and its
• Documents, images, or web pages you've shared or it has searched
• Room for its answer (and its behind-the-scenes "thinking")

How big are context windows?

They've grown dramatically:

Note that the free version of a chatbot app may give you a smaller window than the model's maximum.

What happens when it overflows

Different apps handle a full desk differently:

• Some quietly drop the oldest messages, so the chatbot "forgets" the beginning of a long conversation.
• Some summarize the earlier conversation to make room, keeping the gist but losing details.
• Some simply tell you the conversation is too long and ask you to start a new one.

Big desk, but not perfect attention

Even when everything fits, a chatbot doesn't pay equal attention to all of it. Researchers have found that models can be better at using information near the beginning and end of a long text than in the middle. (This has been nicknamed "lost in the middle.") Newer models are much better at this, but on very long documents, important details can still slip.

Very long conversations can also drift: the chatbot may lose track of instructions you gave early on, or start repeating itself.

Context window vs. training

Keep these separate in your mind:

• Training knowledge is what the model learned months ago, baked into its numbers. Big, blurry, and fixed.
• The context window is what's on the desk right now. Small by comparison, but sharp and specific.

That's why putting a document on the desk ("answer from this") makes answers much more accurate than asking from memory.

The model itself doesn't remember you, but many chatbot apps now add memory features that save notes about you and slip them onto the "desk" in future conversations, and separately your chats may be stored by the company and possibly used to train future models, depending on your settings.

Four different kinds of "remembering"

People mix these up constantly. They're very different:

The good side of memory

Memory can make chatbots genuinely more helpful. You don't have to re-explain your situation every time. It can remember that you prefer detailed explanations with analogies, that you're working on a particular project, or that you like shorter answers. For people who find it tiring to repeat themselves, that's a real benefit.

The careful side

• It may remember more than you expect, including something you mentioned in passing.
• It may remember things wrong, or apply an old note in a context where it no longer fits.
• Stored data is still data. It sits on a company's servers, where it could be subject to breaches, legal requests, or policy changes.
• Work and personal can blur if you use the same account for both.

What never to share with any chatbot

• Passwords, PINs, or security-question answers
• Full Social Security, passport, or ID numbers
• Bank or credit card numbers
• Other people's private information without their permission
• Confidential work material, unless your employer has approved that specific tool

A handful of companies, mostly in the United States and China, build the most capable language models; they compete fiercely, release new versions every few months, and differ in their priorities, business models, and how open they are.

The big chatbots

Who's actually being used?

One market-research firm's estimate of U.S. chatbot usage share, published September 2026:

Source: First Page Sage, "Top Generative AI Chatbots by Market Share," September 2026 (U.S. only). Different research firms measure differently and get different numbers. The same report found ChatGPT's share falling through 2026 as more people use several chatbots.

These figures don't capture AI people use without opening a chatbot app, such as AI answers at the top of Google searches or assistants built into phones and office software, which reach enormous numbers of people.

Ways these companies differ

• Open vs. closed. Meta, DeepSeek, Alibaba, and Mistral have released many models openly for anyone to download. OpenAI, Google, and Anthropic mostly keep their top models closed (though some have released smaller open ones). See Lesson 2, Section 7.
• Business model. Most offer a free tier with limits and paid subscriptions (commonly around $20 a month for individuals, with pricier tiers for heavy use), plus business and developer plans. Some are exploring advertising.
• Ecosystem. Google and Microsoft build AI into tools billions of people already use. Others rely on their own apps and partnerships.
• Stated priorities. Companies emphasize different things: safety research, speed of release, openness, or integration. Their public statements are a starting point, not proof; watch what they do.

The bigger picture

• A small club. Only a few organizations can afford to train the most capable models, which concentrates enormous influence.
• A US–China race. The two countries lead AI development, and governments treat AI chips and models as matters of national security. Export controls on chips (and, briefly in 2026, on certain AI models) are part of that story.
• Rapid change. Top models are typically replaced within months. Rankings on public leaderboards shift constantly.