What does 'training' a model mean, and what does it learn?
Training is the process of adjusting the model's internal parameters using vast amounts of text data. Think of an LLM as an enormous calculator with tens of billions of numerical values (called parameters or weights). Training works like this: feed it large amounts of text, have it repeatedly predict the next token, adjust the values when the prediction is wrong so it predicts better next time. Repeat this billions of times until those values reach a state that produces accurate predictions.
After training, the model has not learned 'a list of facts' — it has learned 'the statistical patterns of language': what kinds of words follow what kinds of words, what kinds of concepts relate to what kinds of concepts. This lets it generate plausible-sounding answers to questions it has never seen before — because it understands the structure and patterns of language.
Why does Claude give slightly different answers each time, even to the same question?
The main reason is the temperature setting. Even at lower temperatures, the model doesn't always select the single highest-probability token — it samples from a probability distribution. As long as temperature is not set to zero (full determinism), each sample can produce slightly different results.
Another reason is context sensitivity: even when you think you asked 'exactly the same question,' there may be subtle differences in the conversation — things you said earlier, minor timing differences, small system-level variations — all of which influence the model's judgment about the most likely next token. This variability is a feature for creative tasks (different results each time) but something to be aware of for tasks requiring precise consistency (such as code generation).
Does an LLM actually 'understand' what it says?
This is a deeply philosophical question without a settled answer, but we can describe what we do know. LLMs have genuinely learned complex language structure and relationships between concepts, and can perform analogy, reasoning, and combination of knowledge in very sophisticated ways. This makes them appear to 'understand' in many situations.
But their 'understanding' is fundamentally different from human understanding. Human understanding is grounded in bodily perception, social experience, emotion, and real-world interaction. An LLM's 'understanding' is grounded in the statistical relationships of language patterns — it knows the word 'fire' tends to appear alongside 'hot,' 'dangerous,' 'light,' but it has never felt heat. This difference doesn't matter in most practical applications, but it shows up in scenarios requiring emotional resonance, embodied judgment, or genuine perception of the physical world.
Advanced: what does it mean for a model to be 'smarter,' and does more parameters equal smarter?
'Smarter' is a fuzzy word in the LLM context, usually meaning better performance on specific benchmarks — math reasoning, code generation, multi-step logic. More parameters (larger models) do generally correlate with better benchmark performance, but the relationship is not linear and has many exceptions.
First, models with the same number of parameters can differ greatly in outcome based on the quality and diversity of training data and the design of training techniques. Second, larger models perform better on some tasks but on tasks requiring speed and cost efficiency, smaller models may be more appropriate. Most importantly: 'smarter' and 'more suitable for your task' are not the same thing. Claude Sonnet performs well enough on most everyday tasks and is much faster than Opus; Opus's advantages mainly show up on edge cases like complex reasoning and long-form analysis.
'How does AI actually answer your questions?' is a more interesting question than most people expect — and understanding it will help you get more out of it. This article explains the actual mechanism by which LLMs (large language models) generate text, in a way that requires no technical background. Not just 'it learned a lot of things,' but what it is actually doing every time it speaks.
An LLM doesn't generate text the way a person might — by thinking through a full sentence and then saying it. Its process is closer to: look at all the current input, pick the single most likely next token (a language unit), add that token to the output, and repeat until done.
A token is the basic unit a language model works with — think of it as slightly smaller than a word. In English, roughly every 4 characters is one token; in Chinese, roughly 1-2 characters is one token. 'The cat sat' is about 3-4 tokens. The model generates one token at a time, adds it to the context, and generates the next one, repeating until it stops.
Each time it needs to pick the next token, the model doesn't guess randomly or look up a fixed rule. Inside, it runs an enormous numerical computation that takes all the current context and converts it into a probability distribution — essentially, 'what is the probability of each possible next token appearing here.'
For example, if you ask 'The capital of France is,' the model's probability distribution might assign 99% to 'Paris,' with other options near zero. If you ask 'Write a poem about the ocean, first line:' there are many plausible next tokens, and the probability is spread across many candidates.
Once the probability distribution is computed, the model selects one token from it. A parameter called temperature controls how this selection works. At low temperature (near zero), the model almost always picks the highest-probability token — outputs are very consistent and predictable, which suits tasks requiring precision (looking up facts, writing code). At high temperature, the model has a greater chance of picking lower-probability tokens — outputs are more varied and creative, but also more likely to go in unexpected directions. Good for creative writing; potentially prone to drift in factual tasks.
Claude has different default temperature settings for different tasks, and you can adjust it in the API. Understanding this mechanism explains why asking the same question twice sometimes gives slightly different answers.
The way LLMs work determines one of their fundamental characteristics: every token the model generates is 'the most likely continuation given the current context,' not 'a fact that was verified before being stated.' In other words, it has no built-in sense of when to stop and say 'I don't know' — it will keep generating plausible-sounding continuations even when those continuations are wrong.
This is why AI 'hallucinates' — not because it is lying, but because its operating mechanism is fundamentally 'predict the most likely next token,' not 'query a fact database and output a verified answer.' Understanding this tells you why it is still important to verify AI output for high-accuracy tasks.
Two direct practical effects of understanding how LLMs generate text. First, you know it has no memory between conversations — every reply involves reading everything in the current conversation from scratch, not drawing on long-term memory. Second, you know it outputs the most linguistically plausible continuation, not a verified fact. This helps you use it correctly: treat it as a powerful language collaborator, not an infallible encyclopedia with comprehensive knowledge.