Large Language Models (LLMs)
An LLM is a large neural network trained on huge amounts of text to predict the next token in a sequence, which turns out to be a surprisingly capable way to generate and manipulate language.
What Is an LLM?
A large language model is a neural network — almost always a transformer — trained on huge amounts of text to do one core thing: predict the next token given the tokens that came before it. Trained at large enough scale, that simple objective produces a model that can write, summarize, translate, answer questions, and generate code.
Tokens
Tokens So FarThe sequence generated up to this point.
Transformer
The NetworkAlmost always the underlying architecture.
Next-Token Probabilities
One Core ObjectiveA probability for every possible next token.
Generated Sequence
Emergent CapabilitySimple prediction, repeated, becomes writing.
At each step, the model outputs a probability distribution over every possible next token, one is selected, appended to the sequence, and the process repeats. Generating a full response is really just this loop running many times.
What an LLM Is Not
Important
An LLM is not a database, and it does not do deterministic knowledge lookup. It generates the statistically likely continuation of text — which is often correct, because correct text was common in its training data, but it is not the same mechanism as querying a record.
That's exactly why LLMs are usually paired with something else in real applications: retrieval to ground answers in specific documents, or tools to take real actions and fetch live data. An LLM by itself is a powerful text engine — LLM plus retrieval and tools is what makes a genuinely capable application.
Code Example
The specifics vary by provider, but the shape of calling an LLM is generally the same — illustrative pseudocode, not a specific SDK:
# Illustrative pseudocode — not a specific provider's SDK
response = llm.generate(
prompt="Summarize the following text in two sentences:\n\n" + document,
temperature=0.2,
)
print(response.text)Common Mistakes
Thinking of an LLM as a database
It generates plausible text based on patterns, not a lookup against a fixed, verified source of facts.
Confusing training with inference
Training happens once, ahead of time, on a large dataset; inference is generating a response to a specific prompt at request time.
Assuming deterministic output
The same prompt can produce different output on different runs, especially with non-zero temperature — plan for that variability rather than assuming an exact repeat.
Interview Question
What is an LLM, and how does it generate text?
An LLM is a large neural network, typically a transformer, trained to predict the next token given the preceding ones. At inference time it repeatedly predicts a probability distribution over the next token, selects one, and appends it, generating text one token at a time. It's not a database or a deterministic lookup — it produces the statistically likely continuation of the input, which is why pairing it with retrieval or tools makes it far more capable for tasks that need current or private information.
What an interviewer may ask next
- Why can an LLM produce a fluent but factually wrong answer?
- What role does temperature play in generation?
- Why would you combine an LLM with retrieval instead of relying on it alone?
Explain It in 30 Seconds
An LLM is a large neural network trained to predict the next token in a sequence, given everything before it. Repeating that prediction one token at a time is how it generates full responses. It's not a database — it generates statistically likely text rather than looking up verified facts — which is exactly why real applications often pair an LLM with retrieval or tools instead of trusting its internal knowledge alone.