LLM in one minute
An LLM does not keep a complete encyclopedia in which it reliably looks up sentences word for word. It learns patterns distributed across model parameters. Some training passages can nevertheless be memorised. A particular application may also give the model search tools or documents, which is different from relying on the trained model alone.
An LLM through a simple example
After a customer meeting, you have two pages of notes. You want a short list of decisions, tasks and unanswered questions. An LLM can organise that material into a useful format. If the notes say that delivery dates are still being discussed, however, it must not turn that uncertainty into a confirmed deadline. Good formatting and faithful reporting are separate qualities.
With Tool Use, an application can let the model request a search, calculation or file operation. Software executes the permitted request. AI Agents build on this by choosing further steps within limits and checking outcomes. The language model and the system performing actions remain distinct components.
- Provide the source
Supply the relevant notes and identify the current version.
- Define the task
Request confirmed decisions, task owners and unresolved questions.
- Check the result
Compare each commitment with the original notes before sharing it.
An instruction such as “write a concise summary for the project manager and leave missing owners unassigned” gives more direction than “summarise this”. A brief example can establish the desired layout. The resulting change in behaviour comes from information in the current context; it is not proof that the model's base weights have permanently changed during the conversation.
The same LLM might help classify enquiries, simplify a complicated letter or suggest a spreadsheet formula. Success with one task does not establish equal reliability with another. A strong writing performance can coexist with incorrect arithmetic, and a tidy summary can omit an important exception. Evaluate the tasks you actually intend to use, rather than treating one impressive answer as a general capability test.
A closer technical look
Text is commonly split into tokens, then represented numerically. Many LLMs use a Transformer, which processes relationships between parts of a sequence through attention and other network operations. LLM describes a kind of language model; Transformer names a widely used architecture. The two terms are related but do not identify exactly the same thing.
Training and parameters. Training changes numerical weights to improve performance on a learning objective. Many generative language models begin with predicting continuations of text. Additional training may improve instruction following, task performance or preferred behaviour. Parameter count describes one aspect of scale, not a universal ranking of usefulness.
An autoregressive language model calculates a distribution over possible next tokens. A selection procedure chooses a token, and the sequence grows step by step. Sampling settings can affect variation; the system does not necessarily choose only the single most probable token every time. Repeated local predictions can produce extended explanations, while also continuing an incorrect assumption introduced earlier.
The Context Window limits how much information can be included in the relevant processing. Instructions, earlier messages, documents and tool results may occupy that space. The whole training dataset is not inserted again for each question. Missing information in the current context therefore differs from information that was never represented reliably during training.
Reasoning. In this setting, reasoning refers to performance on tasks involving intermediate steps, such as comparison, inference or planning. Training and additional computation at use time can improve some of these performances. A visible step-by-step explanation is not necessarily a faithful account of every internal calculation, nor evidence that the model thinks like a person.
MoE and open weights. Some models use a Mixture of Experts, routing inputs through selected network components. Their total parameter count differs from the number active in a given computation. Another term, open weights, means that weights are made available under specified conditions. It does not by itself make all training data, code and usage rights unrestricted. The actual model description and licence determine what is available. Concrete examples include DeepSeek architecture and the downloadable models within Qwen.
What changes when an LLM can use tools?
An application may allow the model to propose a search or calculator call. Software executes an allowed action and returns the result. The model can then use that result in its response. Merely generating the sentence “I checked the catalogue” does not execute a catalogue search. Evidence of an actual retrieval or calculation is needed to support that statement.
| Setup | Information available | Useful check |
|---|---|---|
| Model without external retrieval | Learned patterns and the current context. | Does the question require current facts or a source? |
| Model with a document | Supplied passages plus model processing. | Does each important claim match the correct version? |
| Model with a tool | An actual retrieved or computed result. | Were the tool input and interpretation correct? |
RAG supplies retrieved material for generation and can support answers about current business information. It does not automatically retrain the model. Wrong selection, misreading and Hallucination remain possible even with sources. The workflow should allow an explicit statement that information is missing instead of demanding a complete-looking answer every time.
- Choose representative tasks with outcomes that can be checked.
- Specify which source takes precedence and how recent it must be.
- Use suitable tools for calculations and changing business data.
- Evaluate factual accuracy separately from writing quality.
- Compare the model, tools and review process as a complete system.
These distinctions help when comparing products from different providers. Ask about the documents, languages, tools and error tolerance relevant to your organisation. A benchmark score can provide context, but it cannot replace a practical evaluation using the work you need to complete and the corrections your team would have to make.
- Brown et al. · Language Models are Few-Shot Learners (2020)
- Vaswani et al. · Attention Is All You Need (2017)
- Stanford CRFM · On the Opportunities and Risks of Foundation Models
- Ji et al. · Survey of Hallucination in Natural Language Generation
- Fedus et al. · Switch Transformers
- Turpin et al. · Language Models Do Not Always Say What They Think