How LLMs Really Work: From Tokens to a Machine That Guesses the Next Word

Why Chat Feels Human

Techquickie's “How Do LLMs Work?” opens with a simple puzzle: why does a ChatGPT reply sometimes feel like a person is on the other side? The host frames the model as a pile of numbers guessing the next word , yet it can write emails, arrange flights and even surface website flaws once that pile is tuned. To explain what happens when you ask ChatGPT or Claude anything, the video turns to Colin Kilty, senior machine learning engineer at Red Hat , whose first line is disarmingly candid: language models are a black box — information goes in, information comes out, what happens in the middle is not fully known. The video then asks “do we know anything at all?” and starts unpacking the parts we do understand, as if learning the legend of a map before reading it.

The Math of Language: Tokens and a Concept Map

At the core is the token , roughly four characters splayed into numbers. Every token is turned into a list of numbers and placed on a huge map of concepts; tokens with related meanings sit nearby. The classic illustration is sea and lake ending up close together because both are large bodies of water. Two technical terms get plain glosses here: vector embedding (placing each token at a numeric coordinate) and latent space (the abstract map where all concepts live) . Think of it as a dictionary for the model: language becomes numbers first, reasoning second, and nearby concepts stay nearby because the geometry captures similarity. Once that picture clicks, the rest of the machinery becomes easier to follow.

The numbers that do the heavy lifting are weights . The analogy is to human neurons whose connections can be stronger or weaker; in artificial nets that strength is a number, the weight, and the whole collection is the “miserable pile” the host keeps returning to. Kilty puts it plainly: weights embody the knowledge the model holds, derived from vast human text. They are the raw material for every calculation that turns your prompt into a reply, and they are the main thing that separates one model from another. Crucially, nobody hand-coded grammar, sarcasm or any character quirks; those patterns emerged as weights shifted to better predict what comes next.

A Factory of Trial and Error: Training

Training is described as trial and error at speed and broken into three steps: 1) take a corpus and hide the next word, 2) let the model guess, 3) nudge the numbers a little toward the right answer when it is wrong. Do that trillions of times across much of the internet and you end up with a machine excellent at next-word guessing. The video drops a dry “Worth it.” but the point is serious: the model does not learn absolute truth, it learns what the survey of everything it has read says is most likely next. As more data pours in and weights keep nudging, the guesses get better, which is why the same architecture sounds smarter at larger scale.

The recap ties that to inference , the act of answering your prompt. The model turns your message into numbers, pushes them through its tuned pile of weights, and computes the most likely next word; if you asked a question, that word is the start of the answer. Then it runs again on the prompt plus the newly generated word, and again, autoregressively , building the reply one token at a time until it stops. The result feels less like a calculator's guaranteed output and more like Family Feud — the survey says . Grammar and tone were never programmed in; they formed because they were in the data and better guessing demanded them.

Why It Doubles Down When Wrong

If guessing is so good, why does it still confabulate? The video defines hallucination not as lying on purpose but as an unintentional error of picking the wrong next token without suitable context . Once that token is emitted it becomes part of the input, and the model is locked onto that path. The host likens it to a person who is bad at admitting a mistake and keeps doubling down; the machine does too, because the erroneous word is now data it must continue from. That same lock-in explains why a challenge often flips the answer instantly: your correction is new input that nudges probabilities back.

That is where reasoning models help. These are systems that show a little “thinking” phase before replying, which the video frames as the model talking to itself for a bit to generate extra context. On hard problems the extra text raises the chance of landing on the right words. Whether that counts as genuine reasoning is, in the host's own shrug, anyone's guess. Researchers are trying to watch those internal steps directly, but the tools are still rudimentary and cannot capture everything happening across trillions of weights at once.

Dark Matter: Mapping What We Built

The most sober stretch is about interpretability. Kilty notes we do know roles for some layers — the attention layer that decides which tokens to focus on , for instance — yet the overall box remains dark. Teams are cracking models open to trace how the calculations move, but one idea does not equal one place . The textbook example is a single spot in a vision model that fires for cat faces, car fronts and cat legs alike, a phenomenon called polysemanticity or superposition . With trillions of weights, the Anthropic researcher leading this line estimates only a small fraction has been mapped; the rest is dubbed dark matter . In other words, the makers are still charting the machine they built, and the map has far more blank than ink.

The closing lifts the view from engineering to philosophy. Some researchers argue that as models scale they may be converging on something like Plato's universal truth — a thought space where every concept that could exist already does , encompassing every discovery, formula, proof and law of our universe and any we might yet touch. The counter-view is blunt: maybe it is just glorified autocomplete . The video is explicit that this has been a super-high-level sweep and points curious viewers to deeper write-ups by its labs team and to a companion piece on running a model at home . The distilled takeaway is compact: we understand the building blocks — tokens, weights, training, inference — but their sum did not produce a thinking machine, it produced a guessing machine; the appearance of thought grew as we fed it the internet.

Put simply, the video leaves you with a usable compass: ask a language model for likelihood, not certainty, just as you would a survey. The concept map, the weight pile and the autoregressive loop explain in one frame why the system is fluent, why it can get stubbornly wrong, and why adding context often repairs it. Trying a small model locally makes those four pieces tangible, while the darker corners remain an active research front. For the next step, the path the video suggests is careful curiosity — read deeper, test locally, stay empirical rather than mythic.