The classroom request behind search sheets

Ben Swift · 17 September 2026

In March, Samantha Ephraims, a teacher from regional Queensland, wrote asking for a hands-on activity that showed a group of about six how a language model generates text. Her lessons ran for 25 minutes. That is tight for either version of the core activity. Students can build a bigram table and roll dice against its counts, or cut a text into token pairs and draw them from buckets, where the cutting alone takes a good part of a lesson.

Sam tried both. In her own classroom, groups worked from pre-generated booklets and D10 dice, one model trained on The Cat in the Hat and one on the much longer Frankenstein, then compared what their models wrote. She also printed and cut out the Cat in the Hat tokens (“Eeek!”). The booklets were easier to manage with groups.

She then presented the activities to 40 teachers at ConASTA, the Australian Science Teachers Association’s national conference, and put a QR code for this website on her slides. The tech teachers were keen on the tokens, but the room struggled to see how to fit them into the spaces they actually teach in. A pile of paper slips needs time, table space and some confidence that none will end up under the furniture. The group suggested posters instead, with the token pairs jumbled and stuck around the walls, and Sam pictured A1 sheets in her own small lab. “Space is always a premium,” she wrote.

Search sheets came out of that suggestion. Each set shuffles every token pair in a text and deals them across a stack of A4 pages. Print the pages large and put them on the walls, and you have Sam’s posters; hand out one page per person instead, and the model exists only across the room. Someone reads out the last word on the board, and everyone with that word on their sheet puts a hand up. Whoever gets picked reads out the word that follows it. Common continuations get more hands than rare ones, so picking a hand at random samples from something close to the model’s probability distribution.¹

Four sets are ready to print from the tools page, ranging from 15 sheets for Green Eggs and Ham to 79 for a Simple English Wikipedia article on Australia. The Cat in the Hat set Sam asked for has 36 sheets, and also comes as A5 sheets printed two-up on A4. The first page of every set is a brief for whoever is running it. The talks use the same sheets, and each comes as a single download with slides, presenter notes, sheets and brief (here’s the pack for the 15-minute talk).

Thanks to Sam for doing the cutting, running the sessions and writing up such detailed notes. If you have adapted LLMs Unplugged for your own classroom, tell me what you changed.