What is an LLM?
A Large Language Model is the AI system behind ChatGPT, Claude, and Gemini. It predicts the next word, very well, across an enormous amount of text it was trained on.
A Large Language Model is a neural network trained to predict the next piece of text given the previous text. Train it on enough text (the whole readable internet, books, code, papers) and the same simple objective produces something that can write essays, debug code, summarize meetings, hold a conversation, and answer questions about almost any topic.
The three layers of an LLM-powered product
- The base model. Claude, GPT-4o, Gemini, Llama, etc. Trained once by Anthropic, OpenAI, Google, or Meta. The weights are fixed at deployment time. You do not retrain a frontier model yourself, you call it through an API.
- The system prompt. The instructions you give the model at the start of every conversation. "You are a phone receptionist for an HVAC business. Speak warmly. Book on the calendar. Escalate emergencies." A good system prompt is 3,000 to 12,000 words for a real agent.
- The tools and memory. APIs the model can call (check calendar, book appointment, send SMS), plus a database of past interactions the model reads before each response. This is what turns a chatbot into an agent.
What an LLM is good at
- Reading natural language and extracting structured facts
- Holding a conversation that feels human
- Summarizing long documents
- Writing first drafts of almost any kind of text
- Translating between languages
- Answering factual questions when the answer is in its training data
- Following multi-step instructions when given a clear system prompt
What an LLM is bad at
- Knowing what it does not know. LLMs sometimes produce confident-sounding wrong answers (hallucinations). The fix is retrieval: give the model real source data instead of relying on training memory.
- Exact arithmetic. Use a calculator tool, not the model itself.
- Real-time information. Training data has a cutoff date. For current info, the model needs a web-search tool or a retrieval system that points at fresh data.
- Persistent memory. Without a memory layer, the model forgets everything between conversations. Memory has to be engineered in.
How LLMs power the Traccion product
Every Traccion agent (receptionist, marketing manager, visibility engine) is an LLM with a custom system prompt, real tools wired up, and a memory layer that grows automatically. The base model is usually Claude or GPT-4o. The differentiation is in the rest of the stack.
For a deeper read on how agents work, see What is an AI agent? and How AI agents learn over time.
Common questions
- What does LLM stand for?
- LLM stands for Large Language Model. It is a type of AI system trained on enormous amounts of text to predict the next word in a sequence. Modern LLMs include OpenAI GPT-4o, Anthropic Claude, and Google Gemini.
- How is an LLM different from a chatbot?
- An LLM is the underlying model. A chatbot is a product that wraps an LLM with a conversation interface, a system prompt, and sometimes tools. ChatGPT is a chatbot product built on top of the GPT family of LLMs.
- How does an LLM learn?
- Training happens once, offline, on terabytes of text. After training, the model weights are frozen. The model does not learn from individual conversations unless explicitly fine-tuned. New "learning" usually comes from new system prompts and from retrieved context, not from the base model changing.
- What is a context window?
- The context window is how much text the LLM can read at once. In 2026, leading models support 200,000 to 2,000,000 tokens (roughly 150,000 to 1,500,000 words). Longer windows let agents process more memory and more documents per call.
- Why are some LLMs better than others?
- Three things: the quality and breadth of training data, the number of parameters in the model (loosely correlates with capability), and the post-training tuning (RLHF and reasoning fine-tunes). In 2026, Claude 4.x and GPT-5 lead on most reasoning benchmarks, with Gemini close behind on multimodal tasks.
- Can an LLM run on my own computer?
- Small open-source LLMs (Llama, Mistral, Phi) run on a laptop with a recent GPU. Frontier models (Claude, GPT-4o, Gemini) are far too large to run locally; they live on cloud infrastructure and you reach them through an API.