Plain-English explanation
An LLM predicts what text is most likely to come next. Everything a chatbot appears to "do" (hold a conversation, stay in character, answer a question, write a story) comes from that model generating one small piece of text at a time, each piece chosen based on everything written so far, then feeding its own output back in to generate the next piece until a response is complete.
"Large" refers to scale: modern LLMs are trained on huge collections of text and contain billions of adjustable internal parameters (their model weights), which is what allows the next-piece-of-text prediction to capture grammar, facts, reasoning patterns, and writing style well enough to hold a convincing conversation rather than just producing plausible-looking gibberish.
How it works
Most LLMs use a transformer architecture, which processes text as a sequence of tokens and uses a mechanism called attention to weigh how relevant every earlier token is to predicting the next one. Building an LLM happens in stages: a base model is first pre-trained on enormous collections of training data to learn language and world patterns, then usually further adjusted through fine-tuning and alignment techniques so it follows instructions, adopts a consistent assistant or character behavior, and (for most commercial models) declines certain requests.
During an LLM chat, the model runs at inference time: the model isn't learning from your conversation in any permanent sense, it's running a fixed set of trained weights forward to generate a response, one token at a time, within the limits of its context window.
Adult-AI use
Every character chatbot, AI companion, and roleplay platform is a product experience layered on top of an LLM: the persona, memory system, and interface are what make it feel like "a character," but the actual sentence-by-sentence generation is the LLM doing next-token prediction under a system prompt. Which specific LLM a platform uses (a mainstream commercial API versus a self-hosted open-weight model, and how heavily it's been fine-tuned or left as an uncensored model) is one of the biggest practical differences between adult AI products: it affects writing quality, consistency, cost, and (most visibly to users) how the platform handles NSFW content.
Limitations and misconceptions
- An LLM doesn't "know" things the way a database does. It generates statistically plausible text, which is usually accurate but can produce confident, false statements: see hallucination.
- It has no persistent memory of you by default. Anything it "remembers" across messages either sits within the current context window or is being reinserted by an external character memory or RAG system: the model itself doesn't update its own weights from your chat.
- Bigger isn't automatically better for every use. Smaller, more efficient models are often preferred for local or low-latency companion apps even though larger models are generally more capable.
Technical context
The local LLM dossier connects weights, GGUF, execution backends, templates, caches and character-chat front ends.