On this page
Definition
An LLM is a neural network trained on very large text datasets to predict the next token in a sequence, which enables general-purpose language understanding and generation.
Simple explanation
Imagine reading every book, article and web page ever written, then playing an infinite game of 'guess the next word'. That is, in essence, what an LLM does.
Because language encodes reasoning, summarisation, code, translation and instruction-following, a model that predicts words well ends up being able to do all of those tasks.
Why it matters
LLMs power almost every modern AI product you have heard of — ChatGPT, Claude, Gemini, Copilot, Perplexity. Understanding how they work explains both their strengths and their failure modes.
For businesses, LLMs are the raw material. Every AI feature — search, chat, summaries, drafts — is a wrapper around an LLM plus your data.
How it works
- 1PretrainingThe model reads a very large text corpus and learns statistical patterns.
- 2Fine-tuningIt is refined on curated data to be helpful, safe and format-aware.
- 3AlignmentHuman feedback teaches it which answers people prefer (RLHF).
- 4InferenceAt use time it predicts the next token, one at a time, given your prompt.
Real examples
Products named for illustration only. Inclusion is not an endorsement.
- GPT-5 (OpenAI)Powers ChatGPT and the OpenAI API.
- Claude (Anthropic)Focus on safety, long context and coding.
- Gemini (Google)Native multimodal reasoning across text, image and video.
- Llama (Meta)Open-weight family used inside many self-hosted stacks.
Advantages
- General-purpose — one model handles many tasks.
- Improves with better prompts and better data.
- Available via API, so it slots into existing software.
- Rapid pace of capability improvement.
Limitations
- Can generate confident but false statements ('hallucinations').
- Knowledge is frozen at training time unless augmented with search or RAG.
- Costs, latency and rate limits still matter at scale.
- Bias and safety issues require ongoing evaluation.
Common misunderstandings
- ClaimLLMs 'know' things the way a person does.RealityThey pattern-match on training data; they do not have beliefs or memory between calls.
- ClaimBigger models are always better.RealityFor many tasks, a smaller fine-tuned model outperforms a giant general one.
Frequently asked questions
Is an LLM the same as AI?
No — AI is the broader field. LLMs are one very successful kind of AI model.
Do LLMs learn from my conversations?
Only if the provider says so. Enterprise plans typically exclude your data from training.
Which LLM is best?
It depends on the task — coding, reasoning, cost and latency all trade off differently.
Can I run an LLM on my laptop?
Yes — open-weight models like Llama and Mistral can run locally, though quality varies with size.
Why do LLMs make mistakes?
They predict likely text, not verified fact. Grounding with RAG or search reduces errors.
The Tool Money Lab perspective
The most important business question about LLMs is not 'which model is best' but 'how do we structure our data so any model can be useful'. Model choice will keep changing; your data will not.
Model prices have fallen roughly 10x per year for the last three years. Assume the LLM you pick today will be cheaper, faster and smarter — but also replaceable — within twelve months.
Conclusion
LLMs are the engine of modern AI. They predict the next word, and out of that emerges everything from writing assistance to coding to reasoning.
The winning use of LLMs in business is boring: connect them to your data, keep humans in the loop, and rebuild slow processes around them.