Large Language Model (LLM)

How it actually works (and the part the demos skip)

Strip away the marketing and an LLM does one thing: given a sequence of tokens, it outputs a probability distribution over the next token, samples one from it, appends it, and repeats. Training is self-supervised — the model is shown ordinary text with the next token hidden, scored on how close its guess was, and corrected across billions of examples until its predictions track real language. There is no lookup table of facts inside it. What looks like knowledge is statistical regularity compressed into the weights. A sampling setting (often called temperature) controls how sharply it favors the single most likely token over a plausible alternative, which is why the same prompt can return different answers. Three consequences follow directly from this design, and they explain almost every LLM behavior you will fight with in production:

• It is stateless. The model remembers nothing between calls. Every 'conversation' is the full history re-sent in the prompt each turn. That context window is finite because attention cost rises with its length, and every token in it is metered — so longer memory means higher latency and higher cost, not just more recall.
• It is frozen in time. The weights stop changing when training ends. Anything after the cutoff — or anything that lives only in your CRM, your tickets, your contracts — it cannot know unless you place it in the prompt at runtime.
• It is confidently approximate. Because it samples plausible text, it will produce fluent, well-formed answers that are wrong (a 'hallucination'). Fluency is not evidence of accuracy. A correct paragraph and a fabricated one come out of the exact same mechanism.

Why it matters for buyers and analysts

The executive takeaway is counterintuitive: in most business systems, the LLM is the least important and most swappable component. Raw model capability is largely a commodity that resets across vendors every few months, and it is usually the cheapest part of the system to change — often a one-line config swap. What is not commodity, and what actually decides whether the system returns value, is everything around the model: the data you feed it, the tools you let it act through, the guardrails on what it is allowed to do, and the measurement of whether its output moved a real number. A frontier model on stale or missing data loses to a modest model on clean, well-governed data, every time. There is a quiet cost lever here too — a smaller model that clears your quality bar can run several times cheaper and faster than the largest one, and at scale that gap is the difference between a pilot and a P&L line. So 'which model should we use?' is rarely the question that determines ROI. 'What does the model see, what can it do, and how do we know it worked?' is.

Where it fits in an AI system

An LLM on its own is a text generator, not a product. Useful systems wrap it. Retrieval (RAG) pulls the right facts from your data and places them into the prompt so answers are grounded in records instead of guessed from training. Tools and function-calling let the model trigger actions — query a record, file a case, send a quote — turning a writer into something that does work. Orchestration decides which model, which data, and which tools handle each step, and where a human signs off. The LLM is the reasoning layer in that stack; the value, and the liability, live in the connections to your systems and the discipline around them. That is the practical line between a clever demo and a system someone is accountable for in production — and it is where most of the real engineering goes.

• LLM alone: drafts, summarizes, classifies — but invents specifics and knows nothing about your business.
• LLM + your data (retrieval): grounded answers that cite the actual record instead of a plausible guess.
• LLM + tools (agent): not just answers but actions inside your systems, which is exactly where outcomes and risk both concentrate.
• LLM + measurement: the only configuration where you can tell whether any of it earned its keep — and the one most teams skip.