How AI and large language models work

The AI Agent that builds and edits your B12 website is powered by a large language model (LLM), a type of artificial intelligence trained to understand and produce human-like text. You don't need to know how it works to use it, but a little background helps you understand why it behaves the way it does, and why the way you ask for things makes such a difference.

You already rely on this kind of technology every day. The predictive text that finishes your sentences, the voice assistant that answers a spoken question, and the customer-service chatbot that helps you track an order all work on the same idea. The AI Agent applies that process to building and editing your website.

What a large language model is

A large language model (often shortened to LLM) is software that has "read" an enormous amount of text and learned the patterns in how people write. From those patterns, it can respond to your requests in natural language.

This is a kind of generative AI, meaning it creates new content, such as text, images, or website code, based on the instructions you give it.

It works differently from traditional software. Traditional software follows fixed rules a developer wrote in advance. An LLM has no built-in checklist of tasks. Instead, it responds flexibly to whatever you ask, which is exactly why a clear request matters so much. For tips on asking well, see Writing clear prompts and context for the AI Agent.

How AI generates text

An LLM does not look up answers in a database the way a search engine does. Instead, it predicts what text is most likely to come next, one piece at a time, based on the patterns it learned. This is called next token prediction. Responses can be very focused and consistent or allow for more variety or creativity. This is one reason the same question or request can generate a response that is worded differently each time.

Because it works from probability and creativity, an LLM's responses are more like a highly probable guess than a verified fact. This is why it can give slightly different answers to the same question, and why its output should be read as a strong draft rather than a guaranteed source of truth.

How AI reads: tokens

An LLM doesn't read whole words the way you do. It breaks text into small chunks called tokens. A token is roughly four characters, or about three-quarters of a word, so a short sentence is a handful of tokens.

Tokens matter for two practical reasons. First, a model can only handle so many tokens at once, which limits how much text it can consider in a single request. Second, AI usage is often measured in tokens, so longer back-and-forth uses more of it. In practice, this means a very long message or document can run past the limit, and the AI may lose track of details from earlier in the conversation.

The context window: the AI's working memory

The context window is the total amount of information the AI can hold in mind while it responds. Think of it as the working memory for your conversation. Everything competes for that space: your messages, the AI Agent's replies, your business details, and any files you've shared.

That space has a limit. When the window gets very full or cluttered with unrelated detail, the AI has a harder time paying attention to everything in it, and the quality of its responses can slip. More information is not always better.

What this means for you:

• Keep your requests focused and specific.
  A narrow request leaves less room for the agent to guess or fill gaps with generic patterns from its training.
• Show an example of what you want.
  If you have a tone, a layout, or wording you like, paste it in, upload it, or describe it concretely. A clear example gives the agent better material to work from than an abstract instruction.
• Put the most important details first.
  The agent weighs everything in the window, so lead with what matters most — the key change, the must-have detail, the non-negotiable. Don't bury it at the end of a long message.
• Break big requests into smaller steps.
  Instead of one long message asking for ten changes at once, make a few focused requests. Each one keeps the working memory clearer and the result more accurate.
• Repeat important details from past conversations.
  Don't assume the agent has context from previous chats. If a past detail matters now, restate it.
• Save consistently important information where the agent can always see it.
  Details that matter for every request belong somewhere persistent. In B12, use your Business description for high-level context.
• Leave out detail that isn't relevant.
  More context isn't always better. Pasting in a long document the agent doesn't need just fills the window with noise. Include only what's useful for the task at hand.

Why AI sometimes gets things wrong

Sometimes an LLM states something that sounds completely convincing but is actually incorrect. This is called a hallucination. It happens for the same reason the AI works at all: it predicts plausible-sounding text rather than looking up verified facts. When it doesn't have the right information, it fills the gap with something that fits the pattern, even if it isn't true. Also, because these models learn from text written by people, they can pick up and repeat human biases.

Key terms

Here are the terms you'll come across most often, in plain language:

Frequently asked questions

→ Does the AI Agent search the internet for answers?

⇒ No. It generates responses by predicting likely text from patterns it learned during training, not by looking up live results like a search engine. That's why your own details and context are so important.

→ Why does the AI Agent sometimes give different answers to the same question?

⇒ Because it works from probability, not a fixed lookup. Each response is a fresh prediction, so the wording and even the approach can vary from one try to the next.

→ Why does the AI Agent seem to forget things?

⇒ The AI works within a limited memory for each conversation, and it doesn't carry details between separate chats. In a long session, restate the key facts so they stay in view.

→ Is everything the AI Agent tells me accurate?

⇒ Not always. It can state incorrect information confidently, so review its work and double-check any facts before you publish.