The Next-Word Game: How Large Language Models Work

Grant Sanderson opens with a thought experiment: a film script where the assistant's lines are missing, plus a hypothetical engine that guesses plausible continuations. Feed the fragment in, take the engine's guess, append it, repeat, and you have, in miniature, the loop behind every chatbot exchange.

An LLM never commits to a single continuation; it spreads belief across the whole vocabulary as probabilities. Builders wrap that core with a dialogue template, then sample from the distribution instead of always taking the top candidate, which is why a fixed model still surprises you on every rerun.

Those probability judgments are earned from text at internet scale. Reading the GPT-3 corpus around the clock would occupy a human for more than two and a half millennia, and frontier training runs since have grown far beyond that.

Training is framed as tuning: hundreds of billions of continuous knobs decide every probability the model emits. Nobody sets them by hand; they start random, producing gibberish, and every text example nudges them a fraction toward better guesses.

The nudge has a name, backpropagation: hide the final token of a passage, compare the model's guess with the truth, shift the weights toward the real ending. Repeated across trillions of passages, the system starts handling passages it has never met.

The arithmetic bill is hard to picture: a billion operations a second would still need on the order of a hundred million years. That covers pre-training only, since completing random internet text is a different goal from assisting a user; a second stage, reinforcement learning from human feedback, reshapes the raw completer into something that behaves like an assistant.

None of this runs without graphics processors, and not every design exploits them: pre-2017 models chewed through words in order until the transformer, introduced by a Google team, absorbed whole passages side by side. Words become number lists, attention lets those lists negotiate meaning from surrounding context, a waterside sense of bank rather than a financial one, feed-forward layers add memory, and the enriched final vector votes on what comes next. The framework is engineered, yet the fluency that falls out of it is not, which is why tracing any single answer back to its causes stays so difficult.