Start here
Large Language Model (LLM) in plain language
Tokens are pieces of text rather than necessarily whole words. During training, model parameters are adjusted to improve predictions. During ordinary use, the model processes the available context and produces output; it does not automatically update its training whenever you send a message.
A chat product can wrap a model with search, file processing, memory and tools. Those application features should not be confused with what the underlying model knows or can access. Fluent wording also does not tell you whether a claim is correct.
How it works
Receive context
The application supplies instructions, conversation and any retrieved information within the model’s context limits.
Generate tokens
The model produces a response based on learned parameters and the supplied context. Output can vary with model and generation settings.
Use the result
The application displays text or interprets an output, such as a proposed tool request. Validation and permissions belong to the surrounding system.
Illustrative example · Fictional data
Turning meeting notes into actions
- Input
- Fictional note: “Maya will send the draft on Tuesday. The budget owner has not been assigned.”
- Process
- Ask for a table of actions, owners and deadlines, using only the note. Require “not specified” for missing fields.
- Expected output
- One row should name Maya and Tuesday. Any budget-related row should leave the owner unassigned rather than invent a name.
- How to check it
- Compare every populated field with the input. Add a contradictory note and see whether the model flags the conflict instead of quietly choosing an answer.
Where it helps
- Drafting and restructuring text for a defined audience.
- Extracting fields from supplied documents with verification.
- Assisting with code, explanations and language tasks within a reviewed workflow.
Model, chatbot and agent
- LLM
- The model that processes context and produces output.
- Chat application
- The interface and supporting services around one or more models.
- Agent system
- An application that uses models and tools to pursue tasks through multiple steps with defined permissions.
Tokens, attention and the training sequence
Text is converted to tokens and numerical representations. In transformer-based models, attention combines information across permitted context positions, while other layers transform those representations. Generation scores possible next tokens and selects a continuation; fluent output does not imply a database lookup or a truth check.
Pretraining develops broad statistical capabilities. Instruction and other post-training stages shape how the model responds. Prompting changes the input context; fine-tuning changes trainable parameters; retrieval adds external evidence at inference time. These mechanisms solve different problems.
Context windows are not permanent memory
A context window limits the tokens available for a request, including applicable instructions, conversation, tool results and output budgeting. File-upload limits and product retrieval behaviour are separate constraints. More context does not guarantee that every detail will be used correctly.
Applications may persist conversations, retrieve stored notes or compact earlier content. Those are system behaviours, not proof the model permanently learned the conversation. Review the actual product’s retention, training-use and access policies before supplying confidential data; policies vary by surface and agreement.
Related reading and learning pathways
Common pitfalls
- An LLM can produce convincing unsupported details.
- A long context allowance does not guarantee that every detail will be used correctly.
- Access to current information depends on the application’s sources and tools; it should not be assumed from fluent output.
Put the idea to work
Try it yourself
Use the meeting note above and request an action table. Then remove Tuesday and repeat. A useful response should preserve the missing deadline instead of filling it in.
Common questions
Is an LLM a database?
No. Its parameters encode learned patterns rather than providing a reliable record lookup. An application can connect it to a database or search system. LLMs can produce plausible but incorrect facts, so connect them to a database or search system when an application needs accurate, current or account-specific information.
Does every LLM browse the internet?
No. Browsing is an application capability that must be provided and enabled. The model alone does not imply live web access. Products such as ChatGPT or Gemini can add web search, but the underlying model only knows its training data up to a cutoff unless a tool supplies fresh information.
Do I need coding skills to use an LLM?
Not for many chat-based tasks. Building an application around model APIs requires additional software and evaluation skills. Tools like ChatGPT, Claude and Copilot are usable through a chat interface. Building automations or integrations typically needs Python or JavaScript and some testing practice.
Your next step
Continue learning
Connected concepts
Sources and further reading
Examples on this page are teaching exercises using fictional information. They are not client case studies or measured outcomes.