Underneath all the marketing language around modern AI, a language model is doing one thing on a loop: looking at the words that came before and making a probability guess about which word should land next. The interactive below lets you watch that guess happen out in the open, so instead of taking it on faith you can see the model's actual candidate list, nudge it toward boring or wild with a dial, and feel what's really happening inside the box. The whole point here is to build the mental picture you'll keep using for every module that follows this one.
Pick one of these starters. You can predict word by word, then watch a probability distribution behind the scenes, authored for teaching in the same shape a real model computes.
One honest simplification: this widget predicts whole words so the display stays readable, while a real model predicts tokens, which are often pieces of words.
Temperature is the dial that controls how confidently the model commits to its top pick on any given step. Set it low and the model almost always grabs the highest-probability word, which produces output that feels predictable and a little flat, and crank it up and the probabilities flatten enough that the model starts reaching for less likely candidates until you tip into creative territory or genuine nonsense depending on how far you push it. The exact same prompt produces different output every time you move this slider, which is one of the underrated reasons two people running the same prompt can get wildly different results.
The model has no internal sense of whether its top pick is actually correct in the way a researcher or fact-checker would use that word, since all it knows is which token sits at the highest probability given the patterns it absorbed during training. That gap is why a language model can produce a beautifully written paragraph about a study that does not exist, complete with confident author names and a plausible journal title, without ever knowing that the whole thing is fabricated. Reading smoothly and being factually right turn out to be two completely separate properties, and the model is only optimizing for the first one.
A language model produces a confident-sounding citation for a research paper that does not exist. Which one of the following best explains why?