An LLM, or large language model, does only one thing. Given a text, it works out which token (a chunk of text, often part of a word) is most likely to come next. It adds that token, then starts again. A three-paragraph answer is hundreds of turns of that loop.
On an example sentence, the model picks one of the likely tokens, adds it to the sentence, then starts again.
A machine for plausible continuations
The model learned those probabilities from a vast amount of text. It has seen how sentences, arguments and code fit together, and it reproduces those patterns very fluently. It can summarize a contract or write a function because those tasks look like millions of examples it has read.
Key concept
An LLM predicts the most plausible continuation of a text. It keeps no memory from one request to the next, and it has no judgment of its own: on its own, it does not check whether its answer is true, or whether it is good for you. How fluent an answer sounds tells you nothing about whether it is right.
Two practical consequences. With no memory, the model only knows about your situation what you send it on each request, in its context window. With no judgment, it gives the answer a text like yours usually calls for: a well-written average. It can also state something false as confidently as something true: those are hallucinations.
Request 1 sends your context and your question; the model answers from what it received.
What it means for you
For founders
An LLM speeds up anything that resembles what already exists: a first draft, a summary, a standard reply. It will not tell you which market to go after. Use it to execute faster, not to decide for you.
For product managers
The model only knows about your user what your product sends it. Specify that context like a feature: which data, at what moment, and what happens when it is missing.
For tech leads
Treat each model call as a non-deterministic function. The same input can return two different outputs, so validate the format, log inputs and outputs, and test on real cases. In a large company, add an architecture question: where the data you send ends up, and under which contract.
For business leaders
An LLM produces a fluent synthesis of what it has read, which makes it an average. For an internal memo or a first triage, that is a gain. For a strategic decision, the value lies in what departs from the average, and that work stays yours.
Go deeperFrom prediction to assistant
A model is first pretrained to predict the next token across a large body of text. At that stage it can complete documents but follows instructions poorly. It is then refined, notably through reinforcement learning from human feedback (RLHF): people rank several answers, and the model learns to prefer the higher-ranked ones. That is what makes it a useful assistant.
Current models can also think before they answer. According to Anthropic's documentation, as of October 2026, they then restate the question, try several approaches and check intermediate results. If the application allows it, they call tools, such as web search or code execution, to check a source or a calculation. The underlying mechanism is still prediction: the model only checks what the application lets it check. Even then, a fluent answer is not necessarily a right one.
The classic mistake
fAIling RULE IX
The trap: Let the AI own the strategy
Ask the model to write your strategy and you get a clean, structured, convincing document. It is also the average of every strategy it has read, without your customers, your margins or your constraints. Use it to stress-test a plan you wrote, not to write it.
Read rule IX in the manifestoTry it
Ask an AI assistant the same business question in three new conversations. Compare the answers: what changes, what stays, and what is missing because the model does not know your context. Then ask again, adding that context in two sentences, and note the difference.
Around this topic
Go deeperFine-tuning
Fine-tuning continues a model's training on your own examples, so it picks up their style, format or vocabulary. It mostly pays off on a narrow, repetitive task backed by many clean examples. It is the wrong tool for facts that change: those belong in the context, sent with the request. Not every vendor offers it, and as of October 2026 the Claude API does not offer it as self-serve. Start with better instructions and good examples, and only move to fine-tuning if a test shows that is not enough.
Go deeperTraining, use and the cut-off date
A model learns during training, then it is frozen. Using it does not change it: every request starts from the same model. What it knows therefore stops at a cut-off date, the limit of its knowledge, which vendors publish. As of October 2026, for example, Anthropic gives June 2026 for its four current Claude models. Anything that happened later, the model only knows if you tell it. The memory features of some assistants, such as Claude's, rely on the same principle: the application keeps notes about you and sends them to the model along with your messages.
Example draw 1: “The meeting moves to Monday morning at…” Each step drew the most likely token, but it was still a draw.
Go deeperWhy answers vary
At each step, the model assigns a probability to every possible token, then draws one at random. That is why the same question can get two different answers. For a long time, a setting called temperature made that draw more or less cautious. Vendors increasingly fix it. As of October 2026, according to Anthropic's documentation, its recent models accept only the default value, the new version of its API rejects the setting, and no other setting replaces it. For consistent answers, rely on precise instructions, examples, an answer in a format fixed in advance, and evals that measure the differences.
Go deeperBeyond text as input
Current LLMs also read images: a screenshot, a chart, a scanned invoice. An image is counted in tokens, like text, so it costs money on every request. The model can misread a blurry image or a small figure, just as it can misquote a sentence. Reading an image is not the same as creating one: image generation is the job of other models, covered in Other kinds of models.
What to do with it
Use it to execute faster (a first draft, a summary) and to stress-test a strategy you wrote; keep the decision for yourself.
Specify the context your product sends it like a feature: which data, at what moment, and what happens when it is missing.
Treat each call as non-deterministic: validate the format, log inputs and outputs, and test on real cases.