Librarian Joss lays a new page from the SH-1 packet on the table. It is a practice quiz SH-1 wrote, and every line reads smoothly.
"It writes like a person!" Comet says. "Let's show the panel today."
Wren taps his clipboard. "What does the evidence say? How does it write anything at all?"
Comet pulls out the original Tiny Text tally table. "Maybe the way Tiny Text would. One word at a time."
Nova hovers over the table and projects a die beside it. "Would you like a hint?" she asks. "Turn each tally into die faces, then roll."
Comet rolls and writes: the library opens at eight. "It works!"
She rolls again and writes: the gym opens at eight.
Wren checks the cards. "T6 says the gym opens at seven."
Why does a model that predicts the next word sometimes say false things confidently?
This week you will run Tiny Text with a die, sort what it writes, and rate SH-1's first risk.
Generative AI makes new "synthetic" content, such as text, images, audio and video. That content imitates the data the model learned from.
Generative models make outputs that follow the statistical patterns of their training data.
Large language models predict the next word, or token, in a sentence.
Tiny Text is far simpler than any real language model. It shares one idea with them: it picks the next word from patterns in its training data.
Last week the crew trained Tiny Text on six cards. Each tally became die faces: the more often a word came next, the more faces it gets.
This week the crew runs the original six TT-1 cards, from before last week's fix for the library's new hours. So the cards still say the library opens at eight.
| Card | Training sentence |
|---|---|
| T1 | the library opens at eight |
| T2 | the library has quiet rooms |
| T3 | the lab opens at nine |
| T4 | the lab has new desks |
| T5 | the library opens at eight |
| T6 | the gym opens at seven |
| After this word | Roll the die | Next word |
|---|---|---|
| the | 1, 2 or 3 | library |
| the | 4 or 5 | lab |
| the | 6 | gym |
| library | 1, 2, 3 or 4 | opens |
| library | 5 or 6 | has |
| lab | 1, 2 or 3 | opens |
| lab | 4, 5 or 6 | has |
| gym | no roll | opens |
| opens | no roll | at |
| at | 1 or 2 | eight |
| at | 3 | nine |
| at | 4 | seven |
| at | 5 or 6 | roll again |
| has | 1, 2 or 3 | quiet (then rooms) |
| has | 4, 5 or 6 | new (then desks) |
A sentence always starts with "the". It ends after eight, nine, seven, rooms or desks. Tiny Text is a made-up example the crew runs on paper.
Comet starts with "the" and rolls a 2, so the next word is library. Then she rolls a 1, so library is followed by opens.
Opens always goes to at, so she does not roll. Then she rolls a 2 after at, which gives eight.
Her sentence is "the library opens at eight." It matches cards T1 and T5.
| Statement | True or false? |
|---|---|
| Generative models make outputs that follow patterns in their training data. | ? |
| Large language models predict the next word or token. | ? |
| Generative AI can make text, images, audio and video. | ? |
| Tiny Text works exactly like a real language model. | ? |
Great start, reviewer. Tomorrow you will list every sentence Tiny Text could ever write and sort them.