Ju Lin's AI Weblog

Next Token Prediction, Explained in Plain Language

An AI writes one little piece at a time. It never knows the whole sentence up front — it only ever picks what comes next.

Words get cut into pieces

unbelievableunbelievable

The pieces are called tokens. A short word is usually one piece. A long word breaks into a few.

It's read a mountain of text

peanut butter andbread?jam?rocks?honey?jelly?

It has seen, over and over, which piece usually follows which. That's all the "knowing" it really does.

It scores every possible piece

58%24%12%1%jellyjambreadrocks

Not one answer — a whole list, each with a score. "jelly" scores highest here. "rocks" barely registers, but it still gets a number.

Pick one

jellyjambreadrocks

Usually the top score wins. Sometimes it rolls dice weighted toward the tall ones instead — just enough to keep the writing from sounding the same every time.

Do that a few hundred times

The sentence "The cat sat on the" grows one word at a time. Beneath it, a scoring table appears each round showing the candidate next words and their percentages, with the winner's bar in the accent color, then fades as that word gets glued onto the sentence above — repeating for mat, and, purred, and a final period, until the whole sentence "The cat sat on the mat and purred." is built and a cursor blinks at the end.The cat sat on thematandpurred.mat64%sofa19%roof11%and52%.28%with13%purred44%slept31%sang16%.71%softly18%again11%score the list, pick a winner, glue it on, ask again

Score the list, pick a winner, glue it onto the end — then ask again with the whole longer line. Watch the table underneath update each round as a new piece gets picked and stitched on. No plan, no outline — just the next little piece, guessed over and over, until it decides to stop. Next token prediction — that's the real name for it.